313 files changed, 101622 insertions, 0 deletions

diff --git a/ml/dlib/dlib/external/cblas/CMakeLists.txt b/ml/dlib/dlib/external/cblas/CMakeLists.txt
new file mode 100644
index 000000000..0d800ae13
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/CMakeLists.txt
@@ -0,0 +1,182 @@
+#
+# This is a CMake makefile.  You can find the cmake utility and
+# information about it at http://www.cmake.org
+#
+
+
+cmake_minimum_required(VERSION 2.8.12)
+project(cblas)
+
+
+enable_language (Fortran)
+
+set(CMAKE_POSITION_INDEPENDENT_CODE True)
+add_definitions(-DADD_ -DF77_INT=ptrdiff_t)
+
+add_library(cblas STATIC 
+   cblas_caxpy.c
+   #cblas_ccopy.c
+   cblas_cdotc_sub.c
+   cblas_cdotu_sub.c
+   #cblas_cgbmv.c
+   cblas_cgemm.c
+   cblas_cgemv.c
+   cblas_cgerc.c
+   cblas_cgeru.c
+   #cblas_chbmv.c
+   #cblas_chemm.c
+   #cblas_chemv.c
+   #cblas_cher2.c
+   #cblas_cher2k.c
+   #cblas_cher.c
+   #cblas_cherk.c
+   #cblas_chpmv.c
+   #cblas_chpr2.c
+   #cblas_chpr.c
+   cblas_cscal.c
+   #cblas_csscal.c
+   #cblas_cswap.c
+   #cblas_csymm.c
+   #cblas_csyr2k.c
+   #cblas_csyrk.c
+   #cblas_ctbmv.c
+   #cblas_ctbsv.c
+   #cblas_ctpmv.c
+   #cblas_ctpsv.c
+   #cblas_ctrmm.c
+   #cblas_ctrmv.c
+   cblas_ctrsm.c
+   #cblas_ctrsv.c
+   #cblas_dasum.c
+   cblas_daxpy.c
+   #cblas_dcopy.c
+   cblas_ddot.c
+   #cblas_dgbmv.c
+   cblas_dgemm.c
+   cblas_dgemv.c
+   cblas_dger.c
+   #cblas_dnrm2.c
+   #cblas_drot.c
+   #cblas_drotg.c
+   #cblas_drotm.c
+   #cblas_drotmg.c
+   #cblas_dsbmv.c
+   cblas_dscal.c
+   #cblas_dsdot.c
+   #cblas_dspmv.c
+   #cblas_dspr2.c
+   #cblas_dspr.c
+   #cblas_dswap.c
+   #cblas_dsymm.c
+   #cblas_dsymv.c
+   #cblas_dsyr2.c
+   #cblas_dsyr2k.c
+   #cblas_dsyr.c
+   #cblas_dsyrk.c
+   #cblas_dtbmv.c
+   #cblas_dtbsv.c
+   #cblas_dtpmv.c
+   #cblas_dtpsv.c
+   #cblas_dtrmm.c
+   #cblas_dtrmv.c
+   cblas_dtrsm.c
+   #cblas_dtrsv.c
+   #cblas_dzasum.c
+   #cblas_dznrm2.c
+   #cblas_icamax.c
+   #cblas_idamax.c
+   #cblas_isamax.c
+   #cblas_izamax.c
+   #cblas_sasum.c
+   cblas_saxpy.c
+   #cblas_scasum.c
+   #cblas_scnrm2.c
+   #cblas_scopy.c
+   cblas_sdot.c
+   #cblas_sdsdot.c
+   #cblas_sgbmv.c
+   cblas_sgemm.c
+   cblas_sgemv.c
+   cblas_sger.c
+   #cblas_snrm2.c
+   #cblas_srot.c
+   #cblas_srotg.c
+   #cblas_srotm.c
+   #cblas_srotmg.c
+   #cblas_ssbmv.c
+   cblas_sscal.c
+   #cblas_sspmv.c
+   #cblas_sspr2.c
+   #cblas_sspr.c
+   #cblas_sswap.c
+   #cblas_ssymm.c
+   #cblas_ssymv.c
+   #cblas_ssyr2.c
+   #cblas_ssyr2k.c
+   #cblas_ssyr.c
+   #cblas_ssyrk.c
+   #cblas_stbmv.c
+   #cblas_stbsv.c
+   #cblas_stpmv.c
+   #cblas_stpsv.c
+   #cblas_strmm.c
+   #cblas_strmv.c
+   cblas_strsm.c
+   #cblas_strsv.c
+   cblas_xerbla.c
+   cblas_zaxpy.c
+   #cblas_zcopy.c
+   cblas_zdotc_sub.c
+   cblas_zdotu_sub.c
+   #cblas_zdscal.c
+   #cblas_zgbmv.c
+   cblas_zgemm.c
+   cblas_zgemv.c
+   cblas_zgerc.c
+   cblas_zgeru.c
+   #cblas_zhbmv.c
+   #cblas_zhemm.c
+   #cblas_zhemv.c
+   #cblas_zher2.c
+   #cblas_zher2k.c
+   #cblas_zher.c
+   #cblas_zherk.c
+   #cblas_zhpmv.c
+   #cblas_zhpr2.c
+   #cblas_zhpr.c
+   cblas_zscal.c
+   #cblas_zswap.c
+   #cblas_zsymm.c
+   #cblas_zsyr2k.c
+   #cblas_zsyrk.c
+   #cblas_ztbmv.c
+   #cblas_ztbsv.c
+   #cblas_ztpmv.c
+   #cblas_ztpsv.c
+   #cblas_ztrmm.c
+   #cblas_ztrmv.c
+   cblas_ztrsm.c
+   #cblas_ztrsv.c
+
+   cdotcsub.f
+   cdotusub.f
+   dasumsub.f
+   ddotsub.f
+   dnrm2sub.f
+   dsdotsub.f
+   dzasumsub.f
+   dznrm2sub.f
+   icamaxsub.f
+   idamaxsub.f
+   isamaxsub.f
+   izamaxsub.f
+   sasumsub.f
+   scasumsub.f
+   scnrm2sub.f
+   sdotsub.f
+   sdsdotsub.f
+   snrm2sub.f
+   zdotcsub.f
+   zdotusub.f
+      )
+
diff --git a/ml/dlib/dlib/external/cblas/README b/ml/dlib/dlib/external/cblas/README
new file mode 100644
index 000000000..a89feaffe
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/README
@@ -0,0 +1,7 @@
+This folder contains a copy of CBLAS (from http://www.netlib.org/blas/) which
+has been setup so you can compile it with CMake.  It also only compiles the
+part of CBLAS needed by dlib.
+
+Most BLAS libraries come with CBLAS, however, some don't.  In particular, if
+you are using the BLAS that comes with MATLAB then you will need this CBLAS
+code linked into your own to get dlib working with MATLAB's built in BLAS.
diff --git a/ml/dlib/dlib/external/cblas/cblas.h b/ml/dlib/dlib/external/cblas/cblas.h
new file mode 100644
index 000000000..f91557e74
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas.h
@@ -0,0 +1,575 @@
+#ifndef CBLAS_H
+#define CBLAS_H
+#include <stddef.h>
+
+/*
+ * Enumerated and derived types
+ */
+#define CBLAS_INDEX size_t  /* this may vary between platforms */
+
+enum CBLAS_ORDER {CblasRowMajor=101, CblasColMajor=102};
+enum CBLAS_TRANSPOSE {CblasNoTrans=111, CblasTrans=112, CblasConjTrans=113};
+enum CBLAS_UPLO {CblasUpper=121, CblasLower=122};
+enum CBLAS_DIAG {CblasNonUnit=131, CblasUnit=132};
+enum CBLAS_SIDE {CblasLeft=141, CblasRight=142};
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * ===========================================================================
+ * Prototypes for level 1 BLAS functions (complex are recast as routines)
+ * ===========================================================================
+ */
+float  cblas_sdsdot(const int N, const float alpha, const float *X,
+                    const int incX, const float *Y, const int incY);
+double cblas_dsdot(const int N, const float *X, const int incX, const float *Y,
+                   const int incY);
+float  cblas_sdot(const int N, const float  *X, const int incX,
+                  const float  *Y, const int incY);
+double cblas_ddot(const int N, const double *X, const int incX,
+                  const double *Y, const int incY);
+
+/*
+ * Functions having prefixes Z and C only
+ */
+void   cblas_cdotu_sub(const int N, const void *X, const int incX,
+                       const void *Y, const int incY, void *dotu);
+void   cblas_cdotc_sub(const int N, const void *X, const int incX,
+                       const void *Y, const int incY, void *dotc);
+
+void   cblas_zdotu_sub(const int N, const void *X, const int incX,
+                       const void *Y, const int incY, void *dotu);
+void   cblas_zdotc_sub(const int N, const void *X, const int incX,
+                       const void *Y, const int incY, void *dotc);
+
+
+/*
+ * Functions having prefixes S D SC DZ
+ */
+float  cblas_snrm2(const int N, const float *X, const int incX);
+float  cblas_sasum(const int N, const float *X, const int incX);
+
+double cblas_dnrm2(const int N, const double *X, const int incX);
+double cblas_dasum(const int N, const double *X, const int incX);
+
+float  cblas_scnrm2(const int N, const void *X, const int incX);
+float  cblas_scasum(const int N, const void *X, const int incX);
+
+double cblas_dznrm2(const int N, const void *X, const int incX);
+double cblas_dzasum(const int N, const void *X, const int incX);
+
+
+/*
+ * Functions having standard 4 prefixes (S D C Z)
+ */
+CBLAS_INDEX cblas_isamax(const int N, const float  *X, const int incX);
+CBLAS_INDEX cblas_idamax(const int N, const double *X, const int incX);
+CBLAS_INDEX cblas_icamax(const int N, const void   *X, const int incX);
+CBLAS_INDEX cblas_izamax(const int N, const void   *X, const int incX);
+
+/*
+ * ===========================================================================
+ * Prototypes for level 1 BLAS routines
+ * ===========================================================================
+ */
+
+/* 
+ * Routines with standard 4 prefixes (s, d, c, z)
+ */
+void cblas_sswap(const int N, float *X, const int incX, 
+                 float *Y, const int incY);
+void cblas_scopy(const int N, const float *X, const int incX, 
+                 float *Y, const int incY);
+void cblas_saxpy(const int N, const float alpha, const float *X,
+                 const int incX, float *Y, const int incY);
+
+void cblas_dswap(const int N, double *X, const int incX, 
+                 double *Y, const int incY);
+void cblas_dcopy(const int N, const double *X, const int incX, 
+                 double *Y, const int incY);
+void cblas_daxpy(const int N, const double alpha, const double *X,
+                 const int incX, double *Y, const int incY);
+
+void cblas_cswap(const int N, void *X, const int incX, 
+                 void *Y, const int incY);
+void cblas_ccopy(const int N, const void *X, const int incX, 
+                 void *Y, const int incY);
+void cblas_caxpy(const int N, const void *alpha, const void *X,
+                 const int incX, void *Y, const int incY);
+
+void cblas_zswap(const int N, void *X, const int incX, 
+                 void *Y, const int incY);
+void cblas_zcopy(const int N, const void *X, const int incX, 
+                 void *Y, const int incY);
+void cblas_zaxpy(const int N, const void *alpha, const void *X,
+                 const int incX, void *Y, const int incY);
+
+
+/* 
+ * Routines with S and D prefix only
+ */
+void cblas_srotg(float *a, float *b, float *c, float *s);
+void cblas_srotmg(float *d1, float *d2, float *b1, const float b2, float *P);
+void cblas_srot(const int N, float *X, const int incX,
+                float *Y, const int incY, const float c, const float s);
+void cblas_srotm(const int N, float *X, const int incX,
+                float *Y, const int incY, const float *P);
+
+void cblas_drotg(double *a, double *b, double *c, double *s);
+void cblas_drotmg(double *d1, double *d2, double *b1, const double b2, double *P);
+void cblas_drot(const int N, double *X, const int incX,
+                double *Y, const int incY, const double c, const double  s);
+void cblas_drotm(const int N, double *X, const int incX,
+                double *Y, const int incY, const double *P);
+
+
+/* 
+ * Routines with S D C Z CS and ZD prefixes
+ */
+void cblas_sscal(const int N, const float alpha, float *X, const int incX);
+void cblas_dscal(const int N, const double alpha, double *X, const int incX);
+void cblas_cscal(const int N, const void *alpha, void *X, const int incX);
+void cblas_zscal(const int N, const void *alpha, void *X, const int incX);
+void cblas_csscal(const int N, const float alpha, void *X, const int incX);
+void cblas_zdscal(const int N, const double alpha, void *X, const int incX);
+
+/*
+ * ===========================================================================
+ * Prototypes for level 2 BLAS
+ * ===========================================================================
+ */
+
+/* 
+ * Routines with standard 4 prefixes (S, D, C, Z)
+ */
+void cblas_sgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const float alpha, const float *A, const int lda,
+                 const float *X, const int incX, const float beta,
+                 float *Y, const int incY);
+void cblas_sgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU, const float alpha,
+                 const float *A, const int lda, const float *X,
+                 const int incX, const float beta, float *Y, const int incY);
+void cblas_strmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float *A, const int lda, 
+                 float *X, const int incX);
+void cblas_stbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const float *A, const int lda, 
+                 float *X, const int incX);
+void cblas_stpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float *Ap, float *X, const int incX);
+void cblas_strsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float *A, const int lda, float *X,
+                 const int incX);
+void cblas_stbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const float *A, const int lda,
+                 float *X, const int incX);
+void cblas_stpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float *Ap, float *X, const int incX);
+
+void cblas_dgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const double alpha, const double *A, const int lda,
+                 const double *X, const int incX, const double beta,
+                 double *Y, const int incY);
+void cblas_dgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU, const double alpha,
+                 const double *A, const int lda, const double *X,
+                 const int incX, const double beta, double *Y, const int incY);
+void cblas_dtrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double *A, const int lda, 
+                 double *X, const int incX);
+void cblas_dtbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const double *A, const int lda, 
+                 double *X, const int incX);
+void cblas_dtpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double *Ap, double *X, const int incX);
+void cblas_dtrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double *A, const int lda, double *X,
+                 const int incX);
+void cblas_dtbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const double *A, const int lda,
+                 double *X, const int incX);
+void cblas_dtpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double *Ap, double *X, const int incX);
+
+void cblas_cgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *X, const int incX, const void *beta,
+                 void *Y, const int incY);
+void cblas_cgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU, const void *alpha,
+                 const void *A, const int lda, const void *X,
+                 const int incX, const void *beta, void *Y, const int incY);
+void cblas_ctrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *A, const int lda, 
+                 void *X, const int incX);
+void cblas_ctbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void *A, const int lda, 
+                 void *X, const int incX);
+void cblas_ctpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *Ap, void *X, const int incX);
+void cblas_ctrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *A, const int lda, void *X,
+                 const int incX);
+void cblas_ctbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void *A, const int lda,
+                 void *X, const int incX);
+void cblas_ctpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *Ap, void *X, const int incX);
+
+void cblas_zgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *X, const int incX, const void *beta,
+                 void *Y, const int incY);
+void cblas_zgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU, const void *alpha,
+                 const void *A, const int lda, const void *X,
+                 const int incX, const void *beta, void *Y, const int incY);
+void cblas_ztrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *A, const int lda, 
+                 void *X, const int incX);
+void cblas_ztbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void *A, const int lda, 
+                 void *X, const int incX);
+void cblas_ztpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *Ap, void *X, const int incX);
+void cblas_ztrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *A, const int lda, void *X,
+                 const int incX);
+void cblas_ztbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void *A, const int lda,
+                 void *X, const int incX);
+void cblas_ztpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void *Ap, void *X, const int incX);
+
+
+/* 
+ * Routines with S and D prefixes only
+ */
+void cblas_ssymv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const float alpha, const float *A,
+                 const int lda, const float *X, const int incX,
+                 const float beta, float *Y, const int incY);
+void cblas_ssbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const int K, const float alpha, const float *A,
+                 const int lda, const float *X, const int incX,
+                 const float beta, float *Y, const int incY);
+void cblas_sspmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const float alpha, const float *Ap,
+                 const float *X, const int incX,
+                 const float beta, float *Y, const int incY);
+void cblas_sger(const enum CBLAS_ORDER order, const int M, const int N,
+                const float alpha, const float *X, const int incX,
+                const float *Y, const int incY, float *A, const int lda);
+void cblas_ssyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const float *X,
+                const int incX, float *A, const int lda);
+void cblas_sspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const float *X,
+                const int incX, float *Ap);
+void cblas_ssyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const float *X,
+                const int incX, const float *Y, const int incY, float *A,
+                const int lda);
+void cblas_sspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const float *X,
+                const int incX, const float *Y, const int incY, float *A);
+
+void cblas_dsymv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const double alpha, const double *A,
+                 const int lda, const double *X, const int incX,
+                 const double beta, double *Y, const int incY);
+void cblas_dsbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const int K, const double alpha, const double *A,
+                 const int lda, const double *X, const int incX,
+                 const double beta, double *Y, const int incY);
+void cblas_dspmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const double alpha, const double *Ap,
+                 const double *X, const int incX,
+                 const double beta, double *Y, const int incY);
+void cblas_dger(const enum CBLAS_ORDER order, const int M, const int N,
+                const double alpha, const double *X, const int incX,
+                const double *Y, const int incY, double *A, const int lda);
+void cblas_dsyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const double *X,
+                const int incX, double *A, const int lda);
+void cblas_dspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const double *X,
+                const int incX, double *Ap);
+void cblas_dsyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const double *X,
+                const int incX, const double *Y, const int incY, double *A,
+                const int lda);
+void cblas_dspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const double *X,
+                const int incX, const double *Y, const int incY, double *A);
+
+
+/* 
+ * Routines with C and Z prefixes only
+ */
+void cblas_chemv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *A,
+                 const int lda, const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_chbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const int K, const void *alpha, const void *A,
+                 const int lda, const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_chpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *Ap,
+                 const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_cgeru(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda);
+void cblas_cgerc(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda);
+void cblas_cher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const void *X, const int incX,
+                void *A, const int lda);
+void cblas_chpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const void *X,
+                const int incX, void *A);
+void cblas_cher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
+                const void *alpha, const void *X, const int incX,
+                const void *Y, const int incY, void *A, const int lda);
+void cblas_chpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
+                const void *alpha, const void *X, const int incX,
+                const void *Y, const int incY, void *Ap);
+
+void cblas_zhemv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *A,
+                 const int lda, const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_zhbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const int K, const void *alpha, const void *A,
+                 const int lda, const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_zhpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *Ap,
+                 const void *X, const int incX,
+                 const void *beta, void *Y, const int incY);
+void cblas_zgeru(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda);
+void cblas_zgerc(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda);
+void cblas_zher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const void *X, const int incX,
+                void *A, const int lda);
+void cblas_zhpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const void *X,
+                const int incX, void *A);
+void cblas_zher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
+                const void *alpha, const void *X, const int incX,
+                const void *Y, const int incY, void *A, const int lda);
+void cblas_zhpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo, const int N,
+                const void *alpha, const void *X, const int incX,
+                const void *Y, const int incY, void *Ap);
+
+/*
+ * ===========================================================================
+ * Prototypes for level 3 BLAS
+ * ===========================================================================
+ */
+
+/* 
+ * Routines with standard 4 prefixes (S, D, C, Z)
+ */
+void cblas_sgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const float alpha, const float *A,
+                 const int lda, const float *B, const int ldb,
+                 const float beta, float *C, const int ldc);
+void cblas_ssymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const float alpha, const float *A, const int lda,
+                 const float *B, const int ldb, const float beta,
+                 float *C, const int ldc);
+void cblas_ssyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const float alpha, const float *A, const int lda,
+                 const float beta, float *C, const int ldc);
+void cblas_ssyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const float alpha, const float *A, const int lda,
+                  const float *B, const int ldb, const float beta,
+                  float *C, const int ldc);
+void cblas_strmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const float alpha, const float *A, const int lda,
+                 float *B, const int ldb);
+void cblas_strsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const float alpha, const float *A, const int lda,
+                 float *B, const int ldb);
+
+void cblas_dgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const double alpha, const double *A,
+                 const int lda, const double *B, const int ldb,
+                 const double beta, double *C, const int ldc);
+void cblas_dsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const double alpha, const double *A, const int lda,
+                 const double *B, const int ldb, const double beta,
+                 double *C, const int ldc);
+void cblas_dsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const double alpha, const double *A, const int lda,
+                 const double beta, double *C, const int ldc);
+void cblas_dsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const double alpha, const double *A, const int lda,
+                  const double *B, const int ldb, const double beta,
+                  double *C, const int ldc);
+void cblas_dtrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const double alpha, const double *A, const int lda,
+                 double *B, const int ldb);
+void cblas_dtrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const double alpha, const double *A, const int lda,
+                 double *B, const int ldb);
+
+void cblas_cgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const void *alpha, const void *A,
+                 const int lda, const void *B, const int ldb,
+                 const void *beta, void *C, const int ldc);
+void cblas_csymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc);
+void cblas_csyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const void *alpha, const void *A, const int lda,
+                 const void *beta, void *C, const int ldc);
+void cblas_csyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const void *beta,
+                  void *C, const int ldc);
+void cblas_ctrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 void *B, const int ldb);
+void cblas_ctrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 void *B, const int ldb);
+
+void cblas_zgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const void *alpha, const void *A,
+                 const int lda, const void *B, const int ldb,
+                 const void *beta, void *C, const int ldc);
+void cblas_zsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc);
+void cblas_zsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const void *alpha, const void *A, const int lda,
+                 const void *beta, void *C, const int ldc);
+void cblas_zsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const void *beta,
+                  void *C, const int ldc);
+void cblas_ztrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 void *B, const int ldb);
+void cblas_ztrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 void *B, const int ldb);
+
+
+/* 
+ * Routines with prefixes C and Z only
+ */
+void cblas_chemm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc);
+void cblas_cherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const float alpha, const void *A, const int lda,
+                 const float beta, void *C, const int ldc);
+void cblas_cher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const float beta,
+                  void *C, const int ldc);
+
+void cblas_zhemm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc);
+void cblas_zherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const double alpha, const void *A, const int lda,
+                 const double beta, void *C, const int ldc);
+void cblas_zher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const double beta,
+                  void *C, const int ldc);
+
+void cblas_xerbla(int p, const char *rout, const char *form, ...);
+
+#ifdef __cplusplus
+}
+#endif
+#endif
diff --git a/ml/dlib/dlib/external/cblas/cblas_caxpy.c b/ml/dlib/dlib/external/cblas/cblas_caxpy.c
new file mode 100644
index 000000000..7579aa707
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_caxpy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_caxpy.c
+ *
+ * The program is a C interface to caxpy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_caxpy( const int N, const void *alpha, const void *X,
+                       const int incX, void *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_caxpy( &F77_N, alpha, X, &F77_incX, Y, &F77_incY);
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_ccopy.c b/ml/dlib/dlib/external/cblas/cblas_ccopy.c
new file mode 100644
index 000000000..b7bc42847
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ccopy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_ccopy.c
+ *
+ * The program is a C interface to ccopy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ccopy( const int N, const void *X,
+                      const int incX, void *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_ccopy( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cdotc_sub.c b/ml/dlib/dlib/external/cblas/cblas_cdotc_sub.c
new file mode 100644
index 000000000..d6086814e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cdotc_sub.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_cdotc_sub.c
+ *
+ * The program is a C interface to cdotc.
+ * It calls the fortran wrapper before calling cdotc.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cdotc_sub( const int N, const void *X, const int incX,
+                    const void *Y, const int incY,void *dotc)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_cdotc_sub( &F77_N, X, &F77_incX, Y, &F77_incY, dotc);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cdotu_sub.c b/ml/dlib/dlib/external/cblas/cblas_cdotu_sub.c
new file mode 100644
index 000000000..d06e4e5fa
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cdotu_sub.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_cdotu_sub.f
+ *
+ * The program is a C interface to cdotu.
+ * It calls the forteran wrapper before calling cdotu.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cdotu_sub( const int N, const void *X,
+                     const int incX, const void *Y, const int incY,void *dotu)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_cdotu_sub( &F77_N, X, &F77_incX, Y, &F77_incY, dotu);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cgbmv.c b/ml/dlib/dlib/external/cblas/cblas_cgbmv.c
new file mode 100644
index 000000000..94cc175f3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cgbmv.c
@@ -0,0 +1,154 @@
+/*
+ * cblas_cgbmv.c
+ * The program is a C interface of cgbmv
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+   F77_INT F77_KL=KL,F77_KU=KU;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_KL KL
+   #define F77_KU KU
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n=0, i=0, incx=incX;
+   const float *xx= (float *)X, *alp= (float *)alpha, *bet = (float *)beta;
+   float ALPHA[2],BETA[2];
+   int tincY, tincx;
+   float *x=(float *)X, *y=(float *)Y, *st=0, *tx=0;
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_cgbmv(F77_TA, &F77_M, &F77_N, &F77_KL, &F77_KU, alpha,  
+                     A, &F77_lda, X, &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         ALPHA[0]= *alp;
+         ALPHA[1]= -alp[1];
+         BETA[0]= *bet;
+         BETA[1]= -bet[1];
+         TA = 'N';
+         if (M > 0)
+         {
+            n = M << 1;
+            x = malloc(n*sizeof(float));
+            tx = x;
+
+            if( incX > 0 ) {
+               i = incX << 1 ;
+               tincx = 2;
+               st= x+n;
+            } else {
+               i = incX *(-2);
+               tincx = -2;
+               st = x-2;
+               x +=(n-2);
+            }
+            do
+            {
+               *x = *xx;
+               x[1] = -xx[1];
+               x += tincx ;
+               xx += i;
+            }
+            while (x != st);
+            x=tx;
+
+            #ifdef F77_INT
+               F77_incX = 1;
+            #else
+               incx = 1;
+            #endif
+
+            if( incY > 0 )
+              tincY = incY;
+            else
+              tincY = -incY;
+
+            y++;
+ 
+            if (N > 0)
+            {
+               i = tincY << 1;
+               n = i * N ;
+               st = y + n;
+               do {
+                  *y = -(*y);
+                  y += i;
+               } while(y != st);
+               y -= n;
+            }
+         }
+         else x = (float *) X;
+
+ 
+      }
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      if (TransA == CblasConjTrans)
+         F77_cgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, ALPHA, 
+                        A ,&F77_lda, x,&F77_incX, BETA, Y, &F77_incY);
+      else
+         F77_cgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, alpha, 
+                        A ,&F77_lda, x,&F77_incX, beta, Y, &F77_incY);
+      if (TransA == CblasConjTrans)
+      {
+         if (x != X) free(x);
+         if (N > 0)
+         {
+            do
+            {
+               *y = -(*y);
+               y += i;
+            }
+            while (y != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_cgbmv", "Illegal Order setting, %d\n", order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cgemm.c b/ml/dlib/dlib/external/cblas/cblas_cgemm.c
new file mode 100644
index 000000000..c11641023
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cgemm.c
@@ -0,0 +1,94 @@
+/*
+ *
+ * cblas_cgemm.c
+ * This program is a C interface to cgemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const void *alpha, const void  *A,
+                 const int lda, const void  *B, const int ldb,
+                 const void *beta, void  *C, const int ldc)
+{
+   char TA, TB;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_TB;
+#else
+   #define F77_TA &TA  
+   #define F77_TB &TB  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if(TransA == CblasTrans) TA='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgemm", "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if(TransB == CblasTrans) TB='T';
+      else if ( TransB == CblasConjTrans ) TB='C';
+      else if ( TransB == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cgemm", "Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_cgemm(F77_TA, F77_TB, &F77_M, &F77_N, &F77_K, alpha, A,
+                     &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if(TransA == CblasTrans) TB='T';
+      else if ( TransA == CblasConjTrans ) TB='C';
+      else if ( TransA == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgemm", "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if(TransB == CblasTrans) TA='T';
+      else if ( TransB == CblasConjTrans ) TA='C';
+      else if ( TransB == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgemm", "Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_cgemm(F77_TA, F77_TB, &F77_N, &F77_M, &F77_K, alpha, B,
+                  &F77_ldb, A, &F77_lda, beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_cgemm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cgemv.c b/ml/dlib/dlib/external/cblas/cblas_cgemv.c
new file mode 100644
index 000000000..a1cbb94ee
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cgemv.c
@@ -0,0 +1,151 @@
+/*
+ * cblas_cgemv.c
+ * The program is a C interface of cgemv
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+
+   int n=0, i=0, incx=incX;
+   const float *xx= (const float *)X;
+   float ALPHA[2],BETA[2];
+   int tincY, tincx;
+   float *x=(float *)X, *y=(float *)Y, *st=0, *tx=0;
+   const float *stx = x;
+
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_cgemv(F77_TA, &F77_M, &F77_N, alpha, A, &F77_lda, X, &F77_incX, 
+                beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+         
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         ALPHA[0]=    *( (const float *)  alpha    );
+         ALPHA[1]= -( *( (const float *)  alpha+1) );
+         BETA[0]=     *( (const float *)  beta     );
+         BETA[1]= -(  *( (const float *)  beta+1 ) );
+         TA = 'N';
+         if (M > 0)
+         {
+            n = M << 1;
+            x = malloc(n*sizeof(float));
+            tx = x;
+            if( incX > 0 ) {
+               i = incX << 1 ;
+               tincx = 2;
+               st= x+n;
+            } else { 
+               i = incX *(-2);
+               tincx = -2;
+               st = x-2; 
+               x +=(n-2); 
+            }
+
+            do
+            {
+               *x = *xx;
+               x[1] = -xx[1];
+               x += tincx ;
+               xx += i;
+            }
+            while (x != st);
+            x=tx;
+
+            F77_incX = 1;
+
+            if(incY > 0)
+               tincY = incY; 
+            else
+               tincY = -incY; 
+
+            y++;
+
+            if (N > 0)
+            {
+               i = tincY << 1;
+               n = i * N ;
+               st = y + n;
+               do {
+                  *y = -(*y);
+                  y += i;
+               } while(y != st); 
+               y -= n;
+            }
+            stx = x;
+         }
+         else stx = (const float *)X;
+      }
+      else 
+      {
+         cblas_xerbla(2, "cblas_cgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      if (TransA == CblasConjTrans)
+         F77_cgemv(F77_TA, &F77_N, &F77_M, ALPHA, A, &F77_lda, stx,
+                &F77_incX, BETA, Y, &F77_incY);
+      else
+         F77_cgemv(F77_TA, &F77_N, &F77_M, alpha, A, &F77_lda, x,
+                &F77_incX, beta, Y, &F77_incY);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (x != (const float *)X) free(x);
+         if (N > 0)
+         {
+            do
+            {
+               *y = -(*y);
+               y += i;
+            }
+            while (y != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_cgemv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cgerc.c b/ml/dlib/dlib/external/cblas/cblas_cgerc.c
new file mode 100644
index 000000000..e843f099b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cgerc.c
@@ -0,0 +1,77 @@
+/*
+ * cblas_cgerc.c
+ * The program is a C interface to cgerc.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cgerc(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incy
+   #define F77_lda lda   
+#endif
+
+   int n, i, tincy, incy=incY;
+   float *y=(float *)Y, *yy=(float *)Y, *ty, *st;
+
+
+   if (order == CblasColMajor)
+   {
+      F77_cgerc( &F77_M, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, A, 
+                      &F77_lda);
+   }  else if (order == CblasRowMajor)   
+   {
+      if (N > 0)
+      {
+         n = N << 1;
+         y = malloc(n*sizeof(float));
+
+         ty = y;
+         if( incY > 0 ) {
+            i = incY << 1;
+            tincy = 2;
+            st= y+n;
+         } else { 
+            i = incY *(-2);
+            tincy = -2;
+            st = y-2; 
+            y +=(n-2); 
+         }
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += tincy ;
+            yy += i;
+         }
+         while (y != st);
+         y = ty;
+
+         #ifdef F77_INT
+            F77_incY = 1;
+         #else
+            incy = 1;
+         #endif
+      }
+      else y = (float *) Y;
+
+      F77_cgeru( &F77_N, &F77_M, alpha, y, &F77_incY, X, &F77_incX, A, 
+                      &F77_lda);
+      if(Y!=y)
+         free(y);
+
+   } else cblas_xerbla(1, "cblas_cgerc", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cgeru.c b/ml/dlib/dlib/external/cblas/cblas_cgeru.c
new file mode 100644
index 000000000..4471d1f80
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cgeru.c
@@ -0,0 +1,38 @@
+/*
+ * cblas_cgeru.c
+ * The program is a C interface to cgeru.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cgeru(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda lda
+#endif
+
+
+
+   if (order == CblasColMajor)
+   {
+      F77_cgeru( &F77_M, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, A,
+                      &F77_lda);
+   }
+   else if (order == CblasRowMajor)
+   {
+      F77_cgeru( &F77_N, &F77_M, alpha, Y, &F77_incY, X, &F77_incX, A, 
+                      &F77_lda);
+   }
+   else cblas_xerbla(1, "cblas_cgeru","Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chbmv.c b/ml/dlib/dlib/external/cblas/cblas_chbmv.c
new file mode 100644
index 000000000..8681fa345
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chbmv.c
@@ -0,0 +1,145 @@
+/*
+ * cblas_chbmv.c
+ * The program is a C interface to chbmv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+#include <stdio.h>
+#include <stdlib.h>
+void cblas_chbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo,const int N,const int K,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const float *xx= (float *)X, *alp= (float *)alpha, *bet = (float *)beta;
+   float ALPHA[2],BETA[2];
+   int tincY, tincx;
+   float *x=(float *)X, *y=(float *)Y, *st=0, *tx;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chbmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_chbmv(F77_UL, &F77_N, &F77_K, alpha, A, &F77_lda, X,  
+                     &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1 ;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (float *) X; 
+
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_chbmv(F77_UL, &F77_N, &F77_K, ALPHA, 
+                     A ,&F77_lda, x,&F77_incX, BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_chbmv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor )
+   {
+      if(X!=x)
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+      }
+   }
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chemm.c b/ml/dlib/dlib/external/cblas/cblas_chemm.c
new file mode 100644
index 000000000..e64f0d0f3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chemm.c
@@ -0,0 +1,91 @@
+/*
+ *
+ * cblas_chemm.c
+ * This program is a C interface to chemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_chemm(const enum CBLAS_ORDER Order, const  enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chemm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_chemm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_chemm(F77_SD, F77_UL, &F77_M, &F77_N, alpha, A, &F77_lda, 
+                     B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chemm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_chemm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_chemm(F77_SD, F77_UL, &F77_N, &F77_M, alpha, A,
+                 &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_chemm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chemv.c b/ml/dlib/dlib/external/cblas/cblas_chemv.c
new file mode 100644
index 000000000..bbaaefdc2
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chemv.c
@@ -0,0 +1,146 @@
+/*
+ * cblas_chemv.c
+ * The program is a C interface to chemv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_chemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n=0, i=0, incx=incX;
+   const float *xx= (float *)X, *alp= (float *)alpha, *bet = (float *)beta;
+   float ALPHA[2],BETA[2];
+   int tincY, tincx;
+   float *x=(float *)X, *y=(float *)Y, *st=0, *tx;
+
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chemv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_chemv(F77_UL, &F77_N, alpha, A, &F77_lda, X, &F77_incX, 
+                beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1 ;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (float *) X;
+
+          
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chemv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_chemv(F77_UL, &F77_N, ALPHA, A, &F77_lda, x, &F77_incX, 
+                BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_chemv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor )
+   {
+      if ( X != x )
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+     }
+   }
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cher.c b/ml/dlib/dlib/external/cblas/cblas_cher.c
new file mode 100644
index 000000000..580413b02
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cher.c
@@ -0,0 +1,103 @@
+/*
+ * cblas_cher.c
+ * The program is a C interface to cher.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const void *X, const int incX
+                ,void *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+#endif
+   int n, i, tincx, incx=incX;
+   float *x=(float *)X, *xx=(float *)X, *tx, *st;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_cher(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+         tx = x;
+         if( incX > 0 ) {
+            i = incX << 1 ;
+            tincx = 2;
+            st= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            st = x-2; 
+            x +=(n-2); 
+         }
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+         #ifdef F77_INT
+           F77_incX = 1;
+         #else
+           incx = 1;
+         #endif
+      }
+      else x = (float *) X;
+      F77_cher(F77_UL, &F77_N, &alpha, x, &F77_incX, A, &F77_lda);
+   } else 
+   {
+      cblas_xerbla(1, "cblas_cher","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x) 
+      free(x);
+   
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cher2.c b/ml/dlib/dlib/external/cblas/cblas_cher2.c
new file mode 100644
index 000000000..89d36a0ee
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cher2.c
@@ -0,0 +1,139 @@
+/*
+ * cblas_cher2.c
+ * The program is a C interface to cher2.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incy
+#endif
+   int n, i, j, tincx, tincy, incx=incX, incy=incY;
+   float *x=(float *)X, *xx=(float *)X, *y=(float *)Y, 
+         *yy=(float *)Y, *tx, *ty, *stx, *sty;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_cher2(F77_UL, &F77_N, alpha, X, &F77_incX, 
+                                            Y, &F77_incY, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher2","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+         y = malloc(n*sizeof(float));         
+         tx = x;
+         ty = y;
+         if( incX > 0 ) {
+            i = incX << 1 ;
+            tincx = 2;
+            stx= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            stx = x-2; 
+            x +=(n-2); 
+         }
+         
+         if( incY > 0 ) {
+            j = incY << 1;
+            tincy = 2;
+            sty= y+n;
+         } else { 
+            j = incY *(-2);
+            tincy = -2;
+            sty = y-2; 
+            y +=(n-2); 
+         }
+
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != stx);
+
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += tincy ;
+            yy += j;
+         }
+         while (y != sty);
+
+         x=tx;
+         y=ty;
+
+         #ifdef F77_INT
+            F77_incX = 1;
+            F77_incY = 1;
+         #else
+            incx = 1;
+            incy = 1;
+         #endif
+      }  else 
+      {
+         x = (float *) X;
+         y = (float *) Y;
+      }
+      F77_cher2(F77_UL, &F77_N, alpha, y, &F77_incY, x, 
+                                      &F77_incX, A, &F77_lda);
+   } else 
+   {
+      cblas_xerbla(1, "cblas_cher2","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+      free(x);
+   if(Y!=y)
+      free(y);
+
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cher2k.c b/ml/dlib/dlib/external/cblas/cblas_cher2k.c
new file mode 100644
index 000000000..cad1c432e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cher2k.c
@@ -0,0 +1,96 @@
+/*
+ *
+ * cblas_cher2k.c
+ * This program is a C interface to cher2k.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const float beta,
+                  void *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+   float ALPHA[2]; 
+   const float *alp=(float *)alpha;
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cher2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_cher2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cher2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cher2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      F77_cher2k(F77_UL,F77_TR, &F77_N, &F77_K, ALPHA, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_cher2k", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cherk.c b/ml/dlib/dlib/external/cblas/cblas_cherk.c
new file mode 100644
index 000000000..0b6362db7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cherk.c
@@ -0,0 +1,90 @@
+/*
+ *
+ * cblas_cherk.c
+ * This program is a C interface to cherk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const float alpha, const void *A, const int lda,
+                 const float beta, void *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_cherk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cherk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_cherk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                     &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cherk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_cherk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_cherk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                &beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_cherk", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chpmv.c b/ml/dlib/dlib/external/cblas/cblas_chpmv.c
new file mode 100644
index 000000000..048734760
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chpmv.c
@@ -0,0 +1,146 @@
+/*
+ * cblas_chpmv.c
+ * The program is a C interface of chpmv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_chpmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo,const int N,
+                 const void *alpha, const void  *AP,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const float *xx= (float *)X, *alp= (float *)alpha, *bet = (float *)beta;
+   float ALPHA[2],BETA[2];
+   int tincY, tincx;
+   float *x=(float *)X, *y=(float *)Y, *st=0, *tx;
+
+   if (order == CblasColMajor)
+   { 
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_chpmv(F77_UL, &F77_N, alpha, AP, X,  
+                     &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (float *) X;
+
+
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpmv","Illegal Uplo setting, %d\n", Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_chpmv(F77_UL, &F77_N, ALPHA, 
+                     AP, x, &F77_incX, BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_chpmv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor ) 
+   {
+      if(X!=x)
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+     }
+  }
+
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chpr.c b/ml/dlib/dlib/external/cblas/cblas_chpr.c
new file mode 100644
index 000000000..72796b158
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chpr.c
@@ -0,0 +1,102 @@
+/*
+ * cblas_chpr.c
+ * The program is a C interface to chpr.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_chpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float alpha, const void *X,
+                const int incX, void *A)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incx
+#endif
+   int n, i, tincx, incx=incX;
+   float *x=(float *)X, *xx=(float *)X, *tx, *st;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_chpr(F77_UL, &F77_N, &alpha, X, &F77_incX, A);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpr","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+         tx = x;
+         if( incX > 0 ) {
+            i = incX << 1;
+            tincx = 2;
+            st= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            st = x-2; 
+            x +=(n-2); 
+         }
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != st);
+         x=tx;
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+      }
+      else x = (float *) X;
+
+      F77_chpr(F77_UL, &F77_N, &alpha, x, &F77_incX, A);
+
+   } else 
+   {
+      cblas_xerbla(1, "cblas_chpr","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+     free(x);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_chpr2.c b/ml/dlib/dlib/external/cblas/cblas_chpr2.c
new file mode 100644
index 000000000..f80d087aa
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_chpr2.c
@@ -0,0 +1,136 @@
+/*
+ * cblas_chpr2.c
+ * The program is a C interface to chpr2.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_chpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                      const int N,const void *alpha, const void *X, 
+                      const int incX,const void *Y, const int incY, void *Ap)
+
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N,  F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incx
+   #define F77_incY incy
+#endif
+   int n, i, j, tincx, tincy, incx=incX, incy=incY;
+   float *x=(float *)X, *xx=(float *)X, *y=(float *)Y,
+         *yy=(float *)Y, *tx, *ty, *stx, *sty;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_chpr2(F77_UL, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, Ap);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_chpr2","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(float));
+         y = malloc(n*sizeof(float));
+         tx = x;
+         ty = y;
+         if( incX > 0 ) {
+            i = incX << 1 ;
+            tincx = 2;
+            stx= x+n;
+         } else {
+            i = incX *(-2);
+            tincx = -2;
+            stx = x-2;
+            x +=(n-2);
+         }
+ 
+         if( incY > 0 ) {
+            j = incY << 1;
+            tincy = 2;
+            sty= y+n;
+         } else {
+            j = incY *(-2);
+            tincy = -2;
+            sty = y-2;
+            y +=(n-2);
+         }
+ 
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != stx);
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += tincy ;
+            yy += j;
+         }
+         while (y != sty);
+ 
+         x=tx;
+         y=ty;
+ 
+         #ifdef F77_INT
+            F77_incX = 1;
+            F77_incY = 1;
+         #else
+            incx = 1;
+            incy = 1;
+         #endif
+
+      }  else 
+      {
+         x = (float *) X;
+         y = (void  *) Y;
+      }
+      F77_chpr2(F77_UL, &F77_N, alpha, y, &F77_incY, x, &F77_incX, Ap);
+   } else 
+   {
+      cblas_xerbla(1, "cblas_chpr2","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+      free(x);
+   if(Y!=y)
+      free(y);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cscal.c b/ml/dlib/dlib/external/cblas/cblas_cscal.c
new file mode 100644
index 000000000..a23e6ee57
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_cscal.c
+ *
+ * The program is a C interface to cscal.f.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cscal( const int N, const void *alpha, void *X, 
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_cscal( &F77_N, alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_csscal.c b/ml/dlib/dlib/external/cblas/cblas_csscal.c
new file mode 100644
index 000000000..39983fe07
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_csscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_csscal.c
+ *
+ * The program is a C interface to csscal.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_csscal( const int N, const float alpha, void *X,
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_csscal( &F77_N, &alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_cswap.c b/ml/dlib/dlib/external/cblas/cblas_cswap.c
new file mode 100644
index 000000000..127282072
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_cswap.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_cswap.c
+ *
+ * The program is a C interface to cswap.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_cswap( const int N, void *X, const int incX, void *Y,
+                       const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_cswap( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_csymm.c b/ml/dlib/dlib/external/cblas/cblas_csymm.c
new file mode 100644
index 000000000..a462b5ebd
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_csymm.c
@@ -0,0 +1,91 @@
+/*
+ *
+ * cblas_csymm.c
+ * This program is a C interface to csymm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_csymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *B, const int ldb, const void *beta,
+                 void  *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_csymm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csymm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_csymm(F77_SD, F77_UL, &F77_M, &F77_N, alpha, A, &F77_lda,
+                      B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_csymm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csymm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_csymm(F77_SD, F77_UL, &F77_N, &F77_M, alpha, A, &F77_lda,
+                     B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_csymm", "Illegal Order setting, %d\n", Order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_csyr2k.c b/ml/dlib/dlib/external/cblas/cblas_csyr2k.c
new file mode 100644
index 000000000..c93facb2b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_csyr2k.c
@@ -0,0 +1,93 @@
+/*
+ *
+ * cblas_csyr2k.c
+ * This program is a C interface to csyr2k.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_csyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void  *A, const int lda,
+                  const void  *B, const int ldb, const void *beta,
+                  void  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_csyr2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyr2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_csyr2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                      B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyr2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyr2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_csyr2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_csyr2k", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_csyrk.c b/ml/dlib/dlib/external/cblas/cblas_csyrk.c
new file mode 100644
index 000000000..4ff0bd535
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_csyrk.c
@@ -0,0 +1,93 @@
+/*
+ *
+ * cblas_csyrk.c
+ * This program is a C interface to csyrk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_csyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const void *alpha, const void  *A, const int lda,
+                 const void *beta, void  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_csyrk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyrk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_csyrk(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyrk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_csyrk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_csyrk(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                     beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_csyrk", "Illegal Order setting, %d\n", Order);
+   return;
+}
+
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctbmv.c b/ml/dlib/dlib/external/cblas/cblas_ctbmv.c
new file mode 100644
index 000000000..0b313d858
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctbmv.c
@@ -0,0 +1,139 @@
+/*
+ * cblas_ctbmv.c
+ * The program is a C interface to ctbmv.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void  *A, const int lda,
+                 void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0, *x=(float *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctbmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            x++;
+            st = x + n;
+            do
+            {
+               *x = -(*x);
+               x+= i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctbsv.c b/ml/dlib/dlib/external/cblas/cblas_ctbsv.c
new file mode 100644
index 000000000..31f3f5bb0
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctbsv.c
@@ -0,0 +1,143 @@
+/*
+ * cblas_ctbsv.c
+ * The program is a C interface to ctbsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void  *A, const int lda,
+                 void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0,*x=(float *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+  
+            x++;
+
+            st=x+n;
+
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x+= i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctbsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctpmv.c b/ml/dlib/dlib/external/cblas/cblas_ctpmv.c
new file mode 100644
index 000000000..03dad131e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctpmv.c
@@ -0,0 +1,133 @@
+/*
+ * cblas_ctpmv.c
+ * The program is a C interface to ctpmv.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *Ap, void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0,*x=(float *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            x++;
+            st = x + n;
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctpmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctpsv.c b/ml/dlib/dlib/external/cblas/cblas_ctpsv.c
new file mode 100644
index 000000000..3023306bd
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctpsv.c
@@ -0,0 +1,138 @@
+/*
+ * cblas_ctpsv.c
+ * The program is a C interface to ctpsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *Ap, void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0, *x=(float*)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+ 
+            x++;
+
+            st=x+n;
+
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctpsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctrmm.c b/ml/dlib/dlib/external/cblas/cblas_ctrmm.c
new file mode 100644
index 000000000..ceeb2a5ff
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctrmm.c
@@ -0,0 +1,123 @@
+/*
+ *
+ * cblas_ctrmm.c
+ * This program is a C interface to ctrmm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const  enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 void  *B, const int ldb)
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight ) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrmm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper ) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrmm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans ) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrmm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else cblas_xerbla(5, "cblas_ctrmm", 
+                       "Illegal Diag setting, %d\n", Diag);
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, alpha, A, &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight ) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrmm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper ) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrmm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans ) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrmm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ctrmm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, alpha, A, &F77_lda, B, &F77_ldb);
+   } 
+   else  cblas_xerbla(1, "cblas_ctrmm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctrmv.c b/ml/dlib/dlib/external/cblas/cblas_ctrmv.c
new file mode 100644
index 000000000..dc590fe76
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctrmv.c
@@ -0,0 +1,136 @@
+/*
+ * cblas_ctrmv.c
+ * The program is a C interface to ctrmv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *A, const int lda,
+                 void  *X, const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0,*x=(float *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            st = x + n;
+            do
+            {
+               x[1] = -x[1];
+               x+= i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+         F77_ctrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               x[1] = -x[1];
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctrmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctrsm.c b/ml/dlib/dlib/external/cblas/cblas_ctrsm.c
new file mode 100644
index 000000000..ae2d8796e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctrsm.c
@@ -0,0 +1,132 @@
+/*
+ *
+ * cblas_ctrsm.c
+ * This program is a C interface to ctrsm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 void  *B, const int ldb)
+{
+   char UL, TA, SD, DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ctrsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ctrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, alpha, A,
+                &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ctrsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+
+      F77_ctrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, alpha, A,
+                &F77_lda, B, &F77_ldb);
+   } 
+   else cblas_xerbla(1, "cblas_ctrsm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ctrsv.c b/ml/dlib/dlib/external/cblas/cblas_ctrsv.c
new file mode 100644
index 000000000..8bfacd913
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ctrsv.c
@@ -0,0 +1,137 @@
+/*
+ * cblas_ctrsv.c
+ * The program is a C interface to ctrsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ctrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *A, const int lda, void  *X,
+                 const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   float *st=0,*x=(float *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ctrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+            x++;
+            st=x+n; 
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ctrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ctrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ctrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ctrsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dasum.c b/ml/dlib/dlib/external/cblas/cblas_dasum.c
new file mode 100644
index 000000000..1a3667f2d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dasum.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_dasum.c
+ *
+ * The program is a C interface to dasum.
+ * It calls the fortran wrapper before calling dasum.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double cblas_dasum( const int N, const double *X, const int incX) 
+{
+   double asum;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_dasum_sub( &F77_N, X, &F77_incX, &asum);
+   return asum;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_daxpy.c b/ml/dlib/dlib/external/cblas/cblas_daxpy.c
new file mode 100644
index 000000000..3678137fb
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_daxpy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_daxpy.c
+ *
+ * The program is a C interface to daxpy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_daxpy( const int N, const double alpha, const double *X,
+                       const int incX, double *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_daxpy( &F77_N, &alpha, X, &F77_incX, Y, &F77_incY);
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_dcopy.c b/ml/dlib/dlib/external/cblas/cblas_dcopy.c
new file mode 100644
index 000000000..422a55e51
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dcopy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_dcopy.c
+ *
+ * The program is a C interface to dcopy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dcopy( const int N, const double *X,
+                      const int incX, double *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_dcopy( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ddot.c b/ml/dlib/dlib/external/cblas/cblas_ddot.c
new file mode 100644
index 000000000..d77343403
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ddot.c
@@ -0,0 +1,25 @@
+/*
+ * cblas_ddot.c
+ *
+ * The program is a C interface to ddot.
+ * It calls the fortran wrapper before calling ddot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double cblas_ddot( const int N, const double *X,
+                      const int incX, const double *Y, const int incY)
+{
+   double dot;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_ddot_sub( &F77_N, X, &F77_incX, Y, &F77_incY, &dot);
+   return dot;
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_dgbmv.c b/ml/dlib/dlib/external/cblas/cblas_dgbmv.c
new file mode 100644
index 000000000..886dab740
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dgbmv.c
@@ -0,0 +1,70 @@
+/*
+ *
+ * cblas_dgbmv.c
+ * This program is a C interface to dgbmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU,
+                 const double alpha, const double  *A, const int lda,
+                 const double  *X, const int incX, const double beta,
+                 double  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+   F77_INT F77_KL=KL,F77_KU=KU;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_KL KL
+   #define F77_KU KU
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_dgbmv(F77_TA, &F77_M, &F77_N, &F77_KL, &F77_KU, &alpha,  
+                     A, &F77_lda, X, &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_dgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, &alpha, 
+                     A ,&F77_lda, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_dgbmv", "Illegal Order setting, %d\n", order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dgemm.c b/ml/dlib/dlib/external/cblas/cblas_dgemm.c
new file mode 100644
index 000000000..4fa9d8603
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dgemm.c
@@ -0,0 +1,94 @@
+/*
+ *
+ * cblas_dgemm.c
+ * This program is a C interface to dgemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const double alpha, const double  *A,
+                 const int lda, const double  *B, const int ldb,
+                 const double beta, double  *C, const int ldc)
+{
+   char TA, TB;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_TB;
+#else
+   #define F77_TA &TA  
+   #define F77_TB &TB  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if(TransA == CblasTrans) TA='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgemm","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if(TransB == CblasTrans) TB='T';
+      else if ( TransB == CblasConjTrans ) TB='C';
+      else if ( TransB == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dgemm","Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_dgemm(F77_TA, F77_TB, &F77_M, &F77_N, &F77_K, &alpha, A,
+       &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if(TransA == CblasTrans) TB='T';
+      else if ( TransA == CblasConjTrans ) TB='C';
+      else if ( TransA == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgemm","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if(TransB == CblasTrans) TA='T';
+      else if ( TransB == CblasConjTrans ) TA='C';
+      else if ( TransB == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgemm","Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_dgemm(F77_TA, F77_TB, &F77_N, &F77_M, &F77_K, &alpha, B,
+                  &F77_ldb, A, &F77_lda, &beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_dgemm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dgemv.c b/ml/dlib/dlib/external/cblas/cblas_dgemv.c
new file mode 100644
index 000000000..23a0f51e7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dgemv.c
@@ -0,0 +1,67 @@
+/*
+ *
+ * cblas_dgemv.c
+ * This program is a C interface to dgemv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const double alpha, const double  *A, const int lda,
+                 const double  *X, const int incX, const double beta,
+                 double  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_dgemv(F77_TA, &F77_M, &F77_N, &alpha, A, &F77_lda, X, &F77_incX, 
+                &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_dgemv(F77_TA, &F77_N, &F77_M, &alpha, A, &F77_lda, X,
+                &F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_dgemv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dger.c b/ml/dlib/dlib/external/cblas/cblas_dger.c
new file mode 100644
index 000000000..d021cc401
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dger.c
@@ -0,0 +1,40 @@
+/*
+ *
+ * cblas_dger.c
+ * This program is a C interface to dger.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dger(const enum CBLAS_ORDER order, const int M, const int N,
+                const double alpha, const double  *X, const int incX,
+                const double  *Y, const int incY, double  *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda lda
+#endif
+
+
+   if (order == CblasColMajor)
+   {
+      F77_dger( &F77_M, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A, 
+                      &F77_lda);
+   }
+   else if (order == CblasRowMajor)
+   {
+      F77_dger( &F77_N, &F77_M ,&alpha, Y, &F77_incY, X, &F77_incX, A, 
+                      &F77_lda);
+
+   }
+   else cblas_xerbla(1, "cblas_dger", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dnrm2.c b/ml/dlib/dlib/external/cblas/cblas_dnrm2.c
new file mode 100644
index 000000000..fe46ad484
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dnrm2.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_dnrm2.c
+ *
+ * The program is a C interface to dnrm2.
+ * It calls the fortranwrapper before calling dnrm2.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double cblas_dnrm2( const int N, const double *X, const int incX) 
+{
+   double nrm2;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_dnrm2_sub( &F77_N, X, &F77_incX, &nrm2);
+   return nrm2;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_drot.c b/ml/dlib/dlib/external/cblas/cblas_drot.c
new file mode 100644
index 000000000..51dc4ad5e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_drot.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_drot.c
+ *
+ * The program is a C interface to drot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_drot(const int N, double *X, const int incX,
+   double *Y, const int incY, const double c, const double s)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N 
+   #define F77_incX incX 
+   #define F77_incY incY 
+#endif
+   F77_drot(&F77_N, X, &F77_incX, Y, &F77_incY, &c, &s);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_drotg.c b/ml/dlib/dlib/external/cblas/cblas_drotg.c
new file mode 100644
index 000000000..0cbbd8bc0
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_drotg.c
@@ -0,0 +1,14 @@
+/*
+ * cblas_drotg.c
+ *
+ * The program is a C interface to drotg.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_drotg(  double *a, double *b, double *c, double *s)
+{
+   F77_drotg(a,b,c,s);    
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_drotm.c b/ml/dlib/dlib/external/cblas/cblas_drotm.c
new file mode 100644
index 000000000..ebe20ad62
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_drotm.c
@@ -0,0 +1,14 @@
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_drotm( const int N, double *X, const int incX, double *Y, 
+                       const int incY, const double *P)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_drotm( &F77_N, X, &F77_incX, Y, &F77_incY, P);
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_drotmg.c b/ml/dlib/dlib/external/cblas/cblas_drotmg.c
new file mode 100644
index 000000000..13a2208e5
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_drotmg.c
@@ -0,0 +1,15 @@
+/*
+ * cblas_drotmg.c
+ *
+ * The program is a C interface to drotmg.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_drotmg( double *d1, double *d2, double *b1, 
+                        const double b2, double *p)
+{
+   F77_drotmg(d1,d2,b1,&b2,p);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsbmv.c b/ml/dlib/dlib/external/cblas/cblas_dsbmv.c
new file mode 100644
index 000000000..c2f1a71c3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsbmv.c
@@ -0,0 +1,66 @@
+/*
+ *
+ * cblas_dsbmv.c
+ * This program is a C interface to dsbmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N, const int K,
+                 const double alpha, const double  *A, const int lda,
+                 const double  *X, const int incX, const double beta,
+                 double  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsbmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dsbmv(F77_UL, &F77_N, &F77_K, &alpha, A, &F77_lda, X,  
+                     &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dsbmv(F77_UL, &F77_N, &F77_K, &alpha, 
+                     A ,&F77_lda, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_dsbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dscal.c b/ml/dlib/dlib/external/cblas/cblas_dscal.c
new file mode 100644
index 000000000..bd04de77d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_dscal.c
+ *
+ * The program is a C interface to dscal.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dscal( const int N, const double alpha, double *X, 
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_dscal( &F77_N, &alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsdot.c b/ml/dlib/dlib/external/cblas/cblas_dsdot.c
new file mode 100644
index 000000000..52cd877a2
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsdot.c
@@ -0,0 +1,25 @@
+/*
+ * cblas_dsdot.c
+ *
+ * The program is a C interface to dsdot.
+ * It calls fthe fortran wrapper before calling dsdot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double  cblas_dsdot( const int N, const float *X,
+                      const int incX, const float *Y, const int incY)
+{
+   double dot;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_dsdot_sub( &F77_N, X, &F77_incX, Y, &F77_incY, &dot);
+   return dot;
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_dspmv.c b/ml/dlib/dlib/external/cblas/cblas_dspmv.c
new file mode 100644
index 000000000..ecdf081cb
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dspmv.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_dspmv.c
+ * This program is a C interface to dspmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+ 
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dspmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const double alpha, const double  *AP,
+                 const double  *X, const int incX, const double beta,
+                 double  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dspmv(F77_UL, &F77_N, &alpha, AP, X,  
+                     &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dspmv(F77_UL, &F77_N, &alpha, 
+                     AP, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_dspmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dspr.c b/ml/dlib/dlib/external/cblas/cblas_dspr.c
new file mode 100644
index 000000000..9e40cc11a
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dspr.c
@@ -0,0 +1,59 @@
+/*
+ *
+ * cblas_dspr.c
+ * This program is a C interface to dspr.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const double *X,
+                const int incX, double *Ap)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_dspr(F77_UL, &F77_N, &alpha, X, &F77_incX, Ap);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_dspr(F77_UL, &F77_N, &alpha, X, &F77_incX, Ap); 
+   } else cblas_xerbla(1, "cblas_dspr", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dspr2.c b/ml/dlib/dlib/external/cblas/cblas_dspr2.c
new file mode 100644
index 000000000..4ebbbbd52
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dspr2.c
@@ -0,0 +1,59 @@
+/*
+ * cblas_dspr2.c
+ * The program is a C interface to dspr2.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double  alpha, const double  *X,
+                const int incX, const double  *Y, const int incY, double  *A)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_dspr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dspr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_dspr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY,  A); 
+   } else cblas_xerbla(1, "cblas_dspr2", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dswap.c b/ml/dlib/dlib/external/cblas/cblas_dswap.c
new file mode 100644
index 000000000..9ae5bb93c
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dswap.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_dswap.c
+ *
+ * The program is a C interface to dswap.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dswap( const int N, double *X, const int incX, double *Y,
+                       const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_dswap( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsymm.c b/ml/dlib/dlib/external/cblas/cblas_dsymm.c
new file mode 100644
index 000000000..99b3858a1
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsymm.c
@@ -0,0 +1,91 @@
+/*
+ *
+ * cblas_dsymm.c
+ * This program is a C interface to dsymm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const double alpha, const double  *A, const int lda,
+                 const double  *B, const int ldb, const double beta,
+                 double  *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsymm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsymm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_dsymm(F77_SD, F77_UL, &F77_M, &F77_N, &alpha, A, &F77_lda,
+                      B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsymm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsymm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_dsymm(F77_SD, F77_UL, &F77_N, &F77_M, &alpha, A, &F77_lda, B,
+                 &F77_ldb, &beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_dsymm","Illegal Order setting, %d\n", Order); 
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsymv.c b/ml/dlib/dlib/external/cblas/cblas_dsymv.c
new file mode 100644
index 000000000..f0d32398a
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsymv.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_dsymv.c
+ * This program is a C interface to dsymv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsymv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const double alpha, const double  *A, const int lda,
+                 const double  *X, const int incX, const double beta,
+                 double  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsymv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dsymv(F77_UL, &F77_N, &alpha, A, &F77_lda, X,  
+                     &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsymv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_dsymv(F77_UL, &F77_N, &alpha, 
+                     A ,&F77_lda, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_dsymv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsyr.c b/ml/dlib/dlib/external/cblas/cblas_dsyr.c
new file mode 100644
index 000000000..d21b846e3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsyr.c
@@ -0,0 +1,60 @@
+/*
+ *
+ * cblas_dsyr.c
+ * This program is a C interface to dsyr.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double  alpha, const double  *X,
+                const int incX, double  *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_lda=lda;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_lda  lda
+#endif
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_dsyr(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_dsyr(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda); 
+   } else cblas_xerbla(1, "cblas_dsyr", "Illegal Order setting, %d\n", order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsyr2.c b/ml/dlib/dlib/external/cblas/cblas_dsyr2.c
new file mode 100644
index 000000000..7ce59657d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsyr2.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_dsyr2.c
+ * This program is a C interface to dsyr2.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double  alpha, const double  *X,
+                const int incX, const double  *Y, const int incY, double  *A,
+                const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY, F77_lda=lda;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda  lda
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_dsyr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A, 
+                    &F77_lda);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_dsyr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY,  A, 
+                    &F77_lda); 
+   } else cblas_xerbla(1, "cblas_dsyr2", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsyr2k.c b/ml/dlib/dlib/external/cblas/cblas_dsyr2k.c
new file mode 100644
index 000000000..dc11e9549
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsyr2k.c
@@ -0,0 +1,94 @@
+/*
+ *
+ * cblas_dsyr2k.c
+ * This program is a C interface to dsyr2k.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const double alpha, const double  *A, const int lda,
+                  const double  *B, const int ldb, const double beta,
+                  double  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyr2k","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyr2k","Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_dsyr2k(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                      B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyr2k","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyr2k","Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_dsyr2k(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda, B, 
+                &F77_ldb, &beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_dsyr2k","Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dsyrk.c b/ml/dlib/dlib/external/cblas/cblas_dsyrk.c
new file mode 100644
index 000000000..7ee834ea6
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dsyrk.c
@@ -0,0 +1,93 @@
+/*
+ *
+ * cblas_dsyrk.c
+ * This program is a C interface to dsyrk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const double alpha, const double  *A, const int lda,
+                 const double beta, double  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dsyrk","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyrk","Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_dsyrk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyrk","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dsyrk","Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_dsyrk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                     &beta, C, &F77_ldc);
+   } 
+   else cblas_xerbla(1, "cblas_dsyrk","Illegal Order setting, %d\n", Order);
+   return;
+}
+
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtbmv.c b/ml/dlib/dlib/external/cblas/cblas_dtbmv.c
new file mode 100644
index 000000000..1a06d1886
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtbmv.c
@@ -0,0 +1,103 @@
+/*
+ * cblas_dtbmv.c
+ * The program is a C interface to dtbmv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const double  *A, const int lda,
+                 double  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtbmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+
+   }
+   else cblas_xerbla(1, "cblas_dtbmv", "Illegal Order setting, %d\n", order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtbsv.c b/ml/dlib/dlib/external/cblas/cblas_dtbsv.c
new file mode 100644
index 000000000..aaf4a8d4b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtbsv.c
@@ -0,0 +1,103 @@
+/*
+ * cblas_dtbsv.c
+ * The program is a C interface to dtbsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const double  *A, const int lda,
+                 double  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_dtbsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtpmv.c b/ml/dlib/dlib/external/cblas/cblas_dtpmv.c
new file mode 100644
index 000000000..565f97a68
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtpmv.c
@@ -0,0 +1,98 @@
+/*
+ * cblas_dtpmv.c
+ * The program is a C interface to dtpmv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double  *Ap, double  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_dtpmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtpsv.c b/ml/dlib/dlib/external/cblas/cblas_dtpsv.c
new file mode 100644
index 000000000..4f51ccc56
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtpsv.c
@@ -0,0 +1,99 @@
+/*
+ * cblas_dtpsv.c
+ * The program is a C interface to dtpsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double  *Ap, double  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+
+   }
+   else cblas_xerbla(1, "cblas_dtpsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtrmm.c b/ml/dlib/dlib/external/cblas/cblas_dtrmm.c
new file mode 100644
index 000000000..0f4c0d161
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtrmm.c
@@ -0,0 +1,125 @@
+/*
+ *
+ * cblas_dtrmm.c
+ * This program is a C interface to dtrmm.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const  enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const double alpha, const double  *A, const int lda,
+                 double  *B, const int ldb)
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrmm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrmm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrmm","Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_dtrmm","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, &alpha, A, &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrmm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrmm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrmm","Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_dtrmm","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, &alpha, A, &F77_lda, B, &F77_ldb);
+   } 
+   else cblas_xerbla(1, "cblas_dtrmm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtrmv.c b/ml/dlib/dlib/external/cblas/cblas_dtrmv.c
new file mode 100644
index 000000000..c20ea0626
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtrmv.c
@@ -0,0 +1,103 @@
+/*
+ *
+ * cblas_dtrmv.c
+ * This program is a C interface to sgemv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+ 
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double  *A, const int lda,
+                 double  *X, const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   } else cblas_xerbla(1, "cblas_dtrmv", "Illegal order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtrsm.c b/ml/dlib/dlib/external/cblas/cblas_dtrsm.c
new file mode 100644
index 000000000..986425f60
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtrsm.c
@@ -0,0 +1,130 @@
+/*
+ *
+ * cblas_dtrsm.c
+ * This program is a C interface to dtrsm.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const double alpha, const double  *A, const int lda,
+                 double  *B, const int ldb)
+
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if      ( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrsm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if      ( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrsm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if      ( TransA == CblasTrans    ) TA='T';
+      else if ( TransA == CblasConjTrans) TA='C';
+      else if ( TransA == CblasNoTrans  ) TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrsm","Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if      ( Diag == CblasUnit   ) DI='U';
+      else if ( Diag == CblasNonUnit) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_dtrsm","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, &alpha,
+                A, &F77_lda, B, &F77_ldb);
+   } 
+   else if (Order == CblasRowMajor)
+   {
+      if      ( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrsm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if      ( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrsm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if      ( TransA == CblasTrans    ) TA='T';
+      else if ( TransA == CblasConjTrans) TA='C';
+      else if ( TransA == CblasNoTrans  ) TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrsm","Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if      ( Diag == CblasUnit   ) DI='U';
+      else if ( Diag == CblasNonUnit) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_dtrsm","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_dtrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, &alpha, A, 
+               &F77_lda, B, &F77_ldb);
+   } 
+   else cblas_xerbla(1, "cblas_dtrsm","Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dtrsv.c b/ml/dlib/dlib/external/cblas/cblas_dtrsv.c
new file mode 100644
index 000000000..5c4ed5637
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dtrsv.c
@@ -0,0 +1,102 @@
+/*
+ * cblas_dtrsv.c
+ * The program is a C interface to dtrsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_dtrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const double  *A, const int lda, double  *X,
+                 const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_dtrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_dtrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_dtrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_dtrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_dtrsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dzasum.c b/ml/dlib/dlib/external/cblas/cblas_dzasum.c
new file mode 100644
index 000000000..b32f573e5
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dzasum.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_dzasum.c
+ *
+ * The program is a C interface to dzasum.
+ * It calls the fortran wrapper before calling dzasum.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double cblas_dzasum( const int N, const void *X, const int incX) 
+{
+   double asum;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_dzasum_sub( &F77_N, X, &F77_incX, &asum);
+   return asum;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_dznrm2.c b/ml/dlib/dlib/external/cblas/cblas_dznrm2.c
new file mode 100644
index 000000000..dfa2bfc83
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_dznrm2.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_dznrm2.c
+ *
+ * The program is a C interface to dznrm2.
+ * It calls the fortran wrapper before calling dznrm2.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+double cblas_dznrm2( const int N, const void *X, const int incX) 
+{
+   double nrm2;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_dznrm2_sub( &F77_N, X, &F77_incX, &nrm2);
+   return nrm2;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_f77.h b/ml/dlib/dlib/external/cblas/cblas_f77.h
new file mode 100644
index 000000000..18435cd30
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_f77.h
@@ -0,0 +1,701 @@
+/*
+ * cblas_f77.h
+ * Written by Keita Teranishi
+ *
+ * Updated by Jeff Horner
+ * Merged cblas_f77.h and cblas_fortran_header.h
+ */
+
+#ifndef CBLAS_F77_H
+#define CBLAS_f77_H
+
+#ifdef CRAY
+   #include <fortran.h>
+   #define F77_CHAR _fcd
+   #define C2F_CHAR(a) ( _cptofcd( (a), 1 ) )
+   #define C2F_STR(a, i) ( _cptofcd( (a), (i) ) )
+   #define F77_STRLEN(a) (_fcdlen)
+#endif
+
+#ifdef WeirdNEC
+   #define F77_INT long
+#endif
+
+#ifdef  F77_CHAR
+   #define FCHAR F77_CHAR
+#else
+   #define FCHAR char *
+#endif
+
+#ifdef F77_INT
+   #define FINT const F77_INT *
+   #define FINT2 F77_INT *
+#else
+   #define FINT const int *
+   #define FINT2 int *
+#endif
+
+#if defined(ADD_)
+/*
+ * Level 1 BLAS
+ */
+#define F77_xerbla xerbla_
+   #define F77_srotg      srotg_
+   #define F77_srotmg     srotmg_
+   #define F77_srot       srot_
+   #define F77_srotm      srotm_
+   #define F77_drotg      drotg_
+   #define F77_drotmg     drotmg_
+   #define F77_drot       drot_
+   #define F77_drotm      drotm_
+   #define F77_sswap      sswap_
+   #define F77_scopy      scopy_
+   #define F77_saxpy      saxpy_
+   #define F77_isamax_sub isamaxsub_
+   #define F77_dswap      dswap_
+   #define F77_dcopy      dcopy_
+   #define F77_daxpy      daxpy_
+   #define F77_idamax_sub idamaxsub_
+   #define F77_cswap      cswap_
+   #define F77_ccopy      ccopy_
+   #define F77_caxpy      caxpy_
+   #define F77_icamax_sub icamaxsub_
+   #define F77_zswap      zswap_
+   #define F77_zcopy      zcopy_
+   #define F77_zaxpy      zaxpy_
+   #define F77_izamax_sub izamaxsub_
+   #define F77_sdot_sub   sdotsub_
+   #define F77_ddot_sub   ddotsub_
+   #define F77_dsdot_sub   dsdotsub_
+   #define F77_sscal      sscal_
+   #define F77_dscal      dscal_
+   #define F77_cscal      cscal_
+   #define F77_zscal      zscal_
+   #define F77_csscal      csscal_
+   #define F77_zdscal      zdscal_
+   #define F77_cdotu_sub  cdotusub_
+   #define F77_cdotc_sub  cdotcsub_
+   #define F77_zdotu_sub  zdotusub_
+   #define F77_zdotc_sub  zdotcsub_
+   #define F77_snrm2_sub  snrm2sub_
+   #define F77_sasum_sub  sasumsub_
+   #define F77_dnrm2_sub  dnrm2sub_
+   #define F77_dasum_sub  dasumsub_
+   #define F77_scnrm2_sub  scnrm2sub_
+   #define F77_scasum_sub  scasumsub_
+   #define F77_dznrm2_sub  dznrm2sub_
+   #define F77_dzasum_sub  dzasumsub_
+   #define F77_sdsdot_sub   sdsdotsub_
+/*
+ * Level 2 BLAS
+ */
+   #define F77_ssymv      ssymv_
+   #define F77_ssbmv      ssbmv_
+   #define F77_sspmv      sspmv_
+   #define F77_sger       sger_
+   #define F77_ssyr       ssyr_
+   #define F77_sspr       sspr_
+   #define F77_ssyr2      ssyr2_
+   #define F77_sspr2      sspr2_
+   #define F77_dsymv      dsymv_
+   #define F77_dsbmv      dsbmv_
+   #define F77_dspmv      dspmv_
+   #define F77_dger       dger_
+   #define F77_dsyr       dsyr_
+   #define F77_dspr       dspr_
+   #define F77_dsyr2      dsyr2_
+   #define F77_dspr2      dspr2_
+   #define F77_chemv      chemv_
+   #define F77_chbmv      chbmv_
+   #define F77_chpmv      chpmv_
+   #define F77_cgeru      cgeru_
+   #define F77_cgerc      cgerc_
+   #define F77_cher       cher_
+   #define F77_chpr       chpr_
+   #define F77_cher2      cher2_
+   #define F77_chpr2      chpr2_
+   #define F77_zhemv      zhemv_
+   #define F77_zhbmv      zhbmv_
+   #define F77_zhpmv      zhpmv_
+   #define F77_zgeru      zgeru_
+   #define F77_zgerc      zgerc_
+   #define F77_zher       zher_
+   #define F77_zhpr       zhpr_
+   #define F77_zher2      zher2_
+   #define F77_zhpr2      zhpr2_
+   #define F77_sgemv      sgemv_
+   #define F77_sgbmv      sgbmv_
+   #define F77_strmv      strmv_
+   #define F77_stbmv      stbmv_
+   #define F77_stpmv      stpmv_
+   #define F77_strsv      strsv_
+   #define F77_stbsv      stbsv_
+   #define F77_stpsv      stpsv_
+   #define F77_dgemv      dgemv_
+   #define F77_dgbmv      dgbmv_
+   #define F77_dtrmv      dtrmv_
+   #define F77_dtbmv      dtbmv_
+   #define F77_dtpmv      dtpmv_
+   #define F77_dtrsv      dtrsv_
+   #define F77_dtbsv      dtbsv_
+   #define F77_dtpsv      dtpsv_
+   #define F77_cgemv      cgemv_
+   #define F77_cgbmv      cgbmv_
+   #define F77_ctrmv      ctrmv_
+   #define F77_ctbmv      ctbmv_
+   #define F77_ctpmv      ctpmv_
+   #define F77_ctrsv      ctrsv_
+   #define F77_ctbsv      ctbsv_
+   #define F77_ctpsv      ctpsv_
+   #define F77_zgemv      zgemv_
+   #define F77_zgbmv      zgbmv_
+   #define F77_ztrmv      ztrmv_
+   #define F77_ztbmv      ztbmv_
+   #define F77_ztpmv      ztpmv_
+   #define F77_ztrsv      ztrsv_
+   #define F77_ztbsv      ztbsv_
+   #define F77_ztpsv      ztpsv_
+/*
+ * Level 3 BLAS
+ */
+   #define F77_chemm      chemm_
+   #define F77_cherk      cherk_
+   #define F77_cher2k     cher2k_
+   #define F77_zhemm      zhemm_
+   #define F77_zherk      zherk_
+   #define F77_zher2k     zher2k_
+   #define F77_sgemm      sgemm_
+   #define F77_ssymm      ssymm_
+   #define F77_ssyrk      ssyrk_
+   #define F77_ssyr2k     ssyr2k_
+   #define F77_strmm      strmm_
+   #define F77_strsm      strsm_
+   #define F77_dgemm      dgemm_
+   #define F77_dsymm      dsymm_
+   #define F77_dsyrk      dsyrk_
+   #define F77_dsyr2k     dsyr2k_
+   #define F77_dtrmm      dtrmm_
+   #define F77_dtrsm      dtrsm_
+   #define F77_cgemm      cgemm_
+   #define F77_csymm      csymm_
+   #define F77_csyrk      csyrk_
+   #define F77_csyr2k     csyr2k_
+   #define F77_ctrmm      ctrmm_
+   #define F77_ctrsm      ctrsm_
+   #define F77_zgemm      zgemm_
+   #define F77_zsymm      zsymm_
+   #define F77_zsyrk      zsyrk_
+   #define F77_zsyr2k     zsyr2k_
+   #define F77_ztrmm      ztrmm_
+   #define F77_ztrsm      ztrsm_
+#elif defined(UPCASE)
+/*
+ * Level 1 BLAS
+ */
+#define F77_xerbla  XERBLA
+   #define F77_srotg      SROTG
+   #define F77_srotmg     SROTMG
+   #define F77_srot       SROT
+   #define F77_srotm      SROTM
+   #define F77_drotg      DROTG
+   #define F77_drotmg     DROTMG
+   #define F77_drot       DROT
+   #define F77_drotm      DROTM
+   #define F77_sswap      SSWAP
+   #define F77_scopy      SCOPY
+   #define F77_saxpy      SAXPY
+   #define F77_isamax_sub ISAMAXSUB
+   #define F77_dswap      DSWAP
+   #define F77_dcopy      DCOPY
+   #define F77_daxpy      DAXPY
+   #define F77_idamax_sub IDAMAXSUB
+   #define F77_cswap      CSWAP
+   #define F77_ccopy      CCOPY
+   #define F77_caxpy      CAXPY
+   #define F77_icamax_sub ICAMAXSUB
+   #define F77_zswap      ZSWAP
+   #define F77_zcopy      ZCOPY
+   #define F77_zaxpy      ZAXPY
+   #define F77_izamax_sub IZAMAXSUB
+   #define F77_sdot_sub   SDOTSUB
+   #define F77_ddot_sub   DDOTSUB
+   #define F77_dsdot_sub   DSDOTSUB
+   #define F77_sscal      SSCAL
+   #define F77_dscal      DSCAL
+   #define F77_cscal      CSCAL
+   #define F77_zscal      ZSCAL
+   #define F77_csscal      CSSCAL
+   #define F77_zdscal      ZDSCAL
+   #define F77_cdotu_sub  CDOTUSUB
+   #define F77_cdotc_sub  CDOTCSUB
+   #define F77_zdotu_sub  ZDOTUSUB
+   #define F77_zdotc_sub  ZDOTCSUB
+   #define F77_snrm2_sub  SNRM2SUB
+   #define F77_sasum_sub  SASUMSUB
+   #define F77_dnrm2_sub  DNRM2SUB
+   #define F77_dasum_sub  DASUMSUB
+   #define F77_scnrm2_sub  SCNRM2SUB
+   #define F77_scasum_sub  SCASUMSUB
+   #define F77_dznrm2_sub  DZNRM2SUB
+   #define F77_dzasum_sub  DZASUMSUB
+   #define F77_sdsdot_sub   SDSDOTSUB
+/*
+ * Level 2 BLAS
+ */
+   #define F77_ssymv      SSYMV
+   #define F77_ssbmv      SSBMV
+   #define F77_sspmv      SSPMV
+   #define F77_sger       SGER
+   #define F77_ssyr       SSYR
+   #define F77_sspr       SSPR
+   #define F77_ssyr2      SSYR2
+   #define F77_sspr2      SSPR2
+   #define F77_dsymv      DSYMV
+   #define F77_dsbmv      DSBMV
+   #define F77_dspmv      DSPMV
+   #define F77_dger       DGER
+   #define F77_dsyr       DSYR
+   #define F77_dspr       DSPR
+   #define F77_dsyr2      DSYR2
+   #define F77_dspr2      DSPR2
+   #define F77_chemv      CHEMV
+   #define F77_chbmv      CHBMV
+   #define F77_chpmv      CHPMV
+   #define F77_cgeru      CGERU
+   #define F77_cgerc      CGERC
+   #define F77_cher       CHER
+   #define F77_chpr       CHPR
+   #define F77_cher2      CHER2
+   #define F77_chpr2      CHPR2
+   #define F77_zhemv      ZHEMV
+   #define F77_zhbmv      ZHBMV
+   #define F77_zhpmv      ZHPMV
+   #define F77_zgeru      ZGERU
+   #define F77_zgerc      ZGERC
+   #define F77_zher       ZHER
+   #define F77_zhpr       ZHPR
+   #define F77_zher2      ZHER2
+   #define F77_zhpr2      ZHPR2
+   #define F77_sgemv      SGEMV
+   #define F77_sgbmv      SGBMV
+   #define F77_strmv      STRMV
+   #define F77_stbmv      STBMV
+   #define F77_stpmv      STPMV
+   #define F77_strsv      STRSV
+   #define F77_stbsv      STBSV
+   #define F77_stpsv      STPSV
+   #define F77_dgemv      DGEMV
+   #define F77_dgbmv      DGBMV
+   #define F77_dtrmv      DTRMV
+   #define F77_dtbmv      DTBMV
+   #define F77_dtpmv      DTPMV
+   #define F77_dtrsv      DTRSV
+   #define F77_dtbsv      DTBSV
+   #define F77_dtpsv      DTPSV
+   #define F77_cgemv      CGEMV
+   #define F77_cgbmv      CGBMV
+   #define F77_ctrmv      CTRMV
+   #define F77_ctbmv      CTBMV
+   #define F77_ctpmv      CTPMV
+   #define F77_ctrsv      CTRSV
+   #define F77_ctbsv      CTBSV
+   #define F77_ctpsv      CTPSV
+   #define F77_zgemv      ZGEMV
+   #define F77_zgbmv      ZGBMV
+   #define F77_ztrmv      ZTRMV
+   #define F77_ztbmv      ZTBMV
+   #define F77_ztpmv      ZTPMV
+   #define F77_ztrsv      ZTRSV
+   #define F77_ztbsv      ZTBSV
+   #define F77_ztpsv      ZTPSV
+/*
+ * Level 3 BLAS
+ */
+   #define F77_chemm      CHEMM
+   #define F77_cherk      CHERK
+   #define F77_cher2k     CHER2K
+   #define F77_zhemm      ZHEMM
+   #define F77_zherk      ZHERK
+   #define F77_zher2k     ZHER2K
+   #define F77_sgemm      SGEMM
+   #define F77_ssymm      SSYMM
+   #define F77_ssyrk      SSYRK
+   #define F77_ssyr2k     SSYR2K
+   #define F77_strmm      STRMM
+   #define F77_strsm      STRSM
+   #define F77_dgemm      DGEMM
+   #define F77_dsymm      DSYMM
+   #define F77_dsyrk      DSYRK
+   #define F77_dsyr2k     DSYR2K
+   #define F77_dtrmm      DTRMM
+   #define F77_dtrsm      DTRSM
+   #define F77_cgemm      CGEMM
+   #define F77_csymm      CSYMM
+   #define F77_csyrk      CSYRK
+   #define F77_csyr2k     CSYR2K
+   #define F77_ctrmm      CTRMM
+   #define F77_ctrsm      CTRSM
+   #define F77_zgemm      ZGEMM
+   #define F77_zsymm      ZSYMM
+   #define F77_zsyrk      ZSYRK
+   #define F77_zsyr2k     ZSYR2K
+   #define F77_ztrmm      ZTRMM
+   #define F77_ztrsm      ZTRSM
+#elif defined(NOCHANGE)
+/*
+ * Level 1 BLAS
+ */
+#define F77_xerbla  xerbla
+   #define F77_srotg      srotg
+   #define F77_srotmg     srotmg
+   #define F77_srot       srot
+   #define F77_srotm      srotm
+   #define F77_drotg      drotg
+   #define F77_drotmg     drotmg
+   #define F77_drot       drot
+   #define F77_drotm      drotm
+   #define F77_sswap      sswap
+   #define F77_scopy      scopy
+   #define F77_saxpy      saxpy
+   #define F77_isamax_sub isamaxsub
+   #define F77_dswap      dswap
+   #define F77_dcopy      dcopy
+   #define F77_daxpy      daxpy
+   #define F77_idamax_sub idamaxsub
+   #define F77_cswap      cswap
+   #define F77_ccopy      ccopy
+   #define F77_caxpy      caxpy
+   #define F77_icamax_sub icamaxsub
+   #define F77_zswap      zswap
+   #define F77_zcopy      zcopy
+   #define F77_zaxpy      zaxpy
+   #define F77_izamax_sub izamaxsub
+   #define F77_sdot_sub   sdotsub
+   #define F77_ddot_sub   ddotsub
+   #define F77_dsdot_sub   dsdotsub
+   #define F77_sscal      sscal
+   #define F77_dscal      dscal
+   #define F77_cscal      cscal
+   #define F77_zscal      zscal
+   #define F77_csscal      csscal
+   #define F77_zdscal      zdscal
+   #define F77_cdotu_sub  cdotusub
+   #define F77_cdotc_sub  cdotcsub
+   #define F77_zdotu_sub  zdotusub
+   #define F77_zdotc_sub  zdotcsub
+   #define F77_snrm2_sub  snrm2sub
+   #define F77_sasum_sub  sasumsub
+   #define F77_dnrm2_sub  dnrm2sub
+   #define F77_dasum_sub  dasumsub
+   #define F77_scnrm2_sub  scnrm2sub
+   #define F77_scasum_sub  scasumsub
+   #define F77_dznrm2_sub  dznrm2sub
+   #define F77_dzasum_sub  dzasumsub
+   #define F77_sdsdot_sub   sdsdotsub
+/*
+ * Level 2 BLAS
+ */
+   #define F77_ssymv      ssymv
+   #define F77_ssbmv      ssbmv
+   #define F77_sspmv      sspmv
+   #define F77_sger       sger
+   #define F77_ssyr       ssyr
+   #define F77_sspr       sspr
+   #define F77_ssyr2      ssyr2
+   #define F77_sspr2      sspr2
+   #define F77_dsymv      dsymv
+   #define F77_dsbmv      dsbmv
+   #define F77_dspmv      dspmv
+   #define F77_dger       dger
+   #define F77_dsyr       dsyr
+   #define F77_dspr       dspr
+   #define F77_dsyr2      dsyr2
+   #define F77_dspr2      dspr2
+   #define F77_chemv      chemv
+   #define F77_chbmv      chbmv
+   #define F77_chpmv      chpmv
+   #define F77_cgeru      cgeru
+   #define F77_cgerc      cgerc
+   #define F77_cher       cher
+   #define F77_chpr       chpr
+   #define F77_cher2      cher2
+   #define F77_chpr2      chpr2
+   #define F77_zhemv      zhemv
+   #define F77_zhbmv      zhbmv
+   #define F77_zhpmv      zhpmv
+   #define F77_zgeru      zgeru
+   #define F77_zgerc      zgerc
+   #define F77_zher       zher
+   #define F77_zhpr       zhpr
+   #define F77_zher2      zher2
+   #define F77_zhpr2      zhpr2
+   #define F77_sgemv      sgemv
+   #define F77_sgbmv      sgbmv
+   #define F77_strmv      strmv
+   #define F77_stbmv      stbmv
+   #define F77_stpmv      stpmv
+   #define F77_strsv      strsv
+   #define F77_stbsv      stbsv
+   #define F77_stpsv      stpsv
+   #define F77_dgemv      dgemv
+   #define F77_dgbmv      dgbmv
+   #define F77_dtrmv      dtrmv
+   #define F77_dtbmv      dtbmv
+   #define F77_dtpmv      dtpmv
+   #define F77_dtrsv      dtrsv
+   #define F77_dtbsv      dtbsv
+   #define F77_dtpsv      dtpsv
+   #define F77_cgemv      cgemv
+   #define F77_cgbmv      cgbmv
+   #define F77_ctrmv      ctrmv
+   #define F77_ctbmv      ctbmv
+   #define F77_ctpmv      ctpmv
+   #define F77_ctrsv      ctrsv
+   #define F77_ctbsv      ctbsv
+   #define F77_ctpsv      ctpsv
+   #define F77_zgemv      zgemv
+   #define F77_zgbmv      zgbmv
+   #define F77_ztrmv      ztrmv
+   #define F77_ztbmv      ztbmv
+   #define F77_ztpmv      ztpmv
+   #define F77_ztrsv      ztrsv
+   #define F77_ztbsv      ztbsv
+   #define F77_ztpsv      ztpsv
+/*
+ * Level 3 BLAS
+ */
+   #define F77_chemm      chemm
+   #define F77_cherk      cherk
+   #define F77_cher2k     cher2k
+   #define F77_zhemm      zhemm
+   #define F77_zherk      zherk
+   #define F77_zher2k     zher2k
+   #define F77_sgemm      sgemm
+   #define F77_ssymm      ssymm
+   #define F77_ssyrk      ssyrk
+   #define F77_ssyr2k     ssyr2k
+   #define F77_strmm      strmm
+   #define F77_strsm      strsm
+   #define F77_dgemm      dgemm
+   #define F77_dsymm      dsymm
+   #define F77_dsyrk      dsyrk
+   #define F77_dsyr2k     dsyr2k
+   #define F77_dtrmm      dtrmm
+   #define F77_dtrsm      dtrsm
+   #define F77_cgemm      cgemm
+   #define F77_csymm      csymm
+   #define F77_csyrk      csyrk
+   #define F77_csyr2k     csyr2k
+   #define F77_ctrmm      ctrmm
+   #define F77_ctrsm      ctrsm
+   #define F77_zgemm      zgemm
+   #define F77_zsymm      zsymm
+   #define F77_zsyrk      zsyrk
+   #define F77_zsyr2k     zsyr2k
+   #define F77_ztrmm      ztrmm
+   #define F77_ztrsm      ztrsm
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+   void F77_xerbla(FCHAR, void *);
+/*
+ * Level 1 Fortran Prototypes
+ */
+
+/* Single Precision */
+
+   void F77_srot(FINT, float *, FINT, float *, FINT, const float *, const float *);
+   void F77_srotg(float *,float *,float *,float *);    
+   void F77_srotm( FINT, float *, FINT, float *, FINT, const float *);
+   void F77_srotmg(float *,float *,float *,const float *, float *);
+   void F77_sswap( FINT, float *, FINT, float *, FINT);
+   void F77_scopy( FINT, const float *, FINT, float *, FINT);
+   void F77_saxpy( FINT, const float *, const float *, FINT, float *, FINT);
+   void F77_sdot_sub(FINT, const float *, FINT, const float *, FINT, float *);
+   void F77_sdsdot_sub( FINT, const float *, const float *, FINT, const float *, FINT, float *);
+   void F77_sscal( FINT, const float *, float *, FINT);
+   void F77_snrm2_sub( FINT, const float *, FINT, float *);
+   void F77_sasum_sub( FINT, const float *, FINT, float *);
+   void F77_isamax_sub( FINT, const float * , FINT, FINT2);
+
+/* Double Precision */
+
+   void F77_drot(FINT, double *, FINT, double *, FINT, const double *, const double *);
+   void F77_drotg(double *,double *,double *,double *);    
+   void F77_drotm( FINT, double *, FINT, double *, FINT, const double *);
+   void F77_drotmg(double *,double *,double *,const double *, double *);
+   void F77_dswap( FINT, double *, FINT, double *, FINT);
+   void F77_dcopy( FINT, const double *, FINT, double *, FINT);
+   void F77_daxpy( FINT, const double *, const double *, FINT, double *, FINT);
+   void F77_dswap( FINT, double *, FINT, double *, FINT);
+   void F77_dsdot_sub(FINT, const float *, FINT, const float *, FINT, double *);
+   void F77_ddot_sub( FINT, const double *, FINT, const double *, FINT, double *);
+   void F77_dscal( FINT, const double *, double *, FINT);
+   void F77_dnrm2_sub( FINT, const double *, FINT, double *);
+   void F77_dasum_sub( FINT, const double *, FINT, double *);
+   void F77_idamax_sub( FINT, const double * , FINT, FINT2);
+
+/* Single Complex Precision */
+
+   void F77_cswap( FINT, void *, FINT, void *, FINT);
+   void F77_ccopy( FINT, const void *, FINT, void *, FINT);
+   void F77_caxpy( FINT, const void *, const void *, FINT, void *, FINT);
+   void F77_cswap( FINT, void *, FINT, void *, FINT);
+   void F77_cdotc_sub( FINT, const void *, FINT, const void *, FINT, void *);
+   void F77_cdotu_sub( FINT, const void *, FINT, const void *, FINT, void *);
+   void F77_cscal( FINT, const void *, void *, FINT);
+   void F77_icamax_sub( FINT, const void *, FINT, FINT2);
+   void F77_csscal( FINT, const float *, void *, FINT);
+   void F77_scnrm2_sub( FINT, const void *, FINT, float *);
+   void F77_scasum_sub( FINT, const void *, FINT, float *);
+
+/* Double Complex Precision */
+
+   void F77_zswap( FINT, void *, FINT, void *, FINT);
+   void F77_zcopy( FINT, const void *, FINT, void *, FINT);
+   void F77_zaxpy( FINT, const void *, const void *, FINT, void *, FINT);
+   void F77_zswap( FINT, void *, FINT, void *, FINT);
+   void F77_zdotc_sub( FINT, const void *, FINT, const void *, FINT, void *);
+   void F77_zdotu_sub( FINT, const void *, FINT, const void *, FINT, void *);
+   void F77_zdscal( FINT, const double *, void *, FINT);
+   void F77_zscal( FINT, const void *, void *, FINT);
+   void F77_dznrm2_sub( FINT, const void *, FINT, double *);
+   void F77_dzasum_sub( FINT, const void *, FINT, double *);
+   void F77_izamax_sub( FINT, const void *, FINT, FINT2);
+
+/*
+ * Level 2 Fortran Prototypes
+ */
+
+/* Single Precision */
+
+   void F77_sgemv(FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_sgbmv(FCHAR, FINT, FINT, FINT, FINT, const float *,  const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_ssymv(FCHAR, FINT, const float *, const float *, FINT, const float *,  FINT, const float *, float *, FINT);
+   void F77_ssbmv(FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_sspmv(FCHAR, FINT, const float *, const float *, const float *, FINT, const float *, float *, FINT);
+   void F77_strmv( FCHAR, FCHAR, FCHAR, FINT, const float *, FINT, float *, FINT);
+   void F77_stbmv( FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, FINT, float *, FINT);
+   void F77_strsv( FCHAR, FCHAR, FCHAR, FINT, const float *, FINT, float *, FINT);
+   void F77_stbsv( FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, FINT, float *, FINT);
+   void F77_stpmv( FCHAR, FCHAR, FCHAR, FINT, const float *, float *, FINT);
+   void F77_stpsv( FCHAR, FCHAR, FCHAR, FINT, const float *, float *, FINT);
+   void F77_sger( FINT, FINT, const float *, const float *, FINT, const float *, FINT, float *, FINT);
+   void F77_ssyr(FCHAR, FINT, const float *, const float *, FINT, float *, FINT);
+   void F77_sspr(FCHAR, FINT, const float *, const float *, FINT, float *); 
+   void F77_sspr2(FCHAR, FINT, const float *, const float *, FINT, const float *, FINT,  float *); 
+   void F77_ssyr2(FCHAR, FINT, const float *, const float *, FINT, const float *, FINT,  float *, FINT);
+
+/* Double Precision */
+
+   void F77_dgemv(FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dgbmv(FCHAR, FINT, FINT, FINT, FINT, const double *,  const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dsymv(FCHAR, FINT, const double *, const double *, FINT, const double *,  FINT, const double *, double *, FINT);
+   void F77_dsbmv(FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dspmv(FCHAR, FINT, const double *, const double *, const double *, FINT, const double *, double *, FINT);
+   void F77_dtrmv( FCHAR, FCHAR, FCHAR, FINT, const double *, FINT, double *, FINT);
+   void F77_dtbmv( FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, FINT, double *, FINT);
+   void F77_dtrsv( FCHAR, FCHAR, FCHAR, FINT, const double *, FINT, double *, FINT);
+   void F77_dtbsv( FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, FINT, double *, FINT);
+   void F77_dtpmv( FCHAR, FCHAR, FCHAR, FINT, const double *, double *, FINT);
+   void F77_dtpsv( FCHAR, FCHAR, FCHAR, FINT, const double *, double *, FINT);
+   void F77_dger( FINT, FINT, const double *, const double *, FINT, const double *, FINT, double *, FINT);
+   void F77_dsyr(FCHAR, FINT, const double *, const double *, FINT, double *, FINT);
+   void F77_dspr(FCHAR, FINT, const double *, const double *, FINT, double *); 
+   void F77_dspr2(FCHAR, FINT, const double *, const double *, FINT, const double *, FINT,  double *); 
+   void F77_dsyr2(FCHAR, FINT, const double *, const double *, FINT, const double *, FINT,  double *, FINT);
+
+/* Single Complex Precision */
+
+   void F77_cgemv(FCHAR, FINT, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_cgbmv(FCHAR, FINT, FINT, FINT, FINT, const void *,  const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_chemv(FCHAR, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_chbmv(FCHAR, FINT, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_chpmv(FCHAR, FINT, const void *, const void *, const void *, FINT, const void *, void *, FINT);
+   void F77_ctrmv( FCHAR, FCHAR, FCHAR, FINT, const void *, FINT, void *, FINT);
+   void F77_ctbmv( FCHAR, FCHAR, FCHAR, FINT, FINT, const void *, FINT, void *, FINT);
+   void F77_ctpmv( FCHAR, FCHAR, FCHAR, FINT, const void *, void *, FINT);
+   void F77_ctrsv( FCHAR, FCHAR, FCHAR, FINT, const void *, FINT, void *, FINT);
+   void F77_ctbsv( FCHAR, FCHAR, FCHAR, FINT, FINT, const void *, FINT, void *, FINT);
+   void F77_ctpsv( FCHAR, FCHAR, FCHAR, FINT, const void *, void *,FINT);
+   void F77_cgerc( FINT, FINT, const void *, const void *, FINT, const void *, FINT, void *, FINT);
+   void F77_cgeru( FINT, FINT, const void *, const void *, FINT, const void *, FINT, void *,  FINT);
+   void F77_cher(FCHAR, FINT, const float *, const void *, FINT, void *, FINT);
+   void F77_cher2(FCHAR, FINT, const void *, const void *, FINT, const void *, FINT, void *, FINT);
+   void F77_chpr(FCHAR, FINT, const float *, const void *, FINT, void *);
+   void F77_chpr2(FCHAR, FINT, const float *, const void *, FINT, const void *, FINT, void *);
+
+/* Double Complex Precision */
+
+   void F77_zgemv(FCHAR, FINT, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_zgbmv(FCHAR, FINT, FINT, FINT, FINT, const void *,  const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_zhemv(FCHAR, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_zhbmv(FCHAR, FINT, FINT, const void *, const void *, FINT, const void *, FINT, const void *, void *, FINT);
+   void F77_zhpmv(FCHAR, FINT, const void *, const void *, const void *, FINT, const void *, void *, FINT);
+   void F77_ztrmv( FCHAR, FCHAR, FCHAR, FINT, const void *, FINT, void *, FINT);
+   void F77_ztbmv( FCHAR, FCHAR, FCHAR, FINT, FINT, const void *, FINT, void *, FINT);
+   void F77_ztpmv( FCHAR, FCHAR, FCHAR, FINT, const void *, void *, FINT);
+   void F77_ztrsv( FCHAR, FCHAR, FCHAR, FINT, const void *, FINT, void *, FINT);
+   void F77_ztbsv( FCHAR, FCHAR, FCHAR, FINT, FINT, const void *, FINT, void *, FINT);
+   void F77_ztpsv( FCHAR, FCHAR, FCHAR, FINT, const void *, void *,FINT);
+   void F77_zgerc( FINT, FINT, const void *, const void *, FINT, const void *, FINT, void *, FINT);
+   void F77_zgeru( FINT, FINT, const void *, const void *, FINT, const void *, FINT, void *,  FINT);
+   void F77_zher(FCHAR, FINT, const double *, const void *, FINT, void *, FINT);
+   void F77_zher2(FCHAR, FINT, const void *, const void *, FINT, const void *, FINT, void *, FINT);
+   void F77_zhpr(FCHAR, FINT, const double *, const void *, FINT, void *);
+   void F77_zhpr2(FCHAR, FINT, const double *, const void *, FINT, const void *, FINT, void *);
+
+/*
+ * Level 3 Fortran Prototypes
+ */
+
+/* Single Precision */
+
+   void F77_sgemm(FCHAR, FCHAR, FINT, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_ssymm(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_ssyrk(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, float *, FINT);
+   void F77_ssyr2k(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_strmm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, float *, FINT);
+   void F77_strsm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, float *, FINT);
+
+/* Double Precision */
+
+   void F77_dgemm(FCHAR, FCHAR, FINT, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dsymm(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dsyrk(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, double *, FINT);
+   void F77_dsyr2k(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_dtrmm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, double *, FINT);
+   void F77_dtrsm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, double *, FINT);
+
+/* Single Complex Precision */
+
+   void F77_cgemm(FCHAR, FCHAR, FINT, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_csymm(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_chemm(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_csyrk(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, float *, FINT);
+   void F77_cherk(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, float *, FINT);
+   void F77_csyr2k(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_cher2k(FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, const float *, FINT, const float *, float *, FINT);
+   void F77_ctrmm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, float *, FINT);
+   void F77_ctrsm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const float *, const float *, FINT, float *, FINT);
+
+/* Double Complex Precision */
+
+   void F77_zgemm(FCHAR, FCHAR, FINT, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_zsymm(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_zhemm(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_zsyrk(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, double *, FINT);
+   void F77_zherk(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, double *, FINT);
+   void F77_zsyr2k(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_zher2k(FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, const double *, FINT, const double *, double *, FINT);
+   void F77_ztrmm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, double *, FINT);
+   void F77_ztrsm(FCHAR, FCHAR, FCHAR, FCHAR, FINT, FINT, const double *, const double *, FINT, double *, FINT);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*  CBLAS_F77_H */
diff --git a/ml/dlib/dlib/external/cblas/cblas_icamax.c b/ml/dlib/dlib/external/cblas/cblas_icamax.c
new file mode 100644
index 000000000..b3ffe6eec
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_icamax.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_icamax.c
+ *
+ * The program is a C interface to icamax.
+ * It calls the fortran wrapper before calling icamax.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+CBLAS_INDEX cblas_icamax( const int N, const void *X, const int incX)
+{
+   F77_INT iamax;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_icamax_sub( &F77_N, X, &F77_incX, &iamax);
+   return iamax ? iamax-1 : 0;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_idamax.c b/ml/dlib/dlib/external/cblas/cblas_idamax.c
new file mode 100644
index 000000000..e42e459ea
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_idamax.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_idamax.c
+ *
+ * The program is a C interface to idamax.
+ * It calls the fortran wrapper before calling idamax.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+CBLAS_INDEX cblas_idamax( const int N, const double *X, const int incX)
+{
+   F77_INT iamax;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_idamax_sub( &F77_N, X, &F77_incX, &iamax);
+   return iamax ? iamax-1 : 0;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_isamax.c b/ml/dlib/dlib/external/cblas/cblas_isamax.c
new file mode 100644
index 000000000..63d639c7f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_isamax.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_isamax.c
+ *
+ * The program is a C interface to isamax.
+ * It calls the fortran wrapper before calling isamax.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+CBLAS_INDEX cblas_isamax( const int N, const float *X, const int incX)
+{
+   F77_INT iamax;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_isamax_sub( &F77_N, X, &F77_incX, &iamax);
+   return iamax ? iamax-1 : 0;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_izamax.c b/ml/dlib/dlib/external/cblas/cblas_izamax.c
new file mode 100644
index 000000000..78eda4042
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_izamax.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_izamax.c
+ *
+ * The program is a C interface to izamax.
+ * It calls the fortran wrapper before calling izamax.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+CBLAS_INDEX cblas_izamax( const int N, const void *X, const int incX)
+{
+   F77_INT iamax;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_izamax_sub( &F77_N, X, &F77_incX, &iamax);
+   return (iamax ? iamax-1 : 0);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sasum.c b/ml/dlib/dlib/external/cblas/cblas_sasum.c
new file mode 100644
index 000000000..7d4c32cf9
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sasum.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_sasum.c
+ *
+ * The program is a C interface to sasum.
+ * It calls the fortran wrapper before calling sasum.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_sasum( const int N, const float *X, const int incX) 
+{
+   float asum;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_sasum_sub( &F77_N, X, &F77_incX, &asum);
+   return asum;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_saxpy.c b/ml/dlib/dlib/external/cblas/cblas_saxpy.c
new file mode 100644
index 000000000..2eee8e06e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_saxpy.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_saxpy.c
+ *
+ * The program is a C interface to saxpy.
+ * It calls the fortran wrapper before calling saxpy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_saxpy( const int N, const float alpha, const float *X,
+                       const int incX, float *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_saxpy( &F77_N, &alpha, X, &F77_incX, Y, &F77_incY);
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_scasum.c b/ml/dlib/dlib/external/cblas/cblas_scasum.c
new file mode 100644
index 000000000..e1fa53090
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_scasum.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_scasum.c
+ *
+ * The program is a C interface to scasum.
+ * It calls the fortran wrapper before calling scasum.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_scasum( const int N, const void *X, const int incX) 
+{
+   float asum;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_scasum_sub( &F77_N, X, &F77_incX, &asum);
+   return asum;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_scnrm2.c b/ml/dlib/dlib/external/cblas/cblas_scnrm2.c
new file mode 100644
index 000000000..fa48454ed
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_scnrm2.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_scnrm2.c
+ *
+ * The program is a C interface to scnrm2.
+ * It calls the fortran wrapper before calling scnrm2.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_scnrm2( const int N, const void *X, const int incX) 
+{
+   float nrm2;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_scnrm2_sub( &F77_N, X, &F77_incX, &nrm2);
+   return nrm2;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_scopy.c b/ml/dlib/dlib/external/cblas/cblas_scopy.c
new file mode 100644
index 000000000..7796959f3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_scopy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_scopy.c
+ *
+ * The program is a C interface to scopy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_scopy( const int N, const float *X,
+                      const int incX, float *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_scopy( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sdot.c b/ml/dlib/dlib/external/cblas/cblas_sdot.c
new file mode 100644
index 000000000..baf859272
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sdot.c
@@ -0,0 +1,25 @@
+/*
+ * cblas_sdot.c
+ *
+ * The program is a C interface to sdot.
+ * It calls the fortran wrapper before calling sdot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_sdot( const int N, const float *X,
+                      const int incX, const float *Y, const int incY)
+{
+   float dot;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_sdot_sub( &F77_N, X, &F77_incX, Y, &F77_incY, &dot);
+   return dot;
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_sdsdot.c b/ml/dlib/dlib/external/cblas/cblas_sdsdot.c
new file mode 100644
index 000000000..b824849b9
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sdsdot.c
@@ -0,0 +1,25 @@
+/*
+ * cblas_sdsdot.c
+ *
+ * The program is a C interface to sdsdot.
+ * It calls the fortran wrapper before calling sdsdot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_sdsdot( const int N, const float alpha, const float *X,
+                      const int incX, const float *Y, const int incY)
+{
+   float dot;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_sdsdot_sub( &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, &dot);
+   return dot;
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_sgbmv.c b/ml/dlib/dlib/external/cblas/cblas_sgbmv.c
new file mode 100644
index 000000000..b6de24977
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sgbmv.c
@@ -0,0 +1,72 @@
+/*
+ *
+ * cblas_sgbmv.c
+ * This program is a C interface to sgbmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU,
+                 const float alpha, const float *A, const int lda,
+                 const float  *X, const int incX, const float beta,
+                 float  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+   F77_INT F77_KL=KL,F77_KU=KU;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_KL KL
+   #define F77_KU KU
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_sgbmv(F77_TA, &F77_M, &F77_N, &F77_KL, &F77_KU, &alpha,  
+                     A, &F77_lda, X, &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_sgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, &alpha, 
+                     A ,&F77_lda, X, &F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_sgbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sgemm.c b/ml/dlib/dlib/external/cblas/cblas_sgemm.c
new file mode 100644
index 000000000..f8adeda49
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sgemm.c
@@ -0,0 +1,95 @@
+/*
+ *
+ * cblas_sgemm.c
+ * This program is a C interface to sgemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const float alpha, const float  *A,
+                 const int lda, const float  *B, const int ldb,
+                 const float beta, float  *C, const int ldc)
+{
+   char TA, TB;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_TB;
+#else
+   #define F77_TA &TA  
+   #define F77_TB &TB  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+   
+   if( Order == CblasColMajor )
+   {
+      if(TransA == CblasTrans) TA='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgemm", 
+                       "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if(TransB == CblasTrans) TB='T';
+      else if ( TransB == CblasConjTrans ) TB='C';
+      else if ( TransB == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_sgemm", 
+                       "Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_sgemm(F77_TA, F77_TB, &F77_M, &F77_N, &F77_K, &alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if(TransA == CblasTrans) TB='T';
+      else if ( TransA == CblasConjTrans ) TB='C';
+      else if ( TransA == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgemm", 
+                       "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if(TransB == CblasTrans) TA='T';
+      else if ( TransB == CblasConjTrans ) TA='C';
+      else if ( TransB == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgemm", 
+                       "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_sgemm(F77_TA, F77_TB, &F77_N, &F77_M, &F77_K, &alpha, B, &F77_ldb, A, &F77_lda, &beta, C, &F77_ldc);
+   } else  
+     cblas_xerbla(1, "cblas_sgemm",
+                     "Illegal Order setting, %d\n", Order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sgemv.c b/ml/dlib/dlib/external/cblas/cblas_sgemv.c
new file mode 100644
index 000000000..d47f3be55
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sgemv.c
@@ -0,0 +1,67 @@
+/*
+ *
+ * cblas_sgemv.c
+ * This program is a C interface to sgemv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ * 
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const float alpha, const float  *A, const int lda,
+                 const float  *X, const int incX, const float beta,
+                 float  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgemv","Illegal TransA setting, %d\n", TransA);
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_sgemv(F77_TA, &F77_M, &F77_N, &alpha, A, &F77_lda, X, &F77_incX, 
+                &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sgemv", "Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_sgemv(F77_TA, &F77_N, &F77_M, &alpha, A, &F77_lda, X,
+                &F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_sgemv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sger.c b/ml/dlib/dlib/external/cblas/cblas_sger.c
new file mode 100644
index 000000000..0313590c7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sger.c
@@ -0,0 +1,39 @@
+/*
+ *
+ * cblas_sger.c
+ * This program is a C interface to sger.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sger(const enum CBLAS_ORDER order, const int M, const int N,
+                const float  alpha, const float  *X, const int incX,
+                const float  *Y, const int incY, float  *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda lda
+#endif
+
+
+   if (order == CblasColMajor)
+   {
+      F77_sger( &F77_M, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A,
+       &F77_lda);
+   }
+   else if (order == CblasRowMajor)
+   {
+      F77_sger( &F77_N, &F77_M, &alpha, Y, &F77_incY, X, &F77_incX, A, 
+        &F77_lda);
+   }
+   else cblas_xerbla(1, "cblas_sger", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_snrm2.c b/ml/dlib/dlib/external/cblas/cblas_snrm2.c
new file mode 100644
index 000000000..18161b4fa
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_snrm2.c
@@ -0,0 +1,23 @@
+/*
+ * cblas_snrm2.c
+ *
+ * The program is a C interface to snrm2.
+ * It calls the fortran wrapper before calling snrm2.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+float cblas_snrm2( const int N, const float *X, const int incX) 
+{
+   float nrm2;
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_snrm2_sub( &F77_N, X, &F77_incX, &nrm2);
+   return nrm2;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_srot.c b/ml/dlib/dlib/external/cblas/cblas_srot.c
new file mode 100644
index 000000000..cbd1c8c90
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_srot.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_srot.c
+ *
+ * The program is a C interface to srot.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_srot( const int N, float *X, const int incX, float *Y, 
+                      const int incY, const float  c, const float  s)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_srot(&F77_N, X, &F77_incX, Y, &F77_incY, &c, &s);
+}  
diff --git a/ml/dlib/dlib/external/cblas/cblas_srotg.c b/ml/dlib/dlib/external/cblas/cblas_srotg.c
new file mode 100644
index 000000000..f6460048d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_srotg.c
@@ -0,0 +1,14 @@
+/*
+ * cblas_srotg.c
+ *
+ * The program is a C interface to srotg.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_srotg(  float *a, float *b, float *c, float *s)
+{
+   F77_srotg(a,b,c,s);    
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_srotm.c b/ml/dlib/dlib/external/cblas/cblas_srotm.c
new file mode 100644
index 000000000..496746454
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_srotm.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_srotm.c
+ *
+ * The program is a C interface to srotm.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_srotm( const int N, float *X, const int incX, float *Y, 
+                       const int incY, const float *P)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_srotm( &F77_N, X, &F77_incX, Y, &F77_incY, P);
+}   
diff --git a/ml/dlib/dlib/external/cblas/cblas_srotmg.c b/ml/dlib/dlib/external/cblas/cblas_srotmg.c
new file mode 100644
index 000000000..04f978b40
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_srotmg.c
@@ -0,0 +1,15 @@
+/*
+ * cblas_srotmg.c
+ *
+ * The program is a C interface to srotmg.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_srotmg( float *d1, float *d2, float *b1, 
+                        const float b2, float *p)
+{
+   F77_srotmg(d1,d2,b1,&b2,p);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssbmv.c b/ml/dlib/dlib/external/cblas/cblas_ssbmv.c
new file mode 100644
index 000000000..69663b619
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssbmv.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_ssbmv.c
+ * This program is a C interface to ssbmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+           const int N, const int K, const float alpha, const float *A,
+           const int lda, const float *X, const int incX,
+           const float beta, float *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+   
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssbmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_ssbmv(F77_UL, &F77_N, &F77_K, &alpha, A, &F77_lda, X,
+      &F77_incX, &beta, Y, &F77_incY);
+   }else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_ssbmv(F77_UL, &F77_N, &F77_K, &alpha, A, &F77_lda, X,
+      &F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_ssbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sscal.c b/ml/dlib/dlib/external/cblas/cblas_sscal.c
new file mode 100644
index 000000000..1f09abe7a
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_sscal.c
+ *
+ * The program is a C interface to sscal.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sscal( const int N, const float alpha, float *X, 
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_sscal( &F77_N, &alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sspmv.c b/ml/dlib/dlib/external/cblas/cblas_sspmv.c
new file mode 100644
index 000000000..e3485e742
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sspmv.c
@@ -0,0 +1,62 @@
+/*
+ *
+ * cblas_sspmv.c
+ * This program is a C interface to sspmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sspmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const float alpha, const float  *AP,
+                 const float  *X, const int incX, const float beta,
+                 float  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_sspmv(F77_UL, &F77_N, &alpha, AP, X,  
+                     &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_sspmv(F77_UL, &F77_N, &alpha, 
+                     AP, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_sspmv", "Illegal Order setting, %d\n", order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sspr.c b/ml/dlib/dlib/external/cblas/cblas_sspr.c
new file mode 100644
index 000000000..75d669b30
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sspr.c
@@ -0,0 +1,61 @@
+/*
+ *
+ * cblas_sspr.c
+ * This program is a C interface to sspr.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sspr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const  float alpha, const float *X,
+                const int incX, float *Ap)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_sspr(F77_UL, &F77_N, &alpha, X, &F77_incX, Ap);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_sspr(F77_UL, &F77_N, &alpha, X, &F77_incX, Ap); 
+   } else cblas_xerbla(1, "cblas_sspr", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sspr2.c b/ml/dlib/dlib/external/cblas/cblas_sspr2.c
new file mode 100644
index 000000000..b6ff9bfa2
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sspr2.c
@@ -0,0 +1,60 @@
+/*
+ *
+ * cblas_sspr2.c
+ * This program is a C interface to sspr2.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sspr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float  alpha, const float  *X,
+                const int incX, const float  *Y, const int incY, float  *A)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_sspr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_sspr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_sspr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY,  A); 
+   } else cblas_xerbla(1, "cblas_sspr2", "Illegal Order setting, %d\n", order);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_sswap.c b/ml/dlib/dlib/external/cblas/cblas_sswap.c
new file mode 100644
index 000000000..b74d8469c
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_sswap.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_sswap.c
+ *
+ * The program is a C interface to sswap.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_sswap( const int N, float *X, const int incX, float *Y,
+                       const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_sswap( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssymm.c b/ml/dlib/dlib/external/cblas/cblas_ssymm.c
new file mode 100644
index 000000000..55c413b48
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssymm.c
@@ -0,0 +1,93 @@
+/*
+ *
+ * cblas_ssymm.c
+ * This program is a C interface to ssymm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const float alpha, const float  *A, const int lda,
+                 const float  *B, const int ldb, const float beta,
+                 float  *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssymm", 
+                       "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssymm", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_ssymm(F77_SD, F77_UL, &F77_M, &F77_N, &alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssymm", 
+                       "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssymm", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_ssymm(F77_SD, F77_UL, &F77_N, &F77_M, &alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else  cblas_xerbla(1, "cblas_ssymm",
+                     "Illegal Order setting, %d\n", Order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssymv.c b/ml/dlib/dlib/external/cblas/cblas_ssymv.c
new file mode 100644
index 000000000..c56d4d457
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssymv.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_ssymv.c
+ * This program is a C interface to ssymv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssymv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const float alpha, const float  *A, const int lda,
+                 const float  *X, const int incX, const float beta,
+                 float  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssymv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_ssymv(F77_UL, &F77_N, &alpha, A, &F77_lda, X,  
+                     &F77_incX, &beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssymv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_ssymv(F77_UL, &F77_N, &alpha, 
+                     A ,&F77_lda, X,&F77_incX, &beta, Y, &F77_incY);
+   }
+   else cblas_xerbla(1, "cblas_ssymv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssyr.c b/ml/dlib/dlib/external/cblas/cblas_ssyr.c
new file mode 100644
index 000000000..4215b9671
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssyr.c
@@ -0,0 +1,59 @@
+/*
+ *
+ * cblas_ssyr.c
+ * This program is a C interface to ssyr.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssyr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float  alpha, const float  *X,
+                const int incX, float  *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_lda=lda;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_lda  lda
+#endif
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_ssyr(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_ssyr(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda); 
+   } else cblas_xerbla(1, "cblas_ssyr", "Illegal Order setting, %d\n", order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssyr2.c b/ml/dlib/dlib/external/cblas/cblas_ssyr2.c
new file mode 100644
index 000000000..9cdaa412d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssyr2.c
@@ -0,0 +1,65 @@
+/*
+ *
+ * cblas_ssyr2.c
+ * This program is a C interface to ssyr2.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssyr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const float  alpha, const float  *X,
+                const int incX, const float  *Y, const int incY, float  *A,
+                const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY, F77_lda=lda;
+#else
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda  lda
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_ssyr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY, A, 
+                    &F77_lda);
+
+   }  else if (order == CblasRowMajor) 
+   {
+      if (Uplo == CblasLower) UL = 'U';
+      else if (Uplo == CblasUpper) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif  
+      F77_ssyr2(F77_UL, &F77_N, &alpha, X, &F77_incX, Y, &F77_incY,  A, 
+                    &F77_lda); 
+   } else cblas_xerbla(1, "cblas_ssyr2", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssyr2k.c b/ml/dlib/dlib/external/cblas/cblas_ssyr2k.c
new file mode 100644
index 000000000..9e9f538df
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssyr2k.c
@@ -0,0 +1,96 @@
+/*
+ *
+ * cblas_ssyr2k.c
+ * This program is a C interface to ssyr2k.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const float alpha, const float  *A, const int lda,
+                  const float  *B, const int ldb, const float beta,
+                  float  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyr2k", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyr2k", 
+                       "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_ssyr2k(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyr2k", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyr2k", 
+                       "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_ssyr2k(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else  cblas_xerbla(1, "cblas_ssyr2k",
+                     "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ssyrk.c b/ml/dlib/dlib/external/cblas/cblas_ssyrk.c
new file mode 100644
index 000000000..55ceb7e3a
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ssyrk.c
@@ -0,0 +1,95 @@
+/*
+ *
+ * cblas_ssyrk.c
+ * This program is a C interface to ssyrk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ssyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const float alpha, const float  *A, const int lda,
+                 const float beta, float  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ssyrk", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyrk", 
+                       "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_ssyrk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyrk", 
+                       "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ssyrk", 
+                       "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_ssyrk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda, &beta, C, &F77_ldc);
+   } else  cblas_xerbla(1, "cblas_ssyrk",
+                     "Illegal Order setting, %d\n", Order);
+   return;
+}
+
diff --git a/ml/dlib/dlib/external/cblas/cblas_stbmv.c b/ml/dlib/dlib/external/cblas/cblas_stbmv.c
new file mode 100644
index 000000000..71ef469a7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_stbmv.c
@@ -0,0 +1,103 @@
+/*
+ * cblas_stbmv.c       
+ * This program is a C interface to stbmv.
+ * Written by Keita Teranishi
+ * 3/3/1998
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+
+void cblas_stbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const float  *A, const int lda,
+                 float  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stbmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_stbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_stbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_stbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_stbsv.c b/ml/dlib/dlib/external/cblas/cblas_stbsv.c
new file mode 100644
index 000000000..96df7c0c1
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_stbsv.c
@@ -0,0 +1,103 @@
+/*
+ * cblas_stbsv.c
+ * The program is a C interface to stbsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_stbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const float  *A, const int lda,
+                 float  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL  = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_stbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_stbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_stbsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_stpmv.c b/ml/dlib/dlib/external/cblas/cblas_stpmv.c
new file mode 100644
index 000000000..5cb5cd29d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_stpmv.c
@@ -0,0 +1,99 @@
+/*
+ *
+ * cblas_stpmv.c
+ * This program is a C interface to stpmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_stpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float  *Ap, float  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_stpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_stpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_stpmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_stpsv.c b/ml/dlib/dlib/external/cblas/cblas_stpsv.c
new file mode 100644
index 000000000..2f0d29c0d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_stpsv.c
@@ -0,0 +1,99 @@
+/*
+ * cblas_stpsv.c
+ * The program is a C interface to stpsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_stpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float  *Ap, float  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_stpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_stpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_stpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_stpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_stpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+
+   }
+   else cblas_xerbla(1, "cblas_stpsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_strmm.c b/ml/dlib/dlib/external/cblas/cblas_strmm.c
new file mode 100644
index 000000000..40d6e23dd
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_strmm.c
@@ -0,0 +1,125 @@
+/*
+ *
+ * cblas_strmm.c
+ * This program is a C interface to strmm.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_strmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const  enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const float alpha, const float *A, const int lda,
+                 float *B, const int ldb)
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strmm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strmm","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strmm","Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_strmm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_strmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, &alpha, A, &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strmm","Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strmm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strmm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_strmm","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+#ifdef F77_CHAR
+    F77_UL = C2F_CHAR(&UL);
+    F77_TA = C2F_CHAR(&TA);
+    F77_SD = C2F_CHAR(&SD);
+    F77_DI = C2F_CHAR(&DI);
+#endif
+      F77_strmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, &alpha, A,
+      &F77_lda, B, &F77_ldb);
+   } 
+   else  cblas_xerbla(1, "cblas_strmm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_strmv.c b/ml/dlib/dlib/external/cblas/cblas_strmv.c
new file mode 100644
index 000000000..4c2f7b6a6
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_strmv.c
@@ -0,0 +1,103 @@
+/*
+ *
+ * cblas_strmv.c
+ * This program is a C interface to strmv.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_strmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float  *A, const int lda,
+                 float  *X, const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_strmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_strmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_strmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_strsm.c b/ml/dlib/dlib/external/cblas/cblas_strsm.c
new file mode 100644
index 000000000..178cf7d06
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_strsm.c
@@ -0,0 +1,120 @@
+/*
+ *
+ * cblas_strsm.c
+ * This program is a C interface to strsm.
+ * Written by Keita Teranishi
+ * 4/6/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_strsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const float alpha, const float  *A, const int lda,
+                 float  *B, const int ldb)
+
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_strsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_strsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, &alpha, A, &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_strsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_strsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, &alpha, A, &F77_lda, B, &F77_ldb);
+   } 
+   else  cblas_xerbla(1, "cblas_strsm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_strsv.c b/ml/dlib/dlib/external/cblas/cblas_strsv.c
new file mode 100644
index 000000000..7c3811974
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_strsv.c
@@ -0,0 +1,102 @@
+/*
+ * cblas_strsv.c
+ * The program is a C interface to strsv.
+ *
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_strsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const float  *A, const int lda, float  *X,
+                 const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_strsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_strsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans) TA = 'N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_strsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_strsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_strsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else cblas_xerbla(1, "cblas_strsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_xerbla.c b/ml/dlib/dlib/external/cblas/cblas_xerbla.c
new file mode 100644
index 000000000..0b5b39f53
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_xerbla.c
@@ -0,0 +1,66 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <stdarg.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+
+void cblas_xerbla(int info, const char *rout, const char *form, ...)
+{
+   char empty[1] = "";
+   va_list argptr;
+
+   va_start(argptr, form);
+
+   {
+      if (strstr(rout,"gemm") != 0)
+      {
+         if      (info == 5 ) info =  4;
+         else if (info == 4 ) info =  5;
+         else if (info == 11) info =  9;
+         else if (info == 9 ) info = 11;
+      }
+      else if (strstr(rout,"symm") != 0 || strstr(rout,"hemm") != 0)
+      {
+         if      (info == 5 ) info =  4;
+         else if (info == 4 ) info =  5;
+      }
+      else if (strstr(rout,"trmm") != 0 || strstr(rout,"trsm") != 0)
+      {
+         if      (info == 7 ) info =  6;
+         else if (info == 6 ) info =  7;
+      }
+      else if (strstr(rout,"gemv") != 0)
+      {
+         if      (info == 4)  info = 3;
+         else if (info == 3)  info = 4;
+      }
+      else if (strstr(rout,"gbmv") != 0)
+      {
+         if      (info == 4)  info = 3;
+         else if (info == 3)  info = 4;
+         else if (info == 6)  info = 5;
+         else if (info == 5)  info = 6;
+      }
+      else if (strstr(rout,"ger") != 0)
+      {
+         if      (info == 3) info = 2;
+         else if (info == 2) info = 3;
+         else if (info == 8) info = 6;
+         else if (info == 6) info = 8;
+      }
+      else if ( (strstr(rout,"her2") != 0 || strstr(rout,"hpr2") != 0)
+                 && strstr(rout,"her2k") == 0 )
+      {
+         if      (info == 8) info = 6;
+         else if (info == 6) info = 8;
+      }
+   }
+   if (info)
+      fprintf(stderr, "Parameter %d to routine %s was incorrect\n", info, rout);
+   vfprintf(stderr, form, argptr);
+   va_end(argptr);
+   if (info && !info) 
+      F77_xerbla(empty, &info); /* Force link of our F77 error handler */
+   exit(-1);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zaxpy.c b/ml/dlib/dlib/external/cblas/cblas_zaxpy.c
new file mode 100644
index 000000000..f63c4c39b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zaxpy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_zaxpy.c
+ *
+ * The program is a C interface to zaxpy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zaxpy( const int N, const void *alpha, const void *X,
+                       const int incX, void *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_zaxpy( &F77_N, alpha, X, &F77_incX, Y, &F77_incY);
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_zcopy.c b/ml/dlib/dlib/external/cblas/cblas_zcopy.c
new file mode 100644
index 000000000..a16be28e7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zcopy.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_zcopy.c
+ *
+ * The program is a C interface to zcopy.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zcopy( const int N, const void *X,
+                      const int incX, void *Y, const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_zcopy( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zdotc_sub.c b/ml/dlib/dlib/external/cblas/cblas_zdotc_sub.c
new file mode 100644
index 000000000..29dec6c57
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zdotc_sub.c
@@ -0,0 +1,24 @@
+/*
+ * cblas_zdotc_sub.c
+ *
+ * The program is a C interface to zdotc.
+ * It calls the fortran wrapper before calling zdotc.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zdotc_sub( const int N, const void *X, const int incX,
+                    const void *Y, const int incY, void *dotc)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_zdotc_sub( &F77_N, X, &F77_incX, Y, &F77_incY, dotc);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zdotu_sub.c b/ml/dlib/dlib/external/cblas/cblas_zdotu_sub.c
new file mode 100644
index 000000000..48a14bf3d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zdotu_sub.c
@@ -0,0 +1,24 @@
+/*
+ * cblas_zdotu_sub.c
+ *
+ * The program is a C interface to zdotu.
+ * It calls the fortran wrapper before calling zdotu.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zdotu_sub( const int N, const void *X, const int incX,
+                      const void *Y, const int incY, void *dotu)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_zdotu_sub( &F77_N, X, &F77_incX, Y, &F77_incY, dotu);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zdscal.c b/ml/dlib/dlib/external/cblas/cblas_zdscal.c
new file mode 100644
index 000000000..788365bef
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zdscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_zdscal.c
+ *
+ * The program is a C interface to zdscal.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zdscal( const int N, const double alpha, void  *X,
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_zdscal( &F77_N, &alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zgbmv.c b/ml/dlib/dlib/external/cblas/cblas_zgbmv.c
new file mode 100644
index 000000000..e5b922429
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zgbmv.c
@@ -0,0 +1,155 @@
+/*
+ * cblas_zgbmv.c
+ * The program is a C interface of zgbmv
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zgbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const int KL, const int KU,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+   F77_INT F77_KL=KL,F77_KU=KU;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_KL KL
+   #define F77_KU KU
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const double *xx= (double *)X, *alp= (double *)alpha, *bet = (double *)beta;
+   double ALPHA[2],BETA[2];
+   int tincY, tincx;
+   double *x=(double *)X, *y=(double *)Y, *st=0, *tx;
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_zgbmv(F77_TA, &F77_M, &F77_N, &F77_KL, &F77_KU, alpha,  
+                     A, &F77_lda, X, &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         ALPHA[0]= *alp;
+         ALPHA[1]= -alp[1];
+         BETA[0]= *bet;
+         BETA[1]= -bet[1];
+         TA = 'N';
+         if (M > 0)
+         {
+            n = M << 1;
+            x = malloc(n*sizeof(double));
+            tx = x;
+
+            if( incX > 0 ) {
+               i = incX << 1 ;
+               tincx = 2;
+               st= x+n;
+            } else {
+               i = incX *(-2);
+               tincx = -2;
+               st = x-2;
+               x +=(n-2);
+            }
+            do
+            {
+               *x = *xx;
+               x[1] = -xx[1];
+               x += tincx ;
+               xx += i;
+            }
+            while (x != st);
+            x=tx;
+
+            #ifdef F77_INT
+               F77_incX = 1;
+            #else
+               incx = 1;
+            #endif
+
+            if( incY > 0 )
+              tincY = incY;
+            else
+              tincY = -incY;
+
+            y++;
+ 
+            if (N > 0)
+            {
+               i = tincY << 1;
+               n = i * N ;
+               st = y + n;
+               do {
+                  *y = -(*y);
+                  y += i;
+               } while(y != st);
+               y -= n;
+            }
+         }
+         else x = (double *) X;
+
+ 
+      }
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      if (TransA == CblasConjTrans)
+         F77_zgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, ALPHA, 
+                        A ,&F77_lda, x,&F77_incX, BETA, Y, &F77_incY);
+      else
+         F77_zgbmv(F77_TA, &F77_N, &F77_M, &F77_KU, &F77_KL, alpha, 
+                        A ,&F77_lda, x,&F77_incX, beta, Y, &F77_incY);
+      if (TransA == CblasConjTrans)
+      {
+         if (x != X) free(x);
+         if (N > 0)
+         {
+            do
+            {
+               *y = -(*y);
+               y += i;
+            }
+            while (y != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_zgbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zgemm.c b/ml/dlib/dlib/external/cblas/cblas_zgemm.c
new file mode 100644
index 000000000..c348afa2d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zgemm.c
@@ -0,0 +1,94 @@
+/*
+ *
+ * cblas_zgemm.c
+ * This program is a C interface to zgemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zgemm(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_TRANSPOSE TransB, const int M, const int N,
+                 const int K, const void *alpha, const void  *A,
+                 const int lda, const void  *B, const int ldb,
+                 const void *beta, void  *C, const int ldc)
+{
+   char TA, TB;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_TB;
+#else
+   #define F77_TA &TA  
+   #define F77_TB &TB  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if(TransA == CblasTrans) TA='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgemm","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if(TransB == CblasTrans) TB='T';
+      else if ( TransB == CblasConjTrans ) TB='C';
+      else if ( TransB == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zgemm","Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_zgemm(F77_TA, F77_TB, &F77_M, &F77_N, &F77_K, alpha, A,
+                     &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if(TransA == CblasTrans) TB='T';
+      else if ( TransA == CblasConjTrans ) TB='C';
+      else if ( TransA == CblasNoTrans )   TB='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgemm","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if(TransB == CblasTrans) TA='T';
+      else if ( TransB == CblasConjTrans ) TA='C';
+      else if ( TransB == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgemm","Illegal TransB setting, %d\n", TransB);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+         F77_TB = C2F_CHAR(&TB);
+      #endif
+
+      F77_zgemm(F77_TA, F77_TB, &F77_N, &F77_M, &F77_K, alpha, B,
+                  &F77_ldb, A, &F77_lda, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zgemm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zgemv.c b/ml/dlib/dlib/external/cblas/cblas_zgemv.c
new file mode 100644
index 000000000..6d5cd0cb2
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zgemv.c
@@ -0,0 +1,153 @@
+/*
+ * cblas_zgemv.c
+ * The program is a C interface of zgemv
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zgemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_TRANSPOSE TransA, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char TA;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA;
+#else
+   #define F77_TA &TA   
+#endif
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+
+   int n, i=0, incx=incX;
+   const double *xx= (double *)X, *alp= (double *)alpha, *bet = (double *)beta;
+   double ALPHA[2],BETA[2];
+   int tincY, tincx;
+   double *x=(double *)X, *y=(double *)Y, *st=0, *tx;
+
+
+   if (order == CblasColMajor)
+   {
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      F77_zgemv(F77_TA, &F77_M, &F77_N, alpha, A, &F77_lda, X, &F77_incX, 
+                beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+         
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         ALPHA[0]= *alp;
+         ALPHA[1]= -alp[1];
+         BETA[0]= *bet;
+         BETA[1]= -bet[1];
+         TA = 'N';
+         if (M > 0)
+         {
+            n = M << 1;
+            x = malloc(n*sizeof(double));
+            tx = x;
+            if( incX > 0 ) {
+               i = incX << 1 ;
+               tincx = 2;
+               st= x+n;
+            } else { 
+               i = incX *(-2);
+               tincx = -2;
+               st = x-2; 
+               x +=(n-2); 
+            }
+
+            do
+            {
+               *x = *xx;
+               x[1] = -xx[1];
+               x += tincx ;
+               xx += i;
+            }
+            while (x != st);
+            x=tx;
+
+            #ifdef F77_INT
+               F77_incX = 1;
+            #else
+               incx = 1;
+            #endif
+
+            if(incY > 0)
+               tincY = incY; 
+            else
+               tincY = -incY; 
+
+            y++;
+
+            if (N > 0)
+            {
+               i = tincY << 1;
+               n = i * N ;
+               st = y + n;
+               do {
+                  *y = -(*y);
+                  y += i;
+               } while(y != st); 
+               y -= n;
+            }
+         }
+         else x = (double *) X;
+      }
+      else 
+      {
+         cblas_xerbla(2, "cblas_zgemv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_TA = C2F_CHAR(&TA);
+      #endif
+      if (TransA == CblasConjTrans)
+         F77_zgemv(F77_TA, &F77_N, &F77_M, ALPHA, A, &F77_lda, x,
+                &F77_incX, BETA, Y, &F77_incY);
+      else
+         F77_zgemv(F77_TA, &F77_N, &F77_M, alpha, A, &F77_lda, x,
+                &F77_incX, beta, Y, &F77_incY);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (x != (double *)X) free(x);
+         if (N > 0)
+         {
+            do
+            {
+               *y = -(*y);
+               y += i;
+            }
+            while (y != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_zgemv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zgerc.c b/ml/dlib/dlib/external/cblas/cblas_zgerc.c
new file mode 100644
index 000000000..2fbbcb028
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zgerc.c
@@ -0,0 +1,77 @@
+/*
+ * cblas_zgerc.c
+ * The program is a C interface to zgerc.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zgerc(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incy
+   #define F77_lda lda   
+#endif
+
+   int n, i, tincy, incy=incY;
+   double *y=(double *)Y, *yy=(double *)Y, *ty, *st;
+
+
+   if (order == CblasColMajor)
+   {
+      F77_zgerc( &F77_M, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, A, 
+                      &F77_lda);
+   }  else if (order == CblasRowMajor)   
+   {
+      if (N > 0)
+      {
+         n = N << 1;
+         y = malloc(n*sizeof(double));
+
+         ty = y;
+         if( incY > 0 ) {
+            i = incY << 1;
+            tincy = 2;
+            st= y+n;
+         } else { 
+            i = incY *(-2);
+            tincy = -2;
+            st = y-2; 
+            y +=(n-2); 
+         }
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += tincy ;
+            yy += i;
+         }
+         while (y != st);
+         y = ty;
+
+         #ifdef F77_INT
+            F77_incY = 1;
+         #else
+            incy = 1;
+         #endif
+      }
+      else y = (double *) Y;
+
+      F77_zgeru( &F77_N, &F77_M, alpha, y, &F77_incY, X, &F77_incX, A, 
+                      &F77_lda);
+      if(Y!=y)
+         free(y);
+
+   } else cblas_xerbla(1, "cblas_zgerc", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zgeru.c b/ml/dlib/dlib/external/cblas/cblas_zgeru.c
new file mode 100644
index 000000000..56c3ded68
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zgeru.c
@@ -0,0 +1,37 @@
+/*
+ * cblas_zgeru.c
+ * The program is a C interface to zgeru.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zgeru(const enum CBLAS_ORDER order, const int M, const int N,
+                 const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+   #define F77_lda lda
+#endif
+
+
+   if (order == CblasColMajor)
+   {
+      F77_zgeru( &F77_M, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, A,
+                      &F77_lda);
+   }
+   else if (order == CblasRowMajor)
+   {
+      F77_zgeru( &F77_N, &F77_M, alpha, Y, &F77_incY, X, &F77_incX, A, 
+                      &F77_lda);
+   }
+   else cblas_xerbla(1, "cblas_zgeru", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhbmv.c b/ml/dlib/dlib/external/cblas/cblas_zhbmv.c
new file mode 100644
index 000000000..491207d3f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhbmv.c
@@ -0,0 +1,145 @@
+/*
+ * cblas_zhbmv.c
+ * The program is a C interface to zhbmv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+#include <stdio.h>
+#include <stdlib.h>
+void cblas_zhbmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo,const int N,const int K,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const double *xx= (double *)X, *alp= (double *)alpha, *bet = (double *)beta;
+   double ALPHA[2],BETA[2];
+   int tincY, tincx;
+   double *x=(double *)X, *y=(double *)Y, *st=0, *tx;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhbmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_zhbmv(F77_UL, &F77_N, &F77_K, alpha, A, &F77_lda, X,  
+                     &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1 ;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (double *) X; 
+
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_zhbmv(F77_UL, &F77_N, &F77_K, ALPHA, 
+                     A ,&F77_lda, x,&F77_incX, BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_zhbmv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor )
+   {
+      if(X!=x)
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+      }
+   }
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhemm.c b/ml/dlib/dlib/external/cblas/cblas_zhemm.c
new file mode 100644
index 000000000..a31e9ae87
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhemm.c
@@ -0,0 +1,91 @@
+/*
+ *
+ * cblas_zhemm.c
+ * This program is a C interface to zhemm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zhemm(const enum CBLAS_ORDER Order, const  enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *B, const int ldb, const void *beta,
+                 void *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhemm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zhemm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_zhemm(F77_SD, F77_UL, &F77_M, &F77_N, alpha, A, &F77_lda, 
+                     B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhemm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zhemm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_zhemm(F77_SD, F77_UL, &F77_N, &F77_M, alpha, A,
+                 &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zhemm", "Illegal Order setting, %d\n", Order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhemv.c b/ml/dlib/dlib/external/cblas/cblas_zhemv.c
new file mode 100644
index 000000000..c3cdb5958
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhemv.c
@@ -0,0 +1,146 @@
+/*
+ * cblas_zhemv.c
+ * The program is a C interface to zhemv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zhemv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo, const int N,
+                 const void *alpha, const void *A, const int lda,
+                 const void *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const double *xx= (double *)X, *alp= (double *)alpha, *bet = (double *)beta;
+   double ALPHA[2],BETA[2];
+   int tincY, tincx;
+   double *x=(double *)X, *y=(double *)Y, *st=0, *tx;
+
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhemv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_zhemv(F77_UL, &F77_N, alpha, A, &F77_lda, X, &F77_incX, 
+                beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1 ;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (double *) X;
+
+          
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhemv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_zhemv(F77_UL, &F77_N, ALPHA, A, &F77_lda, x, &F77_incX, 
+                BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_zhemv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor )
+   {
+      if ( X != x )
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+     }
+   }
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zher.c b/ml/dlib/dlib/external/cblas/cblas_zher.c
new file mode 100644
index 000000000..30453737c
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zher.c
@@ -0,0 +1,99 @@
+/*
+ * cblas_zher.c
+ * The program is a C interface to zher.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zher(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const void *X, const int incX
+                ,void *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+#endif
+   int n, i, tincx, incx=incX;
+   double *x=(double *)X, *xx=(double *)X, *tx, *st;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_zher(F77_UL, &F77_N, &alpha, X, &F77_incX, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+         tx = x;
+         if( incX > 0 ) {
+            i = incX << 1 ;
+            tincx = 2;
+            st= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            st = x-2; 
+            x +=(n-2); 
+         }
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+         #ifdef F77_INT
+           F77_incX = 1;
+         #else
+           incx = 1;
+         #endif
+      }
+      else x = (double *) X;
+      F77_zher(F77_UL, &F77_N, &alpha, x, &F77_incX, A, &F77_lda);
+   } else cblas_xerbla(1, "cblas_zher", "Illegal Order setting, %d\n", order);
+   if(X!=x) 
+      free(x);
+   
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zher2.c b/ml/dlib/dlib/external/cblas/cblas_zher2.c
new file mode 100644
index 000000000..8bf0bd733
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zher2.c
@@ -0,0 +1,140 @@
+/*
+ * cblas_zher2.c
+ * The program is a C interface to zher2.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zher2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const int N, const void *alpha, const void *X, const int incX,
+                 const void *Y, const int incY, void *A, const int lda)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incx
+   #define F77_incY incy
+#endif
+   int n, i, j, tincx, tincy, incx=incX, incy=incY;
+   double *x=(double *)X, *xx=(double *)X, *y=(double *)Y, 
+         *yy=(double *)Y, *tx, *ty, *stx, *sty;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher2", "Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_zher2(F77_UL, &F77_N, alpha, X, &F77_incX, 
+                                            Y, &F77_incY, A, &F77_lda);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher2", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+         y = malloc(n*sizeof(double));         
+         tx = x;
+         ty = y;
+         if( incX > 0 ) {
+            i = incX << 1 ;
+            tincx = 2;
+            stx= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            stx = x-2; 
+            x +=(n-2); 
+         }
+         
+         if( incY > 0 ) {
+            j = incY << 1;
+            tincy = 2;
+            sty= y+n;
+         } else { 
+            j = incY *(-2);
+            tincy = -2;
+            sty = y-2; 
+            y +=(n-2); 
+         }
+
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != stx);
+
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += tincy ;
+            yy += j;
+         }
+         while (y != sty);
+
+         x=tx;
+         y=ty;
+
+         #ifdef F77_INT
+            F77_incX = 1;
+            F77_incY = 1;
+         #else
+            incx = 1;
+            incy = 1;
+         #endif
+      }  else 
+      {
+         x = (double *) X;
+         y = (double *) Y;
+      }
+      F77_zher2(F77_UL, &F77_N, alpha, y, &F77_incY, x, 
+                                      &F77_incX, A, &F77_lda);
+   } 
+   else 
+   {
+      cblas_xerbla(1, "cblas_zher2", "Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+      free(x);
+   if(Y!=y)
+      free(y);
+
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zher2k.c b/ml/dlib/dlib/external/cblas/cblas_zher2k.c
new file mode 100644
index 000000000..96bcfe2a5
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zher2k.c
@@ -0,0 +1,95 @@
+/*
+ *
+ * cblas_zher2k.c
+ * This program is a C interface to zher2k.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zher2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void *A, const int lda,
+                  const void *B, const int ldb, const double beta,
+                  void *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+   double ALPHA[2]; 
+   const double *alp=(double *)alpha;
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zher2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zher2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zher2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zher2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      F77_zher2k(F77_UL,F77_TR, &F77_N, &F77_K, ALPHA, A, &F77_lda, B, &F77_ldb, &beta, C, &F77_ldc);
+   } else  cblas_xerbla(1, "cblas_zher2k", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zherk.c b/ml/dlib/dlib/external/cblas/cblas_zherk.c
new file mode 100644
index 000000000..bddef491b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zherk.c
@@ -0,0 +1,90 @@
+/*
+ *
+ * cblas_zherk.c
+ * This program is a C interface to zherk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zherk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const double alpha, const void *A, const int lda,
+                 const double beta, void *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zherk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zherk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zherk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                     &beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zherk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zherk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_zherk(F77_UL, F77_TR, &F77_N, &F77_K, &alpha, A, &F77_lda,
+                &beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zherk", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhpmv.c b/ml/dlib/dlib/external/cblas/cblas_zhpmv.c
new file mode 100644
index 000000000..1812884fc
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhpmv.c
@@ -0,0 +1,146 @@
+/*
+ * cblas_zhpmv.c
+ * The program is a C interface of zhpmv
+ * 
+ * Keita Teranishi  5/18/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zhpmv(const enum CBLAS_ORDER order,
+                 const enum CBLAS_UPLO Uplo,const int N,
+                 const void *alpha, const void  *AP,
+                 const void  *X, const int incX, const void *beta,
+                 void  *Y, const int incY)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incx
+   #define F77_incY incY
+#endif
+   int n, i=0, incx=incX;
+   const double *xx= (double *)X, *alp= (double *)alpha, *bet = (double *)beta;
+   double ALPHA[2],BETA[2];
+   int tincY, tincx;
+   double *x=(double *)X, *y=(double *)Y, *st=0, *tx;
+
+   if (order == CblasColMajor)
+   { 
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpmv","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      F77_zhpmv(F77_UL, &F77_N, alpha, AP, X,  
+                     &F77_incX, beta, Y, &F77_incY);
+   }
+   else if (order == CblasRowMajor)
+   {
+      ALPHA[0]= *alp;
+      ALPHA[1]= -alp[1];
+      BETA[0]= *bet;
+      BETA[1]= -bet[1];
+
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+ 
+         tx = x;
+         if( incX > 0 ) {
+           i = incX << 1;
+           tincx = 2;
+           st= x+n;
+         } else {
+           i = incX *(-2);
+           tincx = -2;
+           st = x-2;
+           x +=(n-2);
+         }
+
+         do
+         {
+           *x = *xx;
+           x[1] = -xx[1];
+           x += tincx ;
+           xx += i;
+         }
+         while (x != st);
+         x=tx;
+
+
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+ 
+         if(incY > 0)
+           tincY = incY;
+         else
+           tincY = -incY;
+         y++;
+
+         i = tincY << 1;
+         n = i * N ;
+         st = y + n;
+         do {
+            *y = -(*y);
+            y += i;
+         } while(y != st);
+         y -= n;
+      }  else
+         x = (double *) X;
+
+
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpmv","Illegal Uplo setting, %d\n", Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_zhpmv(F77_UL, &F77_N, ALPHA, 
+                     AP, x, &F77_incX, BETA, Y, &F77_incY);
+   }
+   else 
+   {
+      cblas_xerbla(1, "cblas_zhpmv","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if ( order == CblasRowMajor ) 
+   {
+      if(X!=x)
+         free(x);
+      if (N > 0)
+      {
+         do
+         {
+            *y = -(*y);
+            y += i;
+         }
+         while (y != st);
+     }
+  }
+
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhpr.c b/ml/dlib/dlib/external/cblas/cblas_zhpr.c
new file mode 100644
index 000000000..3ed2a8f61
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhpr.c
@@ -0,0 +1,102 @@
+/*
+ * cblas_zhpr.c
+ * The program is a C interface to zhpr.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zhpr(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                const int N, const double alpha, const void *X,
+                const int incX, void *A)
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incx
+#endif
+   int n, i, tincx, incx=incX;
+   double *x=(double *)X, *xx=(double *)X, *tx, *st;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpr","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_zhpr(F77_UL, &F77_N, &alpha, X, &F77_incX, A);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpr","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+         tx = x;
+         if( incX > 0 ) {
+            i = incX << 1;
+            tincx = 2;
+            st= x+n;
+         } else { 
+            i = incX *(-2);
+            tincx = -2;
+            st = x-2; 
+            x +=(n-2); 
+         }
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += tincx ;
+            xx += i;
+         }
+         while (x != st);
+         x=tx;
+         #ifdef F77_INT
+            F77_incX = 1;
+         #else
+            incx = 1;
+         #endif
+      }
+      else x = (double *) X;
+
+      F77_zhpr(F77_UL, &F77_N, &alpha, x, &F77_incX, A);
+
+   } else 
+   {
+      cblas_xerbla(1, "cblas_zhpr","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+     free(x);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zhpr2.c b/ml/dlib/dlib/external/cblas/cblas_zhpr2.c
new file mode 100644
index 000000000..0793a298a
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zhpr2.c
@@ -0,0 +1,137 @@
+/*
+ * cblas_zhpr2.c
+ * The program is a C interface to zhpr2.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zhpr2(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                      const int N,const void *alpha, const void *X, 
+                      const int incX,const void *Y, const int incY, void *Ap)
+
+{
+   char UL;
+#ifdef F77_CHAR
+   F77_CHAR F77_UL;
+#else
+   #define F77_UL &UL
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N,  F77_incX=incX, F77_incY=incY;
+#else
+   #define F77_N N
+   #define F77_incX incx
+   #define F77_incY incy
+#endif
+   int n, i, j, incx=incX, incy=incY;
+   double *x=(double *)X, *xx=(double *)X, *y=(double *)Y,
+         *yy=(double *)Y, *stx, *sty;
+
+ 
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasLower) UL = 'L';
+      else if (Uplo == CblasUpper) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpr2","Illegal Uplo setting, %d\n",Uplo );
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+
+      F77_zhpr2(F77_UL, &F77_N, alpha, X, &F77_incX, Y, &F77_incY, Ap);
+
+   }  else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zhpr2","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+      #endif
+      if (N > 0)
+      {
+         n = N << 1;
+         x = malloc(n*sizeof(double));
+         y = malloc(n*sizeof(double));         
+         stx = x + n;
+         sty = y + n;
+         if( incX > 0 )
+            i = incX << 1;
+         else
+            i = incX *(-2);
+ 
+         if( incY > 0 )
+            j = incY << 1;
+         else
+            j = incY *(-2);
+         do
+         {
+            *x = *xx;
+            x[1] = -xx[1];
+            x += 2;
+            xx += i;
+         } while (x != stx);
+         do
+         {
+            *y = *yy;
+            y[1] = -yy[1];
+            y += 2;
+            yy += j;
+         }
+         while (y != sty);
+         x -= n;
+         y -= n;
+
+         #ifdef F77_INT
+            if(incX > 0 )
+               F77_incX = 1;
+            else
+               F77_incX = -1;
+ 
+            if(incY > 0 )
+               F77_incY = 1;
+            else
+               F77_incY = -1;
+ 
+         #else
+            if(incX > 0 )
+               incx = 1;
+            else
+               incx = -1;
+ 
+            if(incY > 0 )
+               incy = 1;
+            else
+               incy = -1;
+         #endif
+
+      }  else 
+      {
+         x = (double *) X;
+         y = (void  *) Y;
+      }
+      F77_zhpr2(F77_UL, &F77_N, alpha, y, &F77_incY, x, &F77_incX, Ap);
+   } 
+   else 
+   {
+      cblas_xerbla(1, "cblas_zhpr2","Illegal Order setting, %d\n", order);
+      return;
+   }
+   if(X!=x)
+      free(x);
+   if(Y!=y)
+      free(y);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zscal.c b/ml/dlib/dlib/external/cblas/cblas_zscal.c
new file mode 100644
index 000000000..37b319f38
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zscal.c
@@ -0,0 +1,21 @@
+/*
+ * cblas_zscal.c
+ *
+ * The program is a C interface to zscal.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zscal( const int N, const void *alpha, void *X, 
+                       const int incX)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   F77_zscal( &F77_N, alpha, X, &F77_incX);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zswap.c b/ml/dlib/dlib/external/cblas/cblas_zswap.c
new file mode 100644
index 000000000..dfde2cbd0
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zswap.c
@@ -0,0 +1,22 @@
+/*
+ * cblas_zswap.c
+ *
+ * The program is a C interface to zswap.
+ *
+ * Written by Keita Teranishi.  2/11/1998
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zswap( const int N, void  *X, const int incX, void  *Y,
+                       const int incY)
+{
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX, F77_incY=incY;
+#else 
+   #define F77_N N
+   #define F77_incX incX
+   #define F77_incY incY
+#endif
+   F77_zswap( &F77_N, X, &F77_incX, Y, &F77_incY);
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zsymm.c b/ml/dlib/dlib/external/cblas/cblas_zsymm.c
new file mode 100644
index 000000000..85d5e3f49
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zsymm.c
@@ -0,0 +1,91 @@
+/*
+ *
+ * cblas_zsymm.c
+ * This program is a C interface to zsymm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zsymm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 const void  *B, const int ldb, const void *beta,
+                 void  *C, const int ldc)
+{
+   char SD, UL;   
+#ifdef F77_CHAR
+   F77_CHAR F77_SD, F77_UL;
+#else
+   #define F77_SD &SD  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zsymm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsymm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_zsymm(F77_SD, F77_UL, &F77_M, &F77_N, alpha, A, &F77_lda,
+                      B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zsymm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsymm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_SD = C2F_CHAR(&SD);
+      #endif
+
+      F77_zsymm(F77_SD, F77_UL, &F77_N, &F77_M, alpha, A, &F77_lda,
+                     B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zsymm", "Illegal Order setting, %d\n", Order);
+   return;
+} 
diff --git a/ml/dlib/dlib/external/cblas/cblas_zsyr2k.c b/ml/dlib/dlib/external/cblas/cblas_zsyr2k.c
new file mode 100644
index 000000000..ffac33462
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zsyr2k.c
@@ -0,0 +1,93 @@
+/*
+ *
+ * cblas_zsyr2k.c
+ * This program is a C interface to zsyr2k.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zsyr2k(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                  const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                  const void *alpha, const void  *A, const int lda,
+                  const void  *B, const int ldb, const void *beta,
+                  void  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda, F77_ldb=ldb;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldb ldb
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zsyr2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyr2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zsyr2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                      B, &F77_ldb, beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyr2k", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyr2k", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zsyr2k(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda, B, &F77_ldb, beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zsyr2k", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_zsyrk.c b/ml/dlib/dlib/external/cblas/cblas_zsyrk.c
new file mode 100644
index 000000000..45796074f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_zsyrk.c
@@ -0,0 +1,92 @@
+/*
+ *
+ * cblas_zsyrk.c
+ * This program is a C interface to zsyrk.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_zsyrk(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE Trans, const int N, const int K,
+                 const void *alpha, const void  *A, const int lda,
+                 const void *beta, void  *C, const int ldc)
+{
+   char UL, TR;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TR, F77_UL;
+#else
+   #define F77_TR &TR  
+   #define F77_UL &UL  
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_K=K, F77_lda=lda;
+   F77_INT F77_ldc=ldc;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_ldc ldc
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_zsyrk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( Trans == CblasTrans) TR ='T';
+      else if ( Trans == CblasConjTrans ) TR='C';
+      else if ( Trans == CblasNoTrans )   TR='N';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyrk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zsyrk(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                beta, C, &F77_ldc);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyrk", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if( Trans == CblasTrans) TR ='N';
+      else if ( Trans == CblasConjTrans ) TR='N';
+      else if ( Trans == CblasNoTrans )   TR='T';
+      else 
+      {
+         cblas_xerbla(3, "cblas_zsyrk", "Illegal Trans setting, %d\n", Trans);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TR = C2F_CHAR(&TR);
+      #endif
+
+      F77_zsyrk(F77_UL, F77_TR, &F77_N, &F77_K, alpha, A, &F77_lda,
+                     beta, C, &F77_ldc);
+   } 
+   else  cblas_xerbla(1, "cblas_zsyrk", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztbmv.c b/ml/dlib/dlib/external/cblas/cblas_ztbmv.c
new file mode 100644
index 000000000..916774d27
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztbmv.c
@@ -0,0 +1,139 @@
+/*
+ * cblas_ztbmv.c
+ * The program is a C interface to ztbmv.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztbmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void  *A, const int lda,
+                 void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0, *x=(double *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztbmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            x++;
+            st = x + n;
+            do
+            {
+               *x = -(*x);
+               x+= i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztbmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztbmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztbmv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztbmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztbsv.c b/ml/dlib/dlib/external/cblas/cblas_ztbsv.c
new file mode 100644
index 000000000..cc5d3f73f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztbsv.c
@@ -0,0 +1,143 @@
+/*
+ * cblas_ztbsv.c
+ * The program is a C interface to ztbsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztbsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const int K, const void  *A, const int lda,
+                 void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_K=K, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_K K
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0,*x=(double *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztbsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+  
+            x++;
+
+            st=x+n;
+
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztbsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztbsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztbsv( F77_UL, F77_TA, F77_DI, &F77_N, &F77_K, A, &F77_lda, X,
+                      &F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x+= i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztbsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztpmv.c b/ml/dlib/dlib/external/cblas/cblas_ztpmv.c
new file mode 100644
index 000000000..2e7949a25
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztpmv.c
@@ -0,0 +1,133 @@
+/*
+ * cblas_ztpmv.c
+ * The program is a C interface to ztpmv.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztpmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *Ap, void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0,*x=(double *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztpmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            x++;
+            st = x + n;
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztpmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztpmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztpmv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztpmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztpsv.c b/ml/dlib/dlib/external/cblas/cblas_ztpsv.c
new file mode 100644
index 000000000..c41d02016
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztpsv.c
@@ -0,0 +1,138 @@
+/*
+ * cblas_ztpsv.c
+ * The program is a C interface to ztpsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztpsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *Ap, void  *X, const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0, *x=(double*)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X, &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztpsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+ 
+            x++;
+
+            st=x+n;
+
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztpsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztpsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztpsv( F77_UL, F77_TA, F77_DI, &F77_N, Ap, X,&F77_incX);
+
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztpsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztrmm.c b/ml/dlib/dlib/external/cblas/cblas_ztrmm.c
new file mode 100644
index 000000000..4e76377d9
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztrmm.c
@@ -0,0 +1,126 @@
+/*
+ *
+ * cblas_ztrmm.c
+ * This program is a C interface to ztrmm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztrmm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const  enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 void  *B, const int ldb)
+{
+   char UL, TA, SD, DI;   
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA  
+   #define F77_UL &UL  
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+      if( Side == CblasRight ) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrmm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+      if( Uplo == CblasUpper ) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrmm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans ) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrmm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ztrmm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, alpha, A, &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+      if( Side == CblasRight ) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrmm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper ) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrmm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans ) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrmm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ztrmm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztrmm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, alpha, A, &F77_lda, B, &F77_ldb);
+   } 
+   else  cblas_xerbla(1, "cblas_ztrmm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztrmv.c b/ml/dlib/dlib/external/cblas/cblas_ztrmv.c
new file mode 100644
index 000000000..65e760529
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztrmv.c
@@ -0,0 +1,137 @@
+/*
+ * cblas_ztrmv.c
+ * The program is a C interface to ztrmv.
+ * 
+ * Keita Teranishi  5/20/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztrmv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *A, const int lda,
+                 void  *X, const int incX)
+
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0,*x=(double *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrmv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if(incX > 0)
+               tincX = incX;
+            else
+               tincX = -incX;
+            i = tincX << 1;
+            n = i * N;
+            x++;
+            st = x + n;
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrmv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrmv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+         F77_ztrmv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztrmv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztrsm.c b/ml/dlib/dlib/external/cblas/cblas_ztrsm.c
new file mode 100644
index 000000000..7540147c8
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztrsm.c
@@ -0,0 +1,132 @@
+/*
+ *
+ * cblas_ztrsm.c
+ * This program is a C interface to ztrsm.
+ * Written by Keita Teranishi
+ * 4/8/1998
+ *
+ */
+
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
+                 const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
+                 const enum CBLAS_DIAG Diag, const int M, const int N,
+                 const void *alpha, const void  *A, const int lda,
+                 void  *B, const int ldb)
+{
+   char UL, TA, SD, DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_SD, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_SD &SD
+   #define F77_DI &DI
+#endif
+
+#ifdef F77_INT
+   F77_INT F77_M=M, F77_N=N, F77_lda=lda, F77_ldb=ldb;
+#else
+   #define F77_M M
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_ldb ldb
+#endif
+
+
+   if( Order == CblasColMajor )
+   {
+
+      if( Side == CblasRight) SD='R';
+      else if ( Side == CblasLeft ) SD='L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='U';
+      else if ( Uplo == CblasLower ) UL='L';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ztrsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+      F77_ztrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_M, &F77_N, alpha, A,
+                &F77_lda, B, &F77_ldb);
+   } else if (Order == CblasRowMajor)
+   {
+
+      if( Side == CblasRight) SD='L';
+      else if ( Side == CblasLeft ) SD='R';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrsm", "Illegal Side setting, %d\n", Side);
+         return;
+      }
+
+      if( Uplo == CblasUpper) UL='L';
+      else if ( Uplo == CblasLower ) UL='U';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrsm", "Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if( TransA == CblasTrans) TA ='T';
+      else if ( TransA == CblasConjTrans ) TA='C';
+      else if ( TransA == CblasNoTrans )   TA='N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrsm", "Illegal Trans setting, %d\n", TransA);
+         return;
+      }
+
+      if( Diag == CblasUnit ) DI='U';
+      else if ( Diag == CblasNonUnit ) DI='N';
+      else 
+      {
+         cblas_xerbla(5, "cblas_ztrsm", "Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_SD = C2F_CHAR(&SD);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+
+
+      F77_ztrsm(F77_SD, F77_UL, F77_TA, F77_DI, &F77_N, &F77_M, alpha, A,
+                &F77_lda, B, &F77_ldb);
+   } 
+   else  cblas_xerbla(1, "cblas_ztrsm", "Illegal Order setting, %d\n", Order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cblas_ztrsv.c b/ml/dlib/dlib/external/cblas/cblas_ztrsv.c
new file mode 100644
index 000000000..07e2653b4
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cblas_ztrsv.c
@@ -0,0 +1,137 @@
+/*
+ * cblas_ztrsv.c
+ * The program is a C interface to ztrsv.
+ * 
+ * Keita Teranishi  3/23/98
+ *
+ */
+#include "cblas.h"
+#include "cblas_f77.h"
+void cblas_ztrsv(const enum CBLAS_ORDER order, const enum CBLAS_UPLO Uplo,
+                 const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_DIAG Diag,
+                 const int N, const void  *A, const int lda, void  *X,
+                 const int incX)
+{
+   char TA;
+   char UL;
+   char DI;
+#ifdef F77_CHAR
+   F77_CHAR F77_TA, F77_UL, F77_DI;
+#else
+   #define F77_TA &TA
+   #define F77_UL &UL
+   #define F77_DI &DI   
+#endif
+#ifdef F77_INT
+   F77_INT F77_N=N, F77_lda=lda, F77_incX=incX;
+#else
+   #define F77_N N
+   #define F77_lda lda
+   #define F77_incX incX
+#endif
+   int n, i=0, tincX; 
+   double *st=0,*x=(double *)X;
+
+   if (order == CblasColMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'U';
+      else if (Uplo == CblasLower) UL = 'L';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+      if (TransA == CblasNoTrans) TA = 'N';
+      else if (TransA == CblasTrans) TA = 'T';
+      else if (TransA == CblasConjTrans) TA = 'C';
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+   }
+   else if (order == CblasRowMajor)
+   {
+      if (Uplo == CblasUpper) UL = 'L';
+      else if (Uplo == CblasLower) UL = 'U';
+      else 
+      {
+         cblas_xerbla(2, "cblas_ztrsv","Illegal Uplo setting, %d\n", Uplo);
+         return;
+      }
+
+      if (TransA == CblasNoTrans) TA = 'T';
+      else if (TransA == CblasTrans) TA = 'N';
+      else if (TransA == CblasConjTrans)
+      {
+         TA = 'N';
+         if ( N > 0)
+         {
+            if ( incX > 0 )
+               tincX = incX;
+            else
+               tincX = -incX;
+ 
+            n = N*2*(tincX);
+            x++;
+            st=x+n; 
+            i = tincX << 1;
+            do
+            {
+               *x = -(*x);
+               x+=i;
+            }
+            while (x != st);
+            x -= n;
+         }
+      }
+      else 
+      {
+         cblas_xerbla(3, "cblas_ztrsv","Illegal TransA setting, %d\n", TransA);
+         return;
+      }
+
+      if (Diag == CblasUnit) DI = 'U';
+      else if (Diag == CblasNonUnit) DI = 'N';
+      else 
+      {
+         cblas_xerbla(4, "cblas_ztrsv","Illegal Diag setting, %d\n", Diag);
+         return;
+      }
+      #ifdef F77_CHAR
+         F77_UL = C2F_CHAR(&UL);
+         F77_TA = C2F_CHAR(&TA);
+         F77_DI = C2F_CHAR(&DI);
+      #endif
+      F77_ztrsv( F77_UL, F77_TA, F77_DI, &F77_N, A, &F77_lda, X,
+                      &F77_incX);
+      if (TransA == CblasConjTrans)
+      {
+         if (N > 0)
+         {
+            do
+            {
+               *x = -(*x);
+               x += i;
+            }
+            while (x != st);
+         }
+      }
+   }
+   else cblas_xerbla(1, "cblas_ztrsv", "Illegal Order setting, %d\n", order);
+   return;
+}
diff --git a/ml/dlib/dlib/external/cblas/cdotcsub.f b/ml/dlib/dlib/external/cblas/cdotcsub.f
new file mode 100644
index 000000000..f97d7159e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cdotcsub.f
@@ -0,0 +1,15 @@
+c     cdotcsub.f
+c
+c     The program is a fortran wrapper for cdotc.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine cdotcsub(n,x,incx,y,incy,dotc)
+c
+      external cdotc
+      complex cdotc,dotc
+      integer n,incx,incy
+      complex x(*),y(*)
+c
+      dotc=cdotc(n,x,incx,y,incy)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/cdotusub.f b/ml/dlib/dlib/external/cblas/cdotusub.f
new file mode 100644
index 000000000..5107c0402
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/cdotusub.f
@@ -0,0 +1,15 @@
+c     cdotusub.f
+c
+c     The program is a fortran wrapper for cdotu.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine cdotusub(n,x,incx,y,incy,dotu)
+c
+      external cdotu
+      complex cdotu,dotu
+      integer n,incx,incy
+      complex x(*),y(*)
+c
+      dotu=cdotu(n,x,incx,y,incy)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/dasumsub.f b/ml/dlib/dlib/external/cblas/dasumsub.f
new file mode 100644
index 000000000..3d64d17e6
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/dasumsub.f
@@ -0,0 +1,15 @@
+c     dasumsun.f
+c
+c     The program is a fortran wrapper for dasum..
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine dasumsub(n,x,incx,asum)
+c
+      external dasum
+      double precision dasum,asum
+      integer n,incx
+      double precision x(*)
+c
+      asum=dasum(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/ddotsub.f b/ml/dlib/dlib/external/cblas/ddotsub.f
new file mode 100644
index 000000000..205f3b46f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/ddotsub.f
@@ -0,0 +1,15 @@
+c     ddotsub.f
+c
+c     The program is a fortran wrapper for ddot.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine ddotsub(n,x,incx,y,incy,dot)
+c
+      external ddot
+      double precision ddot
+      integer n,incx,incy
+      double precision x(*),y(*),dot
+c
+      dot=ddot(n,x,incx,y,incy)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/dnrm2sub.f b/ml/dlib/dlib/external/cblas/dnrm2sub.f
new file mode 100644
index 000000000..88f17db8b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/dnrm2sub.f
@@ -0,0 +1,15 @@
+c     dnrm2sub.f
+c
+c     The program is a fortran wrapper for dnrm2.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine dnrm2sub(n,x,incx,nrm2)
+c
+      external dnrm2
+      double precision dnrm2,nrm2
+      integer n,incx
+      double precision x(*)
+c
+      nrm2=dnrm2(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/dsdotsub.f b/ml/dlib/dlib/external/cblas/dsdotsub.f
new file mode 100644
index 000000000..e7e872c9e
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/dsdotsub.f
@@ -0,0 +1,15 @@
+c     dsdotsub.f
+c
+c     The program is a fortran wrapper for dsdot.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine dsdotsub(n,x,incx,y,incy,dot)
+c
+      external dsdot
+      double precision dsdot,dot
+      integer n,incx,incy
+      real x(*),y(*)
+c
+      dot=dsdot(n,x,incx,y,incy)
+      return
+      end      
diff --git a/ml/dlib/dlib/external/cblas/dzasumsub.f b/ml/dlib/dlib/external/cblas/dzasumsub.f
new file mode 100644
index 000000000..9aaf16387
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/dzasumsub.f
@@ -0,0 +1,15 @@
+c     dzasumsub.f
+c
+c     The program is a fortran wrapper for dzasum.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine dzasumsub(n,x,incx,asum)
+c
+      external dzasum
+      double precision dzasum,asum
+      integer n,incx
+      double complex x(*)
+c
+      asum=dzasum(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/dznrm2sub.f b/ml/dlib/dlib/external/cblas/dznrm2sub.f
new file mode 100644
index 000000000..45dc599f8
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/dznrm2sub.f
@@ -0,0 +1,15 @@
+c     dznrm2sub.f
+c
+c     The program is a fortran wrapper for dznrm2.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine dznrm2sub(n,x,incx,nrm2)
+c
+      external dznrm2
+      double precision dznrm2,nrm2
+      integer n,incx
+      double complex x(*)
+c
+      nrm2=dznrm2(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/icamaxsub.f b/ml/dlib/dlib/external/cblas/icamaxsub.f
new file mode 100644
index 000000000..3f47071eb
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/icamaxsub.f
@@ -0,0 +1,15 @@
+c     icamaxsub.f
+c
+c     The program is a fortran wrapper for icamax.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine icamaxsub(n,x,incx,iamax)
+c
+      external icamax
+      integer  icamax,iamax
+      integer n,incx
+      complex x(*)
+c
+      iamax=icamax(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/idamaxsub.f b/ml/dlib/dlib/external/cblas/idamaxsub.f
new file mode 100644
index 000000000..3c1ee5c32
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/idamaxsub.f
@@ -0,0 +1,15 @@
+c     icamaxsub.f
+c
+c     The program is a fortran wrapper for idamax.
+c     Witten by Keita Teranishi.  2/22/1998
+c
+      subroutine idamaxsub(n,x,incx,iamax)
+c
+      external idamax
+      integer  idamax,iamax
+      integer n,incx
+      double precision x(*)
+c
+      iamax=idamax(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/isamaxsub.f b/ml/dlib/dlib/external/cblas/isamaxsub.f
new file mode 100644
index 000000000..0faf42fde
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/isamaxsub.f
@@ -0,0 +1,15 @@
+c     isamaxsub.f
+c
+c     The program is a fortran wrapper for isamax.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine isamaxsub(n,x,incx,iamax)
+c
+      external isamax
+      integer  isamax,iamax
+      integer n,incx
+      real x(*)
+c
+      iamax=isamax(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/izamaxsub.f b/ml/dlib/dlib/external/cblas/izamaxsub.f
new file mode 100644
index 000000000..5b15855a7
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/izamaxsub.f
@@ -0,0 +1,15 @@
+c     izamaxsub.f
+c
+c     The program is a fortran wrapper for izamax.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine izamaxsub(n,x,incx,iamax)
+c
+      external izamax
+      integer  izamax,iamax
+      integer n,incx
+      double complex x(*)
+c
+      iamax=izamax(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/sasumsub.f b/ml/dlib/dlib/external/cblas/sasumsub.f
new file mode 100644
index 000000000..955f11e8d
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/sasumsub.f
@@ -0,0 +1,15 @@
+c     sasumsub.f
+c
+c     The program is a fortran wrapper for sasum.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine sasumsub(n,x,incx,asum)
+c
+      external sasum
+      real sasum,asum
+      integer n,incx
+      real x(*)
+c
+      asum=sasum(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/scasumsub.f b/ml/dlib/dlib/external/cblas/scasumsub.f
new file mode 100644
index 000000000..077ace670
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/scasumsub.f
@@ -0,0 +1,15 @@
+c     scasumsub.f
+c
+c     The program is a fortran wrapper for scasum.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine scasumsub(n,x,incx,asum)
+c
+      external scasum
+      real scasum,asum
+      integer n,incx
+      complex x(*)
+c
+      asum=scasum(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/scnrm2sub.f b/ml/dlib/dlib/external/cblas/scnrm2sub.f
new file mode 100644
index 000000000..7242c9742
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/scnrm2sub.f
@@ -0,0 +1,15 @@
+c     scnrm2sub.f
+c
+c     The program is a fortran wrapper for scnrm2.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine scnrm2sub(n,x,incx,nrm2)
+c
+      external scnrm2
+      real scnrm2,nrm2
+      integer n,incx
+      complex x(*)
+c
+      nrm2=scnrm2(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/sdotsub.f b/ml/dlib/dlib/external/cblas/sdotsub.f
new file mode 100644
index 000000000..e1af3c97b
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/sdotsub.f
@@ -0,0 +1,15 @@
+c     sdotsub.f
+c
+c     The program is a fortran wrapper for sdot.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine sdotsub(n,x,incx,y,incy,dot)
+c
+      external sdot
+      real sdot
+      integer n,incx,incy
+      real x(*),y(*),dot
+c
+      dot=sdot(n,x,incx,y,incy)
+      return
+      end      
diff --git a/ml/dlib/dlib/external/cblas/sdsdotsub.f b/ml/dlib/dlib/external/cblas/sdsdotsub.f
new file mode 100644
index 000000000..80008e9ce
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/sdsdotsub.f
@@ -0,0 +1,15 @@
+c     sdsdotsub.f
+c
+c     The program is a fortran wrapper for sdsdot.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine sdsdotsub(n,x,incx,y,incy,dot)
+c
+      external sdsdot
+      real sdsdot,dot
+      integer n,incx,incy
+      real x(*),y(*)
+c
+      dot=sdsdot(n,x,incx,y,incy)
+      return
+      end      
diff --git a/ml/dlib/dlib/external/cblas/snrm2sub.f b/ml/dlib/dlib/external/cblas/snrm2sub.f
new file mode 100644
index 000000000..871a6e49f
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/snrm2sub.f
@@ -0,0 +1,15 @@
+c     snrm2sub.f
+c
+c     The program is a fortran wrapper for snrm2.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine snrm2sub(n,x,incx,nrm2)
+c
+      external snrm2
+      real snrm2,nrm2
+      integer n,incx
+      real x(*)
+c
+      nrm2=snrm2(n,x,incx)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/zdotcsub.f b/ml/dlib/dlib/external/cblas/zdotcsub.f
new file mode 100644
index 000000000..8d483c895
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/zdotcsub.f
@@ -0,0 +1,15 @@
+c     zdotcsub.f
+c
+c     The program is a fortran wrapper for zdotc.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine zdotcsub(n,x,incx,y,incy,dotc)
+c
+      external zdotc
+      double complex zdotc,dotc
+      integer n,incx,incy
+      double complex x(*),y(*)
+c
+      dotc=zdotc(n,x,incx,y,incy)
+      return
+      end
diff --git a/ml/dlib/dlib/external/cblas/zdotusub.f b/ml/dlib/dlib/external/cblas/zdotusub.f
new file mode 100644
index 000000000..23f32dec3
--- /dev/null
+++ b/ml/dlib/dlib/external/cblas/zdotusub.f
@@ -0,0 +1,15 @@
+c     zdotusub.f
+c
+c     The program is a fortran wrapper for zdotu.
+c     Witten by Keita Teranishi.  2/11/1998
+c
+      subroutine zdotusub(n,x,incx,y,incy,dotu)
+c
+      external zdotu
+      double complex zdotu,dotu
+      integer n,incx,incy
+      double complex x(*),y(*)
+c
+      dotu=zdotu(n,x,incx,y,incy)
+      return
+      end
diff --git a/ml/dlib/dlib/external/libjpeg/README b/ml/dlib/dlib/external/libjpeg/README
new file mode 100644
index 000000000..86cc20669
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/README
@@ -0,0 +1,385 @@
+The Independent JPEG Group's JPEG software
+==========================================
+
+README for release 6b of 27-Mar-1998
+====================================
+
+This distribution contains the sixth public release of the Independent JPEG
+Group's free JPEG software.  You are welcome to redistribute this software and
+to use it for any purpose, subject to the conditions under LEGAL ISSUES, below.
+
+Serious users of this software (particularly those incorporating it into
+larger programs) should contact IJG at jpeg-info@uunet.uu.net to be added to
+our electronic mailing list.  Mailing list members are notified of updates
+and have a chance to participate in technical discussions, etc.
+
+This software is the work of Tom Lane, Philip Gladstone, Jim Boucher,
+Lee Crocker, Julian Minguillon, Luis Ortiz, George Phillips, Davide Rossi,
+Guido Vollbeding, Ge' Weijers, and other members of the Independent JPEG
+Group.
+
+IJG is not affiliated with the official ISO JPEG standards committee.
+
+
+DOCUMENTATION ROADMAP
+=====================
+
+This file contains the following sections:
+
+OVERVIEW            General description of JPEG and the IJG software.
+LEGAL ISSUES        Copyright, lack of warranty, terms of distribution.
+REFERENCES          Where to learn more about JPEG.
+ARCHIVE LOCATIONS   Where to find newer versions of this software.
+RELATED SOFTWARE    Other stuff you should get.
+FILE FORMAT WARS    Software *not* to get.
+TO DO               Plans for future IJG releases.
+
+Other documentation files in the distribution are:
+
+User documentation:
+  install.doc       How to configure and install the IJG software.
+  usage.doc         Usage instructions for cjpeg, djpeg, jpegtran,
+                    rdjpgcom, and wrjpgcom.
+  *.1               Unix-style man pages for programs (same info as usage.doc).
+  wizard.doc        Advanced usage instructions for JPEG wizards only.
+  change.log        Version-to-version change highlights.
+Programmer and internal documentation:
+  libjpeg.doc       How to use the JPEG library in your own programs.
+  example.c         Sample code for calling the JPEG library.
+  structure.doc     Overview of the JPEG library's internal structure.
+  filelist.doc      Road map of IJG files.
+  coderules.doc     Coding style rules --- please read if you contribute code.
+
+Please read at least the files install.doc and usage.doc.  Useful information
+can also be found in the JPEG FAQ (Frequently Asked Questions) article.  See
+ARCHIVE LOCATIONS below to find out where to obtain the FAQ article.
+
+If you want to understand how the JPEG code works, we suggest reading one or
+more of the REFERENCES, then looking at the documentation files (in roughly
+the order listed) before diving into the code.
+
+
+OVERVIEW
+========
+
+This package contains C software to implement JPEG image compression and
+decompression.  JPEG (pronounced "jay-peg") is a standardized compression
+method for full-color and gray-scale images.  JPEG is intended for compressing
+"real-world" scenes; line drawings, cartoons and other non-realistic images
+are not its strong suit.  JPEG is lossy, meaning that the output image is not
+exactly identical to the input image.  Hence you must not use JPEG if you
+have to have identical output bits.  However, on typical photographic images,
+very good compression levels can be obtained with no visible change, and
+remarkably high compression levels are possible if you can tolerate a
+low-quality image.  For more details, see the references, or just experiment
+with various compression settings.
+
+This software implements JPEG baseline, extended-sequential, and progressive
+compression processes.  Provision is made for supporting all variants of these
+processes, although some uncommon parameter settings aren't implemented yet.
+For legal reasons, we are not distributing code for the arithmetic-coding
+variants of JPEG; see LEGAL ISSUES.  We have made no provision for supporting
+the hierarchical or lossless processes defined in the standard.
+
+We provide a set of library routines for reading and writing JPEG image files,
+plus two sample applications "cjpeg" and "djpeg", which use the library to
+perform conversion between JPEG and some other popular image file formats.
+The library is intended to be reused in other applications.
+
+In order to support file conversion and viewing software, we have included
+considerable functionality beyond the bare JPEG coding/decoding capability;
+for example, the color quantization modules are not strictly part of JPEG
+decoding, but they are essential for output to colormapped file formats or
+colormapped displays.  These extra functions can be compiled out of the
+library if not required for a particular application.  We have also included
+"jpegtran", a utility for lossless transcoding between different JPEG
+processes, and "rdjpgcom" and "wrjpgcom", two simple applications for
+inserting and extracting textual comments in JFIF files.
+
+The emphasis in designing this software has been on achieving portability and
+flexibility, while also making it fast enough to be useful.  In particular,
+the software is not intended to be read as a tutorial on JPEG.  (See the
+REFERENCES section for introductory material.)  Rather, it is intended to
+be reliable, portable, industrial-strength code.  We do not claim to have
+achieved that goal in every aspect of the software, but we strive for it.
+
+We welcome the use of this software as a component of commercial products.
+No royalty is required, but we do ask for an acknowledgement in product
+documentation, as described under LEGAL ISSUES.
+
+
+LEGAL ISSUES
+============
+
+In plain English:
+
+1. We don't promise that this software works.  (But if you find any bugs,
+   please let us know!)
+2. You can use this software for whatever you want.  You don't have to pay us.
+3. You may not pretend that you wrote this software.  If you use it in a
+   program, you must acknowledge somewhere in your documentation that
+   you've used the IJG code.
+
+In legalese:
+
+The authors make NO WARRANTY or representation, either express or implied,
+with respect to this software, its quality, accuracy, merchantability, or
+fitness for a particular purpose.  This software is provided "AS IS", and you,
+its user, assume the entire risk as to its quality and accuracy.
+
+This software is copyright (C) 1991-1998, Thomas G. Lane.
+All Rights Reserved except as specified below.
+
+Permission is hereby granted to use, copy, modify, and distribute this
+software (or portions thereof) for any purpose, without fee, subject to these
+conditions:
+(1) If any part of the source code for this software is distributed, then this
+README file must be included, with this copyright and no-warranty notice
+unaltered; and any additions, deletions, or changes to the original files
+must be clearly indicated in accompanying documentation.
+(2) If only executable code is distributed, then the accompanying
+documentation must state that "this software is based in part on the work of
+the Independent JPEG Group".
+(3) Permission for use of this software is granted only if the user accepts
+full responsibility for any undesirable consequences; the authors accept
+NO LIABILITY for damages of any kind.
+
+These conditions apply to any software derived from or based on the IJG code,
+not just to the unmodified library.  If you use our work, you ought to
+acknowledge us.
+
+Permission is NOT granted for the use of any IJG author's name or company name
+in advertising or publicity relating to this software or products derived from
+it.  This software may be referred to only as "the Independent JPEG Group's
+software".
+
+We specifically permit and encourage the use of this software as the basis of
+commercial products, provided that all warranty or liability claims are
+assumed by the product vendor.
+
+
+ansi2knr.c is included in this distribution by permission of L. Peter Deutsch,
+sole proprietor of its copyright holder, Aladdin Enterprises of Menlo Park, CA.
+ansi2knr.c is NOT covered by the above copyright and conditions, but instead
+by the usual distribution terms of the Free Software Foundation; principally,
+that you must include source code if you redistribute it.  (See the file
+ansi2knr.c for full details.)  However, since ansi2knr.c is not needed as part
+of any program generated from the IJG code, this does not limit you more than
+the foregoing paragraphs do.
+
+The Unix configuration script "configure" was produced with GNU Autoconf.
+It is copyright by the Free Software Foundation but is freely distributable.
+The same holds for its supporting scripts (config.guess, config.sub,
+ltconfig, ltmain.sh).  Another support script, install-sh, is copyright
+by M.I.T. but is also freely distributable.
+
+It appears that the arithmetic coding option of the JPEG spec is covered by
+patents owned by IBM, AT&T, and Mitsubishi.  Hence arithmetic coding cannot
+legally be used without obtaining one or more licenses.  For this reason,
+support for arithmetic coding has been removed from the free JPEG software.
+(Since arithmetic coding provides only a marginal gain over the unpatented
+Huffman mode, it is unlikely that very many implementations will support it.)
+So far as we are aware, there are no patent restrictions on the remaining
+code.
+
+The IJG distribution formerly included code to read and write GIF files.
+To avoid entanglement with the Unisys LZW patent, GIF reading support has
+been removed altogether, and the GIF writer has been simplified to produce
+"uncompressed GIFs".  This technique does not use the LZW algorithm; the
+resulting GIF files are larger than usual, but are readable by all standard
+GIF decoders.
+
+We are required to state that
+    "The Graphics Interchange Format(c) is the Copyright property of
+    CompuServe Incorporated.  GIF(sm) is a Service Mark property of
+    CompuServe Incorporated."
+
+
+REFERENCES
+==========
+
+We highly recommend reading one or more of these references before trying to
+understand the innards of the JPEG software.
+
+The best short technical introduction to the JPEG compression algorithm is
+	Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+	Communications of the ACM, April 1991 (vol. 34 no. 4), pp. 30-44.
+(Adjacent articles in that issue discuss MPEG motion picture compression,
+applications of JPEG, and related topics.)  If you don't have the CACM issue
+handy, a PostScript file containing a revised version of Wallace's article is
+available at ftp://ftp.uu.net/graphics/jpeg/wallace.ps.gz.  The file (actually
+a preprint for an article that appeared in IEEE Trans. Consumer Electronics)
+omits the sample images that appeared in CACM, but it includes corrections
+and some added material.  Note: the Wallace article is copyright ACM and IEEE,
+and it may not be used for commercial purposes.
+
+A somewhat less technical, more leisurely introduction to JPEG can be found in
+"The Data Compression Book" by Mark Nelson and Jean-loup Gailly, published by
+M&T Books (New York), 2nd ed. 1996, ISBN 1-55851-434-1.  This book provides
+good explanations and example C code for a multitude of compression methods
+including JPEG.  It is an excellent source if you are comfortable reading C
+code but don't know much about data compression in general.  The book's JPEG
+sample code is far from industrial-strength, but when you are ready to look
+at a full implementation, you've got one here...
+
+The best full description of JPEG is the textbook "JPEG Still Image Data
+Compression Standard" by William B. Pennebaker and Joan L. Mitchell, published
+by Van Nostrand Reinhold, 1993, ISBN 0-442-01272-1.  Price US$59.95, 638 pp.
+The book includes the complete text of the ISO JPEG standards (DIS 10918-1
+and draft DIS 10918-2).  This is by far the most complete exposition of JPEG
+in existence, and we highly recommend it.
+
+The JPEG standard itself is not available electronically; you must order a
+paper copy through ISO or ITU.  (Unless you feel a need to own a certified
+official copy, we recommend buying the Pennebaker and Mitchell book instead;
+it's much cheaper and includes a great deal of useful explanatory material.)
+In the USA, copies of the standard may be ordered from ANSI Sales at (212)
+642-4900, or from Global Engineering Documents at (800) 854-7179.  (ANSI
+doesn't take credit card orders, but Global does.)  It's not cheap: as of
+1992, ANSI was charging $95 for Part 1 and $47 for Part 2, plus 7%
+shipping/handling.  The standard is divided into two parts, Part 1 being the
+actual specification, while Part 2 covers compliance testing methods.  Part 1
+is titled "Digital Compression and Coding of Continuous-tone Still Images,
+Part 1: Requirements and guidelines" and has document numbers ISO/IEC IS
+10918-1, ITU-T T.81.  Part 2 is titled "Digital Compression and Coding of
+Continuous-tone Still Images, Part 2: Compliance testing" and has document
+numbers ISO/IEC IS 10918-2, ITU-T T.83.
+
+Some extensions to the original JPEG standard are defined in JPEG Part 3,
+a newer ISO standard numbered ISO/IEC IS 10918-3 and ITU-T T.84.  IJG
+currently does not support any Part 3 extensions.
+
+The JPEG standard does not specify all details of an interchangeable file
+format.  For the omitted details we follow the "JFIF" conventions, revision
+1.02.  A copy of the JFIF spec is available from:
+	Literature Department
+	C-Cube Microsystems, Inc.
+	1778 McCarthy Blvd.
+	Milpitas, CA 95035
+	phone (408) 944-6300,  fax (408) 944-6314
+A PostScript version of this document is available by FTP at
+ftp://ftp.uu.net/graphics/jpeg/jfif.ps.gz.  There is also a plain text
+version at ftp://ftp.uu.net/graphics/jpeg/jfif.txt.gz, but it is missing
+the figures.
+
+The TIFF 6.0 file format specification can be obtained by FTP from
+ftp://ftp.sgi.com/graphics/tiff/TIFF6.ps.gz.  The JPEG incorporation scheme
+found in the TIFF 6.0 spec of 3-June-92 has a number of serious problems.
+IJG does not recommend use of the TIFF 6.0 design (TIFF Compression tag 6).
+Instead, we recommend the JPEG design proposed by TIFF Technical Note #2
+(Compression tag 7).  Copies of this Note can be obtained from ftp.sgi.com or
+from ftp://ftp.uu.net/graphics/jpeg/.  It is expected that the next revision
+of the TIFF spec will replace the 6.0 JPEG design with the Note's design.
+Although IJG's own code does not support TIFF/JPEG, the free libtiff library
+uses our library to implement TIFF/JPEG per the Note.  libtiff is available
+from ftp://ftp.sgi.com/graphics/tiff/.
+
+
+ARCHIVE LOCATIONS
+=================
+
+The "official" archive site for this software is ftp.uu.net (Internet
+address 192.48.96.9).  The most recent released version can always be found
+there in directory graphics/jpeg.  This particular version will be archived
+as ftp://ftp.uu.net/graphics/jpeg/jpegsrc.v6b.tar.gz.  If you don't have
+direct Internet access, UUNET's archives are also available via UUCP; contact
+help@uunet.uu.net for information on retrieving files that way.
+
+Numerous Internet sites maintain copies of the UUNET files.  However, only
+ftp.uu.net is guaranteed to have the latest official version.
+
+You can also obtain this software in DOS-compatible "zip" archive format from
+the SimTel archives (ftp://ftp.simtel.net/pub/simtelnet/msdos/graphics/), or
+on CompuServe in the Graphics Support forum (GO CIS:GRAPHSUP), library 12
+"JPEG Tools".  Again, these versions may sometimes lag behind the ftp.uu.net
+release.
+
+The JPEG FAQ (Frequently Asked Questions) article is a useful source of
+general information about JPEG.  It is updated constantly and therefore is
+not included in this distribution.  The FAQ is posted every two weeks to
+Usenet newsgroups comp.graphics.misc, news.answers, and other groups.
+It is available on the World Wide Web at http://www.faqs.org/faqs/jpeg-faq/
+and other news.answers archive sites, including the official news.answers
+archive at rtfm.mit.edu: ftp://rtfm.mit.edu/pub/usenet/news.answers/jpeg-faq/.
+If you don't have Web or FTP access, send e-mail to mail-server@rtfm.mit.edu
+with body
+	send usenet/news.answers/jpeg-faq/part1
+	send usenet/news.answers/jpeg-faq/part2
+
+
+RELATED SOFTWARE
+================
+
+Numerous viewing and image manipulation programs now support JPEG.  (Quite a
+few of them use this library to do so.)  The JPEG FAQ described above lists
+some of the more popular free and shareware viewers, and tells where to
+obtain them on Internet.
+
+If you are on a Unix machine, we highly recommend Jef Poskanzer's free
+PBMPLUS software, which provides many useful operations on PPM-format image
+files.  In particular, it can convert PPM images to and from a wide range of
+other formats, thus making cjpeg/djpeg considerably more useful.  The latest
+version is distributed by the NetPBM group, and is available from numerous
+sites, notably ftp://wuarchive.wustl.edu/graphics/graphics/packages/NetPBM/.
+Unfortunately PBMPLUS/NETPBM is not nearly as portable as the IJG software is;
+you are likely to have difficulty making it work on any non-Unix machine.
+
+A different free JPEG implementation, written by the PVRG group at Stanford,
+is available from ftp://havefun.stanford.edu/pub/jpeg/.  This program
+is designed for research and experimentation rather than production use;
+it is slower, harder to use, and less portable than the IJG code, but it
+is easier to read and modify.  Also, the PVRG code supports lossless JPEG,
+which we do not.  (On the other hand, it doesn't do progressive JPEG.)
+
+
+FILE FORMAT WARS
+================
+
+Some JPEG programs produce files that are not compatible with our library.
+The root of the problem is that the ISO JPEG committee failed to specify a
+concrete file format.  Some vendors "filled in the blanks" on their own,
+creating proprietary formats that no one else could read.  (For example, none
+of the early commercial JPEG implementations for the Macintosh were able to
+exchange compressed files.)
+
+The file format we have adopted is called JFIF (see REFERENCES).  This format
+has been agreed to by a number of major commercial JPEG vendors, and it has
+become the de facto standard.  JFIF is a minimal or "low end" representation.
+We recommend the use of TIFF/JPEG (TIFF revision 6.0 as modified by TIFF
+Technical Note #2) for "high end" applications that need to record a lot of
+additional data about an image.  TIFF/JPEG is fairly new and not yet widely
+supported, unfortunately.
+
+The upcoming JPEG Part 3 standard defines a file format called SPIFF.
+SPIFF is interoperable with JFIF, in the sense that most JFIF decoders should
+be able to read the most common variant of SPIFF.  SPIFF has some technical
+advantages over JFIF, but its major claim to fame is simply that it is an
+official standard rather than an informal one.  At this point it is unclear
+whether SPIFF will supersede JFIF or whether JFIF will remain the de-facto
+standard.  IJG intends to support SPIFF once the standard is frozen, but we
+have not decided whether it should become our default output format or not.
+(In any case, our decoder will remain capable of reading JFIF indefinitely.)
+
+Various proprietary file formats incorporating JPEG compression also exist.
+We have little or no sympathy for the existence of these formats.  Indeed,
+one of the original reasons for developing this free software was to help
+force convergence on common, open format standards for JPEG files.  Don't
+use a proprietary file format!
+
+
+TO DO
+=====
+
+The major thrust for v7 will probably be improvement of visual quality.
+The current method for scaling the quantization tables is known not to be
+very good at low Q values.  We also intend to investigate block boundary
+smoothing, "poor man's variable quantization", and other means of improving
+quality-vs-file-size performance without sacrificing compatibility.
+
+In future versions, we are considering supporting some of the upcoming JPEG
+Part 3 extensions --- principally, variable quantization and the SPIFF file
+format.
+
+As always, speeding things up is of great interest.
+
+Please send bug reports, offers of help, etc. to jpeg-info@uunet.uu.net.
diff --git a/ml/dlib/dlib/external/libjpeg/jcapimin.cpp b/ml/dlib/dlib/external/libjpeg/jcapimin.cpp
new file mode 100644
index 000000000..bc0ceac5b
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcapimin.cpp
@@ -0,0 +1,280 @@
+/*
+ * jcapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-compression case or the transcoding-only
+ * case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jcapistd.c.  But also see jcparam.c for
+ * parameter-setup helper routines, jcomapi.c for routines shared by
+ * compression and decompression, and jctrans.c for the transcoding case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG compression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateCompress (j_compress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_compress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_compress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_compress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = FALSE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->dest = NULL;
+
+  cinfo->comp_info = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++)
+    cinfo->quant_tbl_ptrs[i] = NULL;
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  cinfo->script_space = NULL;
+
+  cinfo->input_gamma = 1.0;	/* in case application forgets */
+
+  /* OK, I'm ready */
+  cinfo->global_state = CSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG compression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_compress (j_compress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG compression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_compress (j_compress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Forcibly suppress or un-suppress all quantization and Huffman tables.
+ * Marks all currently defined tables as already written (if suppress)
+ * or not written (if !suppress).  This will control whether they get emitted
+ * by a subsequent jpeg_start_compress call.
+ *
+ * This routine is exported for use by applications that want to produce
+ * abbreviated JPEG datastreams.  It logically belongs in jcparam.c, but
+ * since it is called by jpeg_start_compress, we put it here --- otherwise
+ * jcparam.o would be linked whether the application used it or not.
+ */
+
+GLOBAL(void)
+jpeg_suppress_tables (j_compress_ptr cinfo, int suppress)
+{
+  int i;
+  JQUANT_TBL * qtbl;
+  JHUFF_TBL * htbl;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if ((qtbl = cinfo->quant_tbl_ptrs[i]) != NULL)
+      qtbl->sent_table = suppress;
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    if ((htbl = cinfo->dc_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+    if ((htbl = cinfo->ac_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+  }
+}
+
+
+/*
+ * Finish JPEG compression.
+ *
+ * If a multipass operating mode was selected, this may do a great deal of
+ * work including most of the actual output.
+ */
+
+GLOBAL(void)
+jpeg_finish_compress (j_compress_ptr cinfo)
+{
+  JDIMENSION iMCU_row;
+
+  if (cinfo->global_state == CSTATE_SCANNING ||
+      cinfo->global_state == CSTATE_RAW_OK) {
+    /* Terminate first pass */
+    if (cinfo->next_scanline < cinfo->image_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_pass) (cinfo);
+  } else if (cinfo->global_state != CSTATE_WRCOEFS)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any remaining passes */
+  while (! cinfo->master->is_last_pass) {
+    (*cinfo->master->prepare_for_pass) (cinfo);
+    for (iMCU_row = 0; iMCU_row < cinfo->total_iMCU_rows; iMCU_row++) {
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) iMCU_row;
+	cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* We bypass the main controller and invoke coef controller directly;
+       * all work is being done from the coefficient buffer.
+       */
+      if (! (*cinfo->coef->compress_data) (cinfo, (JSAMPIMAGE) NULL))
+	ERREXIT(cinfo, JERR_CANT_SUSPEND);
+    }
+    (*cinfo->master->finish_pass) (cinfo);
+  }
+  /* Write EOI, do final cleanup */
+  (*cinfo->marker->write_file_trailer) (cinfo);
+  (*cinfo->dest->term_destination) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+}
+
+
+/*
+ * Write a special marker.
+ * This is only recommended for writing COM or APPn markers.
+ * Must be called after jpeg_start_compress() and before
+ * first call to jpeg_write_scanlines() or jpeg_write_raw_data().
+ */
+
+GLOBAL(void)
+jpeg_write_marker (j_compress_ptr cinfo, int marker,
+		   const JOCTET *dataptr, unsigned int datalen)
+{
+  JMETHOD(void, write_marker_byte, (j_compress_ptr info, int val));
+
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+  write_marker_byte = cinfo->marker->write_marker_byte;	/* copy for speed */
+  while (datalen--) {
+    (*write_marker_byte) (cinfo, *dataptr);
+    dataptr++;
+  }
+}
+
+/* Same, but piecemeal. */
+
+GLOBAL(void)
+jpeg_write_m_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+{
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+}
+
+GLOBAL(void)
+jpeg_write_m_byte (j_compress_ptr cinfo, int val)
+{
+  (*cinfo->marker->write_marker_byte) (cinfo, val);
+}
+
+
+/*
+ * Alternate compression function: just write an abbreviated table file.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * To produce a pair of files containing abbreviated tables and abbreviated
+ * image data, one would proceed as follows:
+ *
+ *		initialize JPEG object
+ *		set JPEG parameters
+ *		set destination to table file
+ *		jpeg_write_tables(cinfo);
+ *		set destination to image file
+ *		jpeg_start_compress(cinfo, FALSE);
+ *		write data...
+ *		jpeg_finish_compress(cinfo);
+ *
+ * jpeg_write_tables has the side effect of marking all tables written
+ * (same as jpeg_suppress_tables(..., TRUE)).  Thus a subsequent start_compress
+ * will not re-emit the tables unless it is passed write_all_tables=TRUE.
+ */
+
+GLOBAL(void)
+jpeg_write_tables (j_compress_ptr cinfo)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Initialize the marker writer ... bit of a crock to do it here. */
+  jinit_marker_writer(cinfo);
+  /* Write them tables! */
+  (*cinfo->marker->write_tables_only) (cinfo);
+  /* And clean up. */
+  (*cinfo->dest->term_destination) (cinfo);
+  /*
+   * In library releases up through v6a, we called jpeg_abort() here to free
+   * any working memory allocated by the destination manager and marker
+   * writer.  Some applications had a problem with that: they allocated space
+   * of their own from the library memory manager, and didn't want it to go
+   * away during write_tables.  So now we do nothing.  This will cause a
+   * memory leak if an app calls write_tables repeatedly without doing a full
+   * compression cycle or otherwise resetting the JPEG object.  However, that
+   * seems less bad than unexpectedly freeing memory in the normal case.
+   * An app that prefers the old behavior can call jpeg_abort for itself after
+   * each call to jpeg_write_tables().
+   */
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcapistd.cpp b/ml/dlib/dlib/external/libjpeg/jcapistd.cpp
new file mode 100644
index 000000000..3f4e08063
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcapistd.cpp
@@ -0,0 +1,161 @@
+/*
+ * jcapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-compression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_compress, it will end up linking in the entire compressor.
+ * We thus must separate this file from jcapimin.c to avoid linking the
+ * whole compression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Compression initialization.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * We require a write_all_tables parameter as a failsafe check when writing
+ * multiple datastreams from the same compression object.  Since prior runs
+ * will have left all the tables marked sent_table=TRUE, a subsequent run
+ * would emit an abbreviated stream (no tables) by default.  This may be what
+ * is wanted, but for safety's sake it should not be the default behavior:
+ * programmers should have to make a deliberate choice to emit abbreviated
+ * images.  Therefore the documentation and examples should encourage people
+ * to pass write_all_tables=TRUE; then it will take active thought to do the
+ * wrong thing.
+ */
+
+GLOBAL(void)
+jpeg_start_compress (j_compress_ptr cinfo, int write_all_tables)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (write_all_tables)
+    jpeg_suppress_tables(cinfo, FALSE);	/* mark all tables to be written */
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  jinit_compress_master(cinfo);
+  /* Set up for the first pass */
+  (*cinfo->master->prepare_for_pass) (cinfo);
+  /* Ready for application to drive first pass through jpeg_write_scanlines
+   * or jpeg_write_raw_data.
+   */
+  cinfo->next_scanline = 0;
+  cinfo->global_state = (cinfo->raw_data_in ? CSTATE_RAW_OK : CSTATE_SCANNING);
+}
+
+
+/*
+ * Write some scanlines of data to the JPEG compressor.
+ *
+ * The return value will be the number of lines actually written.
+ * This should be less than the supplied num_lines only in case that
+ * the data destination module has requested suspension of the compressor,
+ * or if more than image_height scanlines are passed in.
+ *
+ * Note: we warn about excess calls to jpeg_write_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * excess scanlines passed in the last valid call are *silently* ignored,
+ * so that the application need not adjust num_lines for end-of-image
+ * when using a multiple-scanline buffer.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_scanlines (j_compress_ptr cinfo, JSAMPARRAY scanlines,
+		      JDIMENSION num_lines)
+{
+  JDIMENSION row_ctr, rows_left;
+
+  if (cinfo->global_state != CSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height)
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_scanlines.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_scanlines.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Ignore any extra scanlines at bottom of image. */
+  rows_left = cinfo->image_height - cinfo->next_scanline;
+  if (num_lines > rows_left)
+    num_lines = rows_left;
+
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, num_lines);
+  cinfo->next_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to write raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_raw_data (j_compress_ptr cinfo, JSAMPIMAGE data,
+		     JDIMENSION num_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != CSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_raw_data.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_raw_data.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Verify that at least one iMCU row has been passed. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * DCTSIZE;
+  if (num_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Directly compress the row. */
+  if (! (*cinfo->coef->compress_data) (cinfo, data)) {
+    /* If compressor did not consume the whole row, suspend processing. */
+    return 0;
+  }
+
+  /* OK, we processed one iMCU row. */
+  cinfo->next_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jccoefct.cpp b/ml/dlib/dlib/external/libjpeg/jccoefct.cpp
new file mode 100644
index 000000000..175d7ecd9
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jccoefct.cpp
@@ -0,0 +1,449 @@
+/*
+ * jccoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for compression.
+ * This controller is the top level of the JPEG compressor proper.
+ * The coefficient buffer lies between forward-DCT and entropy encoding steps.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* We use a full-image coefficient buffer when doing Huffman optimization,
+ * and also for writing multiple-scan JPEG files.  In all cases, the DCT
+ * step is run during the first pass, and subsequent passes need only read
+ * the buffered coefficients.
+ */
+#ifdef ENTROPY_OPT_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#else
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#endif
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;	/* iMCU row # within image */
+  JDIMENSION mcu_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* For single-pass compression, it's sufficient to buffer just one MCU
+   * (although this may prove a bit slow in practice).  We allocate a
+   * workspace of C_MAX_BLOCKS_IN_MCU coefficient blocks, and reuse it for each
+   * MCU constructed and sent.  (On 80x86, the workspace is FAR even though
+   * it's not really very big; this is to keep the module interfaces unchanged
+   * when a large coefficient buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays.
+   */
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+
+/* Forward declarations */
+METHODDEF(int) compress_data
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+METHODDEF(int) compress_first_pass
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+METHODDEF(int) compress_output
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (coef->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_data;
+    break;
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_first_pass;
+    break;
+  case JBUF_CRANK_DEST:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_output;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data in the single-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(int)
+compress_data (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, bi, ci, yindex, yoffset, blockcnt;
+  JDIMENSION ypos, xpos;
+  jpeg_component_info *compptr;
+
+  /* Loop to write as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Determine where data comes from in input_buf and do the DCT thing.
+       * Each call on forward_DCT processes a horizontal row of DCT blocks
+       * as wide as an MCU; we rely on having allocated the MCU_buffer[] blocks
+       * sequentially.  Dummy blocks at the right or bottom edge are filled in
+       * specially.  The data in them does not matter for image reconstruction,
+       * so we fill them with values that will encode to the smallest amount of
+       * data, viz: all zeroes in the AC entries, DC entries equal to previous
+       * block's DC value.  (Thanks to Thomas Kinsman for this idea.)
+       */
+      blkn = 0;
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						: compptr->last_col_width;
+	xpos = MCU_col_num * compptr->MCU_sample_width;
+	ypos = yoffset * DCTSIZE; /* ypos == (yoffset+yindex) * DCTSIZE */
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (coef->iMCU_row_num < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    (*cinfo->fdct->forward_DCT) (cinfo, compptr,
+					 input_buf[compptr->component_index],
+					 coef->MCU_buffer[blkn],
+					 ypos, xpos, (JDIMENSION) blockcnt);
+	    if (blockcnt < compptr->MCU_width) {
+	      /* Create some dummy blocks at the right edge of the image. */
+	      jzero_far((void FAR *) coef->MCU_buffer[blkn + blockcnt],
+			(compptr->MCU_width - blockcnt) * SIZEOF(JBLOCK));
+	      for (bi = blockcnt; bi < compptr->MCU_width; bi++) {
+		coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn+bi-1][0][0];
+	      }
+	    }
+	  } else {
+	    /* Create a row of dummy blocks at the bottom of the image. */
+	    jzero_far((void FAR *) coef->MCU_buffer[blkn],
+		      compptr->MCU_width * SIZEOF(JBLOCK));
+	    for (bi = 0; bi < compptr->MCU_width; bi++) {
+	      coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn-1][0][0];
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  ypos += DCTSIZE;
+	}
+      }
+      /* Try to write the MCU.  In event of a suspension failure, we will
+       * re-DCT the MCU on restart (a bit inefficient, could be fixed...)
+       */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+
+/*
+ * Process some data in the first pass of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * This amount of data is read from the source buffer, DCT'd and quantized,
+ * and saved into the virtual arrays.  We also generate suitable dummy blocks
+ * as needed at the right and lower edges.  (The dummy blocks are constructed
+ * in the virtual arrays, which have been padded appropriately.)  This makes
+ * it possible for subsequent passes not to worry about real vs. dummy blocks.
+ *
+ * We must also emit the data to the entropy encoder.  This is conveniently
+ * done by calling compress_output() after we've loaded the current strip
+ * of the virtual arrays.
+ *
+ * NB: input_buf contains a plane for each component in image.  All
+ * components are DCT'd and loaded into the virtual arrays in this pass.
+ * However, it may be that only a subset of the components are emitted to
+ * the entropy encoder during this first pass; be careful about looking
+ * at the scan-dependent variables (MCU dimensions, etc).
+ */
+
+METHODDEF(int)
+compress_first_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION blocks_across, MCUs_across, MCUindex;
+  int bi, ci, h_samp_factor, block_row, block_rows, ndummy;
+  JCOEF lastDC;
+  jpeg_component_info *compptr;
+  JBLOCKARRAY buffer;
+  JBLOCKROW thisblockrow, lastblockrow;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (coef->iMCU_row_num < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here, since may not be set! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    blocks_across = compptr->width_in_blocks;
+    h_samp_factor = compptr->h_samp_factor;
+    /* Count number of dummy blocks to be added at the right margin. */
+    ndummy = (int) (blocks_across % h_samp_factor);
+    if (ndummy > 0)
+      ndummy = h_samp_factor - ndummy;
+    /* Perform DCT for all non-dummy blocks in this iMCU row.  Each call
+     * on forward_DCT processes a complete horizontal row of DCT blocks.
+     */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      thisblockrow = buffer[block_row];
+      (*cinfo->fdct->forward_DCT) (cinfo, compptr,
+				   input_buf[ci], thisblockrow,
+				   (JDIMENSION) (block_row * DCTSIZE),
+				   (JDIMENSION) 0, blocks_across);
+      if (ndummy > 0) {
+	/* Create dummy blocks at the right edge of the image. */
+	thisblockrow += blocks_across; /* => first dummy block */
+	jzero_far((void FAR *) thisblockrow, ndummy * SIZEOF(JBLOCK));
+	lastDC = thisblockrow[-1][0];
+	for (bi = 0; bi < ndummy; bi++) {
+	  thisblockrow[bi][0] = lastDC;
+	}
+      }
+    }
+    /* If at end of image, create dummy block rows as needed.
+     * The tricky part here is that within each MCU, we want the DC values
+     * of the dummy blocks to match the last real block's DC value.
+     * This squeezes a few more bytes out of the resulting file...
+     */
+    if (coef->iMCU_row_num == last_iMCU_row) {
+      blocks_across += ndummy;	/* include lower right corner */
+      MCUs_across = blocks_across / h_samp_factor;
+      for (block_row = block_rows; block_row < compptr->v_samp_factor;
+	   block_row++) {
+	thisblockrow = buffer[block_row];
+	lastblockrow = buffer[block_row-1];
+	jzero_far((void FAR *) thisblockrow,
+		  (size_t) (blocks_across * SIZEOF(JBLOCK)));
+	for (MCUindex = 0; MCUindex < MCUs_across; MCUindex++) {
+	  lastDC = lastblockrow[h_samp_factor-1][0];
+	  for (bi = 0; bi < h_samp_factor; bi++) {
+	    thisblockrow[bi][0] = lastDC;
+	  }
+	  thisblockrow += h_samp_factor; /* advance to next MCU in row */
+	  lastblockrow += h_samp_factor;
+	}
+      }
+    }
+  }
+  /* NB: compress_output will increment iMCU_row_num if successful.
+   * A suspension return will result in redoing all the work above next time.
+   */
+
+  /* Emit data to the entropy encoder, sharing code with subsequent passes */
+  return compress_output(cinfo, input_buf);
+}
+
+
+/*
+ * Process some data in subsequent passes of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(int)
+compress_output (j_compress_ptr cinfo, JSAMPIMAGE )//input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan.
+   * NB: during first pass, this is safe only because the buffers will
+   * already be aligned properly, so jmemmgr.c won't need to do any I/O.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to write the MCU. */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+#endif /* FULL_COEF_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_coef_controller (j_compress_ptr cinfo, int need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *) coef;
+  coef->pub.start_pass = start_pass_coef;
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    int ci;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) compptr->v_samp_factor);
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->whole_image[0] = NULL; /* flag for no virtual arrays */
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jccolor.cpp b/ml/dlib/dlib/external/libjpeg/jccolor.cpp
new file mode 100644
index 000000000..c5cfeded5
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jccolor.cpp
@@ -0,0 +1,459 @@
+/*
+ * jccolor.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_converter pub; /* public fields */
+
+  /* Private state for RGB->YCC conversion */
+  long * rgb_ycc_tab;		/* => table for RGB to YCbCr conversion */
+} my_color_converter;
+
+typedef my_color_converter * my_cconvert_ptr;
+
+
+/**************** RGB -> YCbCr conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *	Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + CENTERJSAMPLE
+ *	Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + CENTERJSAMPLE
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ * Note: older versions of the IJG code used a zero offset of MAXJSAMPLE/2,
+ * rather than CENTERJSAMPLE, for Cb and Cr.  This gave equal positive and
+ * negative swings for Cb/Cr, but meant that grayscale values (Cb=Cr=0)
+ * were not represented exactly.  Now we sacrifice exact representation of
+ * maximum red and maximum blue in order to get exact grayscales.
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times R,G,B for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The CENTERJSAMPLE offsets and the rounding fudge-factor of 0.5 are included
+ * in the tables to save adding them separately in the inner loop.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define CBCR_OFFSET	((long) CENTERJSAMPLE << SCALEBITS)
+#define ONE_HALF	((long) 1 << (SCALEBITS-1))
+#define FIX(x)		((long) ((x) * (1L<<SCALEBITS) + 0.5))
+
+/* We allocate one big table and divide it up into eight parts, instead of
+ * doing eight alloc_small requests.  This lets us use a single table base
+ * address, which can be held in a register in the inner loops on many
+ * machines (more than can hold all eight addresses, anyway).
+ */
+
+#define R_Y_OFF		0			/* offset to R => Y section */
+#define G_Y_OFF		(1*(MAXJSAMPLE+1))	/* offset to G => Y section */
+#define B_Y_OFF		(2*(MAXJSAMPLE+1))	/* etc. */
+#define R_CB_OFF	(3*(MAXJSAMPLE+1))
+#define G_CB_OFF	(4*(MAXJSAMPLE+1))
+#define B_CB_OFF	(5*(MAXJSAMPLE+1))
+#define R_CR_OFF	B_CB_OFF		/* B=>Cb, R=>Cr are the same */
+#define G_CR_OFF	(6*(MAXJSAMPLE+1))
+#define B_CR_OFF	(7*(MAXJSAMPLE+1))
+#define TABLE_SIZE	(8*(MAXJSAMPLE+1))
+
+
+/*
+ * Initialize for RGB->YCC colorspace conversion.
+ */
+
+METHODDEF(void)
+rgb_ycc_start (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  long * rgb_ycc_tab;
+  long i;
+
+  /* Allocate and fill in the conversion tables. */
+  cconvert->rgb_ycc_tab = rgb_ycc_tab = (long *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(TABLE_SIZE * SIZEOF(long)));
+
+  for (i = 0; i <= MAXJSAMPLE; i++) {
+    rgb_ycc_tab[i+R_Y_OFF] = FIX(0.29900) * i;
+    rgb_ycc_tab[i+G_Y_OFF] = FIX(0.58700) * i;
+    rgb_ycc_tab[i+B_Y_OFF] = FIX(0.11400) * i     + ONE_HALF;
+    rgb_ycc_tab[i+R_CB_OFF] = (-FIX(0.16874)) * i;
+    rgb_ycc_tab[i+G_CB_OFF] = (-FIX(0.33126)) * i;
+    /* We use a rounding fudge-factor of 0.5-epsilon for Cb and Cr.
+     * This ensures that the maximum output will round to MAXJSAMPLE
+     * not MAXJSAMPLE+1, and thus that we don't have to range-limit.
+     */
+    rgb_ycc_tab[i+B_CB_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+/*  B=>Cb and R=>Cr tables are the same
+    rgb_ycc_tab[i+R_CR_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+*/
+    rgb_ycc_tab[i+G_CR_OFF] = (-FIX(0.41869)) * i;
+    rgb_ycc_tab[i+B_CR_OFF] = (-FIX(0.08131)) * i;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ *
+ * Note that we change from the application's interleaved-pixel format
+ * to our internal noninterleaved, one-plane-per-component format.
+ * The input buffer is therefore three times as wide as the output buffer.
+ *
+ * A starting row offset is provided only for the output buffer.  The caller
+ * can easily adjust the passed input_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+rgb_ycc_convert (j_compress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		 JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int r, g, b;
+  long * ctab = cconvert->rgb_ycc_tab;
+  JSAMPROW inptr;
+  JSAMPROW outptr0, outptr1, outptr2;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = GETJSAMPLE(inptr[RGB_RED]);
+      g = GETJSAMPLE(inptr[RGB_GREEN]);
+      b = GETJSAMPLE(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/**************** Cases other than RGB -> YCbCr **************/
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles RGB->grayscale conversion, which is the same
+ * as the RGB->Y portion of RGB->YCbCr.
+ * We assume rgb_ycc_start has been called (we only use the Y tables).
+ */
+
+METHODDEF(void)
+rgb_gray_convert (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		  JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int r, g, b;
+  long * ctab = cconvert->rgb_ycc_tab;
+  JSAMPROW inptr;
+  JSAMPROW outptr;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = GETJSAMPLE(inptr[RGB_RED]);
+      g = GETJSAMPLE(inptr[RGB_GREEN]);
+      b = GETJSAMPLE(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* Y */
+      outptr[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles Adobe-style CMYK->YCCK conversion,
+ * where we convert R=1-C, G=1-M, and B=1-Y to YCbCr using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume rgb_ycc_start has been called.
+ */
+
+METHODDEF(void)
+cmyk_ycck_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int r, g, b;
+  long * ctab = cconvert->rgb_ycc_tab;
+  JSAMPROW inptr;
+  JSAMPROW outptr0, outptr1, outptr2, outptr3;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    outptr3 = output_buf[3][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = MAXJSAMPLE - GETJSAMPLE(inptr[0]);
+      g = MAXJSAMPLE - GETJSAMPLE(inptr[1]);
+      b = MAXJSAMPLE - GETJSAMPLE(inptr[2]);
+      /* K passes through as-is */
+      outptr3[col] = inptr[3];	/* don't need GETJSAMPLE here */
+      inptr += 4;
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles grayscale output with no conversion.
+ * The source can be either plain grayscale or YCbCr (since Y == gray).
+ */
+
+METHODDEF(void)
+grayscale_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  JSAMPROW inptr;
+  JSAMPROW outptr;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+  int instride = cinfo->input_components;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[col] = inptr[0];	/* don't need GETJSAMPLE() here */
+      inptr += instride;
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles multi-component colorspaces without conversion.
+ * We assume input_components == num_components.
+ */
+
+METHODDEF(void)
+null_convert (j_compress_ptr cinfo,
+	      JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+	      JDIMENSION output_row, int num_rows)
+{
+  JSAMPROW inptr;
+  JSAMPROW outptr;
+  JDIMENSION col;
+  int ci;
+  int nc = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    /* It seems fastest to make a separate pass for each component. */
+    for (ci = 0; ci < nc; ci++) {
+      inptr = *input_buf;
+      outptr = output_buf[ci][output_row];
+      for (col = 0; col < num_cols; col++) {
+	outptr[col] = inptr[ci]; /* don't need GETJSAMPLE() here */
+	inptr += nc;
+      }
+    }
+    input_buf++;
+    output_row++;
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+null_method (j_compress_ptr )//cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for input colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_converter (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_converter));
+  cinfo->cconvert = (struct jpeg_color_converter *) cconvert;
+  /* set start_pass to null method until we find out differently */
+  cconvert->pub.start_pass = null_method;
+
+  /* Make sure input_components agrees with in_color_space */
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->input_components != 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+#if RGB_PIXELSIZE != 3
+    if (cinfo->input_components != RGB_PIXELSIZE)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+#endif /* else share code with YCbCr */
+
+  case JCS_YCbCr:
+    if (cinfo->input_components != 3)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->input_components != 4)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->input_components < 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+  }
+
+  /* Check num_components, set conversion method based on requested space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_GRAYSCALE)
+      cconvert->pub.color_convert = grayscale_convert;
+    else if (cinfo->in_color_space == JCS_RGB) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = rgb_gray_convert;
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = grayscale_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_RGB && RGB_PIXELSIZE == 3)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_RGB) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = rgb_ycc_convert;
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = cmyk_ycck_convert;
+    } else if (cinfo->in_color_space == JCS_YCCK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:			/* allow null conversion of JCS_UNKNOWN */
+    if (cinfo->jpeg_color_space != cinfo->in_color_space ||
+	cinfo->num_components != cinfo->input_components)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    cconvert->pub.color_convert = null_convert;
+    break;
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcdctmgr.cpp b/ml/dlib/dlib/external/libjpeg/jcdctmgr.cpp
new file mode 100644
index 000000000..cbfc1a857
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcdctmgr.cpp
@@ -0,0 +1,387 @@
+/*
+ * jcdctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the forward-DCT management logic.
+ * This code selects a particular DCT implementation to be used,
+ * and it performs related housekeeping chores including coefficient
+ * quantization.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_forward_dct pub;	/* public fields */
+
+  /* Pointer to the DCT routine actually in use */
+  forward_DCT_method_ptr do_dct;
+
+  /* The actual post-DCT divisors --- not identical to the quant table
+   * entries, because of scaling (especially for an unnormalized DCT).
+   * Each table is given in normal array order.
+   */
+  DCTELEM * divisors[NUM_QUANT_TBLS];
+
+#ifdef DCT_FLOAT_SUPPORTED
+  /* Same as above for the floating-point case. */
+  float_DCT_method_ptr do_float_dct;
+  FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
+#endif
+} my_fdct_controller;
+
+typedef my_fdct_controller * my_fdct_ptr;
+
+
+/*
+ * Initialize for a processing pass.
+ * Verify that all referenced Q-tables are present, and set up
+ * the divisor table for each one.
+ * In the current implementation, DCT of all components is done during
+ * the first pass, even if only some components will be output in the
+ * first scan.  Hence all components should be examined here.
+ */
+
+METHODDEF(void)
+start_pass_fdctmgr (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  int ci, qtblno, i;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtbl;
+  DCTELEM * dtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    qtblno = compptr->quant_tbl_no;
+    /* Make sure specified quantization table is present */
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    qtbl = cinfo->quant_tbl_ptrs[qtblno];
+    /* Compute divisors for this quant table */
+    /* We may do this more than once for same table, but it's not a big deal */
+    switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+    case JDCT_ISLOW:
+      /* For LL&M IDCT method, divisors are equal to raw quantization
+       * coefficients multiplied by 8 (to counteract scaling).
+       */
+      if (fdct->divisors[qtblno] == NULL) {
+	fdct->divisors[qtblno] = (DCTELEM *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      DCTSIZE2 * SIZEOF(DCTELEM));
+      }
+      dtbl = fdct->divisors[qtblno];
+      for (i = 0; i < DCTSIZE2; i++) {
+	dtbl[i] = ((DCTELEM) qtbl->quantval[i]) << 3;
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 */
+#define CONST_BITS 14
+	static const short aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	if (fdct->divisors[qtblno] == NULL) {
+	  fdct->divisors[qtblno] = (DCTELEM *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					DCTSIZE2 * SIZEOF(DCTELEM));
+	}
+	dtbl = fdct->divisors[qtblno];
+	for (i = 0; i < DCTSIZE2; i++) {
+	  dtbl[i] = (DCTELEM)
+	    DESCALE(MULTIPLY16V16((long) qtbl->quantval[i],
+				  (long) aanscales[i]),
+		    CONST_BITS-3);
+	}
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 * What's actually stored is 1/divisor so that the inner loop can
+	 * use a multiplication rather than a division.
+	 */
+	FAST_FLOAT * fdtbl;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	if (fdct->float_divisors[qtblno] == NULL) {
+	  fdct->float_divisors[qtblno] = (FAST_FLOAT *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					DCTSIZE2 * SIZEOF(FAST_FLOAT));
+	}
+	fdtbl = fdct->float_divisors[qtblno];
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fdtbl[i] = (FAST_FLOAT)
+	      (1.0 / (((double) qtbl->quantval[i] *
+		       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
+	    i++;
+	  }
+	}
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Perform forward DCT on one or more blocks of a component.
+ *
+ * The input samples are taken from the sample_data[] array starting at
+ * position start_row/start_col, and moving to the right for any additional
+ * blocks. The quantized coefficients are returned in coef_blocks[].
+ */
+
+METHODDEF(void)
+forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
+	     JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+	     JDIMENSION start_row, JDIMENSION start_col,
+	     JDIMENSION num_blocks)
+/* This version is used for integer DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  forward_DCT_method_ptr do_dct = fdct->do_dct;
+  DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
+  DCTELEM workspace[DCTSIZE2];	/* work area for FDCT subroutine */
+  JDIMENSION bi;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    { DCTELEM *workspaceptr;
+      JSAMPROW elemptr;
+      int elemr;
+
+      workspaceptr = workspace;
+      for (elemr = 0; elemr < DCTSIZE; elemr++) {
+	elemptr = sample_data[elemr] + start_col;
+#if DCTSIZE == 8		/* unroll the inner loop */
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	*workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+#else
+	{ int elemc;
+	  for (elemc = DCTSIZE; elemc > 0; elemc--) {
+	    *workspaceptr++ = GETJSAMPLE(*elemptr++) - CENTERJSAMPLE;
+	  }
+	}
+#endif
+      }
+    }
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    { DCTELEM temp, qval;
+      int i;
+      JCOEFPTR output_ptr = coef_blocks[bi];
+
+      for (i = 0; i < DCTSIZE2; i++) {
+	qval = divisors[i];
+	temp = workspace[i];
+	/* Divide the coefficient value by qval, ensuring proper rounding.
+	 * Since C does not specify the direction of rounding for negative
+	 * quotients, we have to force the dividend positive for portability.
+	 *
+	 * In most files, at least half of the output values will be zero
+	 * (at default quantization settings, more like three-quarters...)
+	 * so we should ensure that this case is fast.  On many machines,
+	 * a comparison is enough cheaper than a divide to make a special test
+	 * a win.  Since both inputs will be nonnegative, we need only test
+	 * for a < b to discover whether a/b is 0.
+	 * If your machine's division is fast enough, define FAST_DIVIDE.
+	 */
+#ifdef FAST_DIVIDE
+#define DIVIDE_BY(a,b)	a /= b
+#else
+#define DIVIDE_BY(a,b)	if (a >= b) a /= b; else a = 0
+#endif
+	if (temp < 0) {
+	  temp = -temp;
+	  temp += qval>>1;	/* for rounding */
+	  DIVIDE_BY(temp, qval);
+	  temp = -temp;
+	} else {
+	  temp += qval>>1;	/* for rounding */
+	  DIVIDE_BY(temp, qval);
+	}
+	output_ptr[i] = (JCOEF) temp;
+      }
+    }
+  }
+}
+
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+METHODDEF(void)
+forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+		   JDIMENSION start_row, JDIMENSION start_col,
+		   JDIMENSION num_blocks)
+/* This version is used for floating-point DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  float_DCT_method_ptr do_dct = fdct->do_float_dct;
+  FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
+  FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
+  JDIMENSION bi;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += DCTSIZE) {
+    /* Load data into workspace, applying unsigned->signed conversion */
+    { FAST_FLOAT *workspaceptr;
+      JSAMPROW elemptr;
+      int elemr;
+
+      workspaceptr = workspace;
+      for (elemr = 0; elemr < DCTSIZE; elemr++) {
+	elemptr = sample_data[elemr] + start_col;
+#if DCTSIZE == 8		/* unroll the inner loop */
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	*workspaceptr++ = (FAST_FLOAT)(GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+#else
+	{ int elemc;
+	  for (elemc = DCTSIZE; elemc > 0; elemc--) {
+	    *workspaceptr++ = (FAST_FLOAT)
+	      (GETJSAMPLE(*elemptr++) - CENTERJSAMPLE);
+	  }
+	}
+#endif
+      }
+    }
+
+    /* Perform the DCT */
+    (*do_dct) (workspace);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    { FAST_FLOAT temp;
+      int i;
+      JCOEFPTR output_ptr = coef_blocks[bi];
+
+      for (i = 0; i < DCTSIZE2; i++) {
+	/* Apply the quantization and scaling factor */
+	temp = workspace[i] * divisors[i];
+	/* Round to nearest integer.
+	 * Since C does not specify the direction of rounding for negative
+	 * quotients, we have to force the dividend positive for portability.
+	 * The maximum coefficient size is +-16K (for 12-bit data), so this
+	 * code should work for either 16-bit or 32-bit ints.
+	 */
+	output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
+      }
+    }
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
+
+
+/*
+ * Initialize FDCT manager.
+ */
+
+GLOBAL(void)
+jinit_forward_dct (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct;
+  int i;
+
+  fdct = (my_fdct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_fdct_controller));
+  cinfo->fdct = (struct jpeg_forward_dct *) fdct;
+  fdct->pub.start_pass = start_pass_fdctmgr;
+
+  switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+  case JDCT_ISLOW:
+    fdct->pub.forward_DCT = forward_DCT;
+    fdct->do_dct = jpeg_fdct_islow;
+    break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+  case JDCT_IFAST:
+    fdct->pub.forward_DCT = forward_DCT;
+    fdct->do_dct = jpeg_fdct_ifast;
+    break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  case JDCT_FLOAT:
+    fdct->pub.forward_DCT = forward_DCT_float;
+    fdct->do_float_dct = jpeg_fdct_float;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+
+  /* Mark divisor tables unallocated */
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    fdct->divisors[i] = NULL;
+#ifdef DCT_FLOAT_SUPPORTED
+    fdct->float_divisors[i] = NULL;
+#endif
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jchuff.cpp b/ml/dlib/dlib/external/libjpeg/jchuff.cpp
new file mode 100644
index 000000000..d543319a6
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jchuff.cpp
@@ -0,0 +1,909 @@
+/*
+ * jchuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy encoding routines.
+ *
+ * Much of the complexity here has to do with supporting output suspension.
+ * If the data destination module demands suspension, we want to be able to
+ * back up to the start of the current MCU.  To do this, we copy state
+ * variables into local working storage, and update them back to the
+ * permanent JPEG objects only upon successful completion of an MCU.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jchuff.h"		/* Declarations shared with jcphuff.c */
+
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  long put_buffer;		/* current bit-accumulation buffer */
+  int put_bits;			/* # of bits now in it */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).put_buffer = (src).put_buffer, \
+	 (dest).put_bits = (src).put_bits, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  savable_state saved;		/* Bit buffer & DC state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+#ifdef ENTROPY_OPT_SUPPORTED	/* Statistics tables for optimization */
+  long * dc_count_ptrs[NUM_HUFF_TBLS];
+  long * ac_count_ptrs[NUM_HUFF_TBLS];
+#endif
+} huff_entropy_encoder;
+
+typedef huff_entropy_encoder * huff_entropy_ptr;
+
+/* Working state while writing an MCU.
+ * This struct contains all the fields that are needed by subroutines.
+ */
+
+typedef struct {
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  savable_state cur;		/* Current bit buffer & DC state */
+  j_compress_ptr cinfo;		/* dump_buffer needs access to this */
+} working_state;
+
+
+/* Forward declarations */
+METHODDEF(int) encode_mcu_huff JPP((j_compress_ptr cinfo,
+					JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_huff JPP((j_compress_ptr cinfo));
+#ifdef ENTROPY_OPT_SUPPORTED
+METHODDEF(int) encode_mcu_gather JPP((j_compress_ptr cinfo,
+					  JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_gather JPP((j_compress_ptr cinfo));
+#endif
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ * If gather_statistics is TRUE, we do not output anything during the scan,
+ * just count the Huffman symbols used and generate Huffman code tables.
+ */
+
+METHODDEF(void)
+start_pass_huff (j_compress_ptr cinfo, int gather_statistics)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info * compptr;
+
+  if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->pub.encode_mcu = encode_mcu_gather;
+    entropy->pub.finish_pass = finish_pass_gather;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    entropy->pub.encode_mcu = encode_mcu_huff;
+    entropy->pub.finish_pass = finish_pass_huff;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (gather_statistics) {
+#ifdef ENTROPY_OPT_SUPPORTED
+      /* Check for invalid table indexes */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (dctbl < 0 || dctbl >= NUM_HUFF_TBLS)
+	ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, dctbl);
+      if (actbl < 0 || actbl >= NUM_HUFF_TBLS)
+	ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, actbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->dc_count_ptrs[dctbl] == NULL)
+	entropy->dc_count_ptrs[dctbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->dc_count_ptrs[dctbl], 257 * SIZEOF(long));
+      if (entropy->ac_count_ptrs[actbl] == NULL)
+	entropy->ac_count_ptrs[actbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->ac_count_ptrs[actbl], 257 * SIZEOF(long));
+#endif
+    } else {
+      /* Compute derived values for Huffman tables */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, TRUE, dctbl,
+			      & entropy->dc_derived_tbls[dctbl]);
+      jpeg_make_c_derived_tbl(cinfo, FALSE, actbl,
+			      & entropy->ac_derived_tbls[actbl]);
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bit buffer to empty */
+  entropy->saved.put_buffer = 0;
+  entropy->saved.put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jcphuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_c_derived_tbl (j_compress_ptr cinfo, int isDC, int tblno,
+			 c_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  c_derived_tbl *dtbl;
+  int p, i, l, lastp, si, maxsymbol;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (c_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(c_derived_tbl));
+  dtbl = *pdtbl;
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  lastp = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((long) code) >= (((long) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+  
+  /* Figure C.3: generate encoding tables */
+  /* These are code and size indexed by symbol value */
+
+  /* Set all codeless symbols to have code length 0;
+   * this lets us detect duplicate VAL entries here, and later
+   * allows emit_bits to detect any attempt to emit such symbols.
+   */
+  MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
+
+  /* This is also a convenient place to check for out-of-range
+   * and duplicated VAL entries.  We allow 0..255 for AC symbols
+   * but only 0..15 for DC.  (We could constrain them further
+   * based on data depth and mode, but this seems enough.)
+   */
+  maxsymbol = isDC ? 15 : 255;
+
+  for (p = 0; p < lastp; p++) {
+    i = htbl->huffval[p];
+    if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    dtbl->ehufco[i] = huffcode[p];
+    dtbl->ehufsi[i] = huffsize[p];
+  }
+}
+
+
+/* Outputting bytes to the file */
+
+/* Emit a byte, taking 'action' if must suspend. */
+#define emit_byte(state,val,action)  \
+	{ *(state)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(state)->free_in_buffer == 0)  \
+	    if (! dump_buffer(state))  \
+	      { action; } }
+
+
+LOCAL(int)
+dump_buffer (working_state * state)
+/* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
+{
+  struct jpeg_destination_mgr * dest = state->cinfo->dest;
+
+  if (! (*dest->empty_output_buffer) (state->cinfo))
+    return FALSE;
+  /* After a successful buffer dump, must reset buffer pointers */
+  state->next_output_byte = dest->next_output_byte;
+  state->free_in_buffer = dest->free_in_buffer;
+  return TRUE;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+inline
+LOCAL(int)
+emit_bits (working_state * state, unsigned int code, int size)
+/* Emit some bits; return TRUE if successful, FALSE if must suspend */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  long put_buffer = (long) code;
+  int put_bits = state->cur.put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(state->cinfo, JERR_HUFF_MISSING_CODE);
+
+  put_buffer &= (((long) 1)<<size) - 1; /* mask off any extra bits in code */
+  
+  put_bits += size;		/* new number of bits in buffer */
+  
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  put_buffer |= state->cur.put_buffer; /* and merge with old buffer contents */
+  
+  while (put_bits >= 8) {
+    int c = (int) ((put_buffer >> 16) & 0xFF);
+    
+    emit_byte(state, c, return FALSE);
+    if (c == 0xFF) {		/* need to stuff a zero byte? */
+      emit_byte(state, 0, return FALSE);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  state->cur.put_buffer = put_buffer; /* update state variables */
+  state->cur.put_bits = put_bits;
+
+  return TRUE;
+}
+
+
+LOCAL(int)
+flush_bits (working_state * state)
+{
+  if (! emit_bits(state, 0x7F, 7)) /* fill any partial byte with ones */
+    return FALSE;
+  state->cur.put_buffer = 0;	/* and reset bit-buffer to empty */
+  state->cur.put_bits = 0;
+  return TRUE;
+}
+
+
+/* Encode a single block's worth of coefficients */
+
+LOCAL(int)
+encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
+		  c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  int temp, temp2;
+  int nbits;
+  int k, r, i;
+  
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+  
+  temp = temp2 = block[0] - last_dc_val;
+
+  if (temp < 0) {
+    temp = -temp;		/* temp is abs value of input */
+    /* For a negative input, want temp2 = bitwise complement of abs(input) */
+    /* This code assumes we are on a two's complement machine */
+    temp2--;
+  }
+  
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > MAX_COEF_BITS+1)
+    ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
+  
+  /* Emit the Huffman-coded symbol for the number of bits */
+  if (! emit_bits(state, dctbl->ehufco[nbits], dctbl->ehufsi[nbits]))
+    return FALSE;
+
+  /* Emit that number of bits of the value, if positive, */
+  /* or the complement of its magnitude, if negative. */
+  if (nbits)			/* emit_bits rejects calls with size 0 */
+    if (! emit_bits(state, (unsigned int) temp2, nbits))
+      return FALSE;
+
+  /* Encode the AC coefficients per section F.1.2.2 */
+  
+  r = 0;			/* r = run length of zeros */
+  
+  for (k = 1; k < DCTSIZE2; k++) {
+    if ((temp = block[jpeg_natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+	if (! emit_bits(state, actbl->ehufco[0xF0], actbl->ehufsi[0xF0]))
+	  return FALSE;
+	r -= 16;
+      }
+
+      temp2 = temp;
+      if (temp < 0) {
+	temp = -temp;		/* temp is abs value of input */
+	/* This code assumes we are on a two's complement machine */
+	temp2--;
+      }
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;		/* there must be at least one 1 bit */
+      while ((temp >>= 1))
+	nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > MAX_COEF_BITS)
+	ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
+      
+      /* Emit Huffman symbol for run length / number of bits */
+      i = (r << 4) + nbits;
+      if (! emit_bits(state, actbl->ehufco[i], actbl->ehufsi[i]))
+	return FALSE;
+
+      /* Emit that number of bits of the value, if positive, */
+      /* or the complement of its magnitude, if negative. */
+      if (! emit_bits(state, (unsigned int) temp2, nbits))
+	return FALSE;
+      
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    if (! emit_bits(state, actbl->ehufco[0], actbl->ehufsi[0]))
+      return FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(int)
+emit_restart (working_state * state, int restart_num)
+{
+  int ci;
+
+  if (! flush_bits(state))
+    return FALSE;
+
+  emit_byte(state, 0xFF, return FALSE);
+  emit_byte(state, JPEG_RST0 + restart_num, return FALSE);
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
+    state->cur.last_dc_val[ci] = 0;
+
+  /* The restart counter is not updated until we successfully write the MCU. */
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of Huffman-compressed coefficients.
+ */
+
+METHODDEF(int)
+encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Load up working state */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  ASSIGN_STATE(state.cur, entropy->saved);
+  state.cinfo = cinfo;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! emit_restart(&state, entropy->next_restart_num))
+	return FALSE;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    if (! encode_one_block(&state,
+			   MCU_data[blkn][0], state.cur.last_dc_val[ci],
+			   entropy->dc_derived_tbls[compptr->dc_tbl_no],
+			   entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+      return FALSE;
+    /* Update last_dc_val */
+    state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  /* Completed MCU, so update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  ASSIGN_STATE(entropy->saved, state.cur);
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass_huff (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+
+  /* Load up working state ... flush_bits needs it */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  ASSIGN_STATE(state.cur, entropy->saved);
+  state.cinfo = cinfo;
+
+  /* Flush out the last data */
+  if (! flush_bits(&state))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  /* Update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  ASSIGN_STATE(entropy->saved, state.cur);
+}
+
+
+/*
+ * Huffman coding optimization.
+ *
+ * We first scan the supplied data and count the number of uses of each symbol
+ * that is to be Huffman-coded. (This process MUST agree with the code above.)
+ * Then we build a Huffman coding tree for the observed counts.
+ * Symbols which are not needed at all for the particular image are not
+ * assigned any code, which saves space in the DHT marker as well as in
+ * the compressed data.
+ */
+
+#ifdef ENTROPY_OPT_SUPPORTED
+
+
+/* Process a single block's worth of coefficients */
+
+LOCAL(void)
+htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
+		 long dc_counts[], long ac_counts[])
+{
+  int temp;
+  int nbits;
+  int k, r;
+  
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+  
+  temp = block[0] - last_dc_val;
+  if (temp < 0)
+    temp = -temp;
+  
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > MAX_COEF_BITS+1)
+    ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+  /* Count the Huffman symbol for the number of bits */
+  dc_counts[nbits]++;
+  
+  /* Encode the AC coefficients per section F.1.2.2 */
+  
+  r = 0;			/* r = run length of zeros */
+  
+  for (k = 1; k < DCTSIZE2; k++) {
+    if ((temp = block[jpeg_natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+	ac_counts[0xF0]++;
+	r -= 16;
+      }
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      if (temp < 0)
+	temp = -temp;
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;		/* there must be at least one 1 bit */
+      while ((temp >>= 1))
+	nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > MAX_COEF_BITS)
+	ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+      
+      /* Count Huffman symbol for run length / number of bits */
+      ac_counts[(r << 4) + nbits]++;
+      
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    ac_counts[0]++;
+}
+
+
+/*
+ * Trial-encode one MCU's worth of Huffman-compressed coefficients.
+ * No data is actually output, so no suspension return is possible.
+ */
+
+METHODDEF(int)
+encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Take care of restart intervals if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      /* Re-initialize DC predictions to 0 */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+	entropy->saved.last_dc_val[ci] = 0;
+      /* Update restart state */
+      entropy->restarts_to_go = cinfo->restart_interval;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
+		    entropy->dc_count_ptrs[compptr->dc_tbl_no],
+		    entropy->ac_count_ptrs[compptr->ac_tbl_no]);
+    entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Generate the best Huffman code table for the given counts, fill htbl.
+ * Note this is also used by jcphuff.c.
+ *
+ * The JPEG standard requires that no symbol be assigned a codeword of all
+ * one bits (so that padding bits added at the end of a compressed segment
+ * can't look like a valid code).  Because of the canonical ordering of
+ * codewords, this just means that there must be an unused slot in the
+ * longest codeword length category.  Section K.2 of the JPEG spec suggests
+ * reserving such a slot by pretending that symbol 256 is a valid symbol
+ * with count 1.  In theory that's not optimal; giving it count zero but
+ * including it in the symbol set anyway should give a better Huffman code.
+ * But the theoretically better code actually seems to come out worse in
+ * practice, because it produces more all-ones bytes (which incur stuffed
+ * zero bytes in the final file).  In any case the difference is tiny.
+ *
+ * The JPEG standard requires Huffman codes to be no more than 16 bits long.
+ * If some symbols have a very small but nonzero probability, the Huffman tree
+ * must be adjusted to meet the code length restriction.  We currently use
+ * the adjustment method suggested in JPEG section K.2.  This method is *not*
+ * optimal; it may not choose the best possible limited-length code.  But
+ * typically only very-low-frequency symbols will be given less-than-optimal
+ * lengths, so the code is almost optimal.  Experimental comparisons against
+ * an optimal limited-length-code algorithm indicate that the difference is
+ * microscopic --- usually less than a hundredth of a percent of total size.
+ * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
+ */
+
+GLOBAL(void)
+jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
+{
+#define MAX_CLEN 32		/* assumed maximum initial code length */
+  unsigned short bits[MAX_CLEN+1];	/* bits[k] = # of symbols with code length k */
+  int codesize[257];		/* codesize[k] = code length of symbol k */
+  int others[257];		/* next symbol in current branch of tree */
+  int c1, c2;
+  int p, i, j;
+  long v;
+
+  /* This algorithm is explained in section K.2 of the JPEG standard */
+
+  MEMZERO(bits, SIZEOF(bits));
+  MEMZERO(codesize, SIZEOF(codesize));
+  for (i = 0; i < 257; i++)
+    others[i] = -1;		/* init links to empty */
+  
+  freq[256] = 1;		/* make sure 256 has a nonzero count */
+  /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
+   * that no real symbol is given code-value of all ones, because 256
+   * will be placed last in the largest codeword category.
+   */
+
+  /* Huffman's basic algorithm to assign optimal code lengths to symbols */
+
+  for (;;) {
+    /* Find the smallest nonzero frequency, set c1 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c1 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v) {
+	v = freq[i];
+	c1 = i;
+      }
+    }
+
+    /* Find the next smallest nonzero frequency, set c2 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c2 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v && i != c1) {
+	v = freq[i];
+	c2 = i;
+      }
+    }
+
+    /* Done if we've merged everything into one frequency */
+    if (c2 < 0)
+      break;
+    
+    /* Else merge the two counts/trees */
+    freq[c1] += freq[c2];
+    freq[c2] = 0;
+
+    /* Increment the codesize of everything in c1's tree branch */
+    codesize[c1]++;
+    while (others[c1] >= 0) {
+      c1 = others[c1];
+      codesize[c1]++;
+    }
+    
+    others[c1] = c2;		/* chain c2 onto c1's tree branch */
+    
+    /* Increment the codesize of everything in c2's tree branch */
+    codesize[c2]++;
+    while (others[c2] >= 0) {
+      c2 = others[c2];
+      codesize[c2]++;
+    }
+  }
+
+  /* Now count the number of symbols of each code length */
+  for (i = 0; i <= 256; i++) {
+    if (codesize[i]) {
+      /* The JPEG standard seems to think that this can't happen, */
+      /* but I'm paranoid... */
+      if (codesize[i] > MAX_CLEN)
+	ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
+
+      bits[codesize[i]]++;
+    }
+  }
+
+  /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
+   * Huffman procedure assigned any such lengths, we must adjust the coding.
+   * Here is what the JPEG spec says about how this next bit works:
+   * Since symbols are paired for the longest Huffman code, the symbols are
+   * removed from this length category two at a time.  The prefix for the pair
+   * (which is one bit shorter) is allocated to one of the pair; then,
+   * skipping the BITS entry for that prefix length, a code word from the next
+   * shortest nonzero BITS entry is converted into a prefix for two code words
+   * one bit longer.
+   */
+  
+  for (i = MAX_CLEN; i > 16; i--) {
+    while (bits[i] > 0) {
+      j = i - 2;		/* find length of new prefix to be used */
+      while (bits[j] == 0)
+	j--;
+      
+      bits[i] -= 2;		/* remove two symbols */
+      bits[i-1]++;		/* one goes in this length */
+      bits[j+1] += 2;		/* two new symbols in this length */
+      bits[j]--;		/* symbol of this length is now a prefix */
+    }
+  }
+
+  /* Remove the count for the pseudo-symbol 256 from the largest codelength */
+  while (bits[i] == 0)		/* find largest codelength still in use */
+    i--;
+  bits[i]--;
+  
+  /* Return final symbol counts (only for lengths 0..16) */
+  MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
+  
+  /* Return a list of the symbols sorted by code length */
+  /* It's not real clear to me why we don't need to consider the codelength
+   * changes made above, but the JPEG spec seems to think this works.
+   */
+  p = 0;
+  for (i = 1; i <= MAX_CLEN; i++) {
+    for (j = 0; j <= 255; j++) {
+      if (codesize[j] == i) {
+	htbl->huffval[p] = (unsigned char) j;
+	p++;
+      }
+    }
+  }
+
+  /* Set sent_table FALSE so updated table will be written to JPEG file. */
+  htbl->sent_table = FALSE;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, dctbl, actbl;
+  jpeg_component_info * compptr;
+  JHUFF_TBL **htblptr;
+  int did_dc[NUM_HUFF_TBLS];
+  int did_ac[NUM_HUFF_TBLS];
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  MEMZERO(did_dc, SIZEOF(did_dc));
+  MEMZERO(did_ac, SIZEOF(did_ac));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    if (! did_dc[dctbl]) {
+      htblptr = & cinfo->dc_huff_tbl_ptrs[dctbl];
+      if (*htblptr == NULL)
+	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[dctbl]);
+      did_dc[dctbl] = TRUE;
+    }
+    if (! did_ac[actbl]) {
+      htblptr = & cinfo->ac_huff_tbl_ptrs[actbl];
+      if (*htblptr == NULL)
+	*htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[actbl]);
+      did_ac[actbl] = TRUE;
+    }
+  }
+}
+
+
+#endif /* ENTROPY_OPT_SUPPORTED */
+
+
+/*
+ * Module initialization routine for Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_huff_encoder (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+#ifdef ENTROPY_OPT_SUPPORTED
+    entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
+#endif
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jchuff.h b/ml/dlib/dlib/external/libjpeg/jchuff.h
new file mode 100644
index 000000000..2d184ec55
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jchuff.h
@@ -0,0 +1,47 @@
+/*
+ * jchuff.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains declarations for Huffman entropy encoding routines
+ * that are shared between the sequential encoder (jchuff.c) and the
+ * progressive encoder (jcphuff.c).  No other modules need to see these.
+ */
+
+/* The legal range of a DCT coefficient is
+ *  -1024 .. +1023  for 8-bit data;
+ * -16384 .. +16383 for 12-bit data.
+ * Hence the magnitude should always fit in 10 or 14 bits respectively.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MAX_COEF_BITS 10
+#else
+#define MAX_COEF_BITS 14
+#endif
+
+/* Derived data constructed for each Huffman table */
+
+typedef struct {
+  unsigned int ehufco[256];	/* code for each symbol */
+  char ehufsi[256];		/* length of code for each symbol */
+  /* If no code has been allocated for a symbol S, ehufsi[S] contains 0 */
+} c_derived_tbl;
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_make_c_derived_tbl	jMkCDerived
+#define jpeg_gen_optimal_table	jGenOptTbl
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_c_derived_tbl
+	JPP((j_compress_ptr cinfo, int isDC, int tblno,
+	     c_derived_tbl ** pdtbl));
+
+/* Generate an optimal table definition given the specified counts */
+EXTERN(void) jpeg_gen_optimal_table
+	JPP((j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[]));
diff --git a/ml/dlib/dlib/external/libjpeg/jcinit.cpp b/ml/dlib/dlib/external/libjpeg/jcinit.cpp
new file mode 100644
index 000000000..2ca809cfd
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcinit.cpp
@@ -0,0 +1,72 @@
+/*
+ * jcinit.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains initialization logic for the JPEG compressor.
+ * This routine is in charge of selecting the modules to be executed and
+ * making an initialization call to each one.
+ *
+ * Logically, this code belongs in jcmaster.c.  It's split out because
+ * linking this routine implies linking the entire compression library.
+ * For a transcoding-only application, we want to be able to use jcmaster.c
+ * without linking in the whole library.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Master selection of compression modules.
+ * This is done once at the start of processing an image.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ */
+
+GLOBAL(void)
+jinit_compress_master (j_compress_ptr cinfo)
+{
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, FALSE /* full compression */);
+
+  /* Preprocessing */
+  if (! cinfo->raw_data_in) {
+    jinit_color_converter(cinfo);
+    jinit_downsampler(cinfo);
+    jinit_c_prep_controller(cinfo, FALSE /* never need full buffer here */);
+  }
+  /* Forward DCT */
+  jinit_forward_dct(cinfo);
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      jinit_phuff_encoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_encoder(cinfo);
+  }
+
+  /* Need a full-image coefficient buffer in any multi-pass mode. */
+  jinit_c_coef_controller(cinfo,
+		(int) (cinfo->num_scans > 1 || cinfo->optimize_coding));
+  jinit_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Write the datastream header (SOI) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcmainct.cpp b/ml/dlib/dlib/external/libjpeg/jcmainct.cpp
new file mode 100644
index 000000000..1e5f97a20
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcmainct.cpp
@@ -0,0 +1,293 @@
+/*
+ * jcmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for compression.
+ * The main buffer lies between the pre-processor and the JPEG
+ * compressor proper; it holds downsampled data in the JPEG colorspace.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Note: currently, there is no operating mode in which a full-image buffer
+ * is needed at this step.  If there were, that mode could not be used with
+ * "raw data" input, since this module is bypassed in that case.  However,
+ * we've left the code here for possible use in special applications.
+ */
+#undef FULL_MAIN_BUFFER_SUPPORTED
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_main_controller pub; /* public fields */
+
+  JDIMENSION cur_iMCU_row;	/* number of current iMCU row */
+  JDIMENSION rowgroup_ctr;	/* counts row groups received in iMCU row */
+  int suspended;		/* remember if we suspended output */
+  J_BUF_MODE pass_mode;		/* current operating mode */
+
+  /* If using just a strip buffer, this points to the entire set of buffers
+   * (we allocate one for each component).  In the full-image case, this
+   * points to the currently accessible strips of the virtual arrays.
+   */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  /* If using full-image storage, this array holds pointers to virtual-array
+   * control blocks for each component.  Unused if not full-image storage.
+   */
+  jvirt_sarray_ptr whole_image[MAX_COMPONENTS];
+#endif
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+METHODDEF(void) process_data_buffer_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Do nothing in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  main->cur_iMCU_row = 0;	/* initialize counters */
+  main->rowgroup_ctr = 0;
+  main->suspended = FALSE;
+  main->pass_mode = pass_mode;	/* save mode for use by process_data */
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    if (main->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+    main->pub.process_data = process_data_simple_main;
+    break;
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  case JBUF_SAVE_SOURCE:
+  case JBUF_CRANK_DEST:
+  case JBUF_SAVE_AND_PASS:
+    if (main->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    main->pub.process_data = process_data_buffer_main;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This routine handles the simple pass-through mode,
+ * where we have only a strip buffer.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Read input data if we haven't filled the main buffer yet */
+    if (main->rowgroup_ctr < DCTSIZE)
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) DCTSIZE);
+
+    /* If we don't have a full iMCU row buffered, return to application for
+     * more data.  Note that preprocessor will always pad to fill the iMCU row
+     * at the bottom of the image.
+     */
+    if (main->rowgroup_ctr != DCTSIZE)
+      return;
+
+    /* Send the completed row to the compressor */
+    if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+      /* If compressor did not consume the whole row, then we must need to
+       * suspend processing and return to the application.  In this situation
+       * we pretend we didn't yet consume the last input row; otherwise, if
+       * it happened to be the last row of the image, the application would
+       * think we were done.
+       */
+      if (! main->suspended) {
+	(*in_row_ctr)--;
+	main->suspended = TRUE;
+      }
+      return;
+    }
+    /* We did finish the row.  Undo our little suspension hack if a previous
+     * call suspended; then mark the main buffer empty.
+     */
+    if (main->suspended) {
+      (*in_row_ctr)++;
+      main->suspended = FALSE;
+    }
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+
+/*
+ * Process some data.
+ * This routine handles all of the modes that use a full-size buffer.
+ */
+
+METHODDEF(void)
+process_data_buffer_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci;
+  jpeg_component_info *compptr;
+  int writing = (main->pass_mode != JBUF_CRANK_DEST);
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Realign the virtual buffers if at the start of an iMCU row. */
+    if (main->rowgroup_ctr == 0) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	main->buffer[ci] = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, main->whole_image[ci],
+	   main->cur_iMCU_row * (compptr->v_samp_factor * DCTSIZE),
+	   (JDIMENSION) (compptr->v_samp_factor * DCTSIZE), writing);
+      }
+      /* In a read pass, pretend we just read some source data. */
+      if (! writing) {
+	*in_row_ctr += cinfo->max_v_samp_factor * DCTSIZE;
+	main->rowgroup_ctr = DCTSIZE;
+      }
+    }
+
+    /* If a write pass, read input data until the current iMCU row is full. */
+    /* Note: preprocessor will pad if necessary to fill the last iMCU row. */
+    if (writing) {
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) DCTSIZE);
+      /* Return to application if we need more data to fill the iMCU row. */
+      if (main->rowgroup_ctr < DCTSIZE)
+	return;
+    }
+
+    /* Emit data, unless this is a sink-only pass. */
+    if (main->pass_mode != JBUF_SAVE_SOURCE) {
+      if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+	/* If compressor did not consume the whole row, then we must need to
+	 * suspend processing and return to the application.  In this situation
+	 * we pretend we didn't yet consume the last input row; otherwise, if
+	 * it happened to be the last row of the image, the application would
+	 * think we were done.
+	 */
+	if (! main->suspended) {
+	  (*in_row_ctr)--;
+	  main->suspended = TRUE;
+	}
+	return;
+      }
+      /* We did finish the row.  Undo our little suspension hack if a previous
+       * call suspended; then mark the main buffer empty.
+       */
+      if (main->suspended) {
+	(*in_row_ctr)++;
+	main->suspended = FALSE;
+      }
+    }
+
+    /* If get here, we are done with this iMCU row.  Mark buffer empty. */
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+#endif /* FULL_MAIN_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_main_controller (j_compress_ptr cinfo, int need_full_buffer)
+{
+  my_main_ptr main;
+  int ci;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_c_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  /* We don't need to create a buffer in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  /* Create the buffer.  It holds downsampled data, so each component
+   * may be of a different size.
+   */
+  if (need_full_buffer) {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component */
+    /* Note we pad the bottom to a multiple of the iMCU height */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->whole_image[ci] = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 compptr->width_in_blocks * DCTSIZE,
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor) * DCTSIZE,
+	 (JDIMENSION) (compptr->v_samp_factor * DCTSIZE));
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    main->whole_image[0] = NULL; /* flag for no virtual arrays */
+#endif
+    /* Allocate a strip buffer for each component */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 compptr->width_in_blocks * DCTSIZE,
+	 (JDIMENSION) (compptr->v_samp_factor * DCTSIZE));
+    }
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcmarker.cpp b/ml/dlib/dlib/external/libjpeg/jcmarker.cpp
new file mode 100644
index 000000000..1bfd15c55
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcmarker.cpp
@@ -0,0 +1,664 @@
+/*
+ * jcmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write JPEG datastream markers.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_writer pub; /* public fields */
+
+  unsigned int last_restart_interval; /* last DRI value emitted; 0 after SOI */
+} my_marker_writer;
+
+typedef my_marker_writer * my_marker_ptr;
+
+
+/*
+ * Basic output routines.
+ *
+ * Note that we do not support suspension while writing a marker.
+ * Therefore, an application using suspension must ensure that there is
+ * enough buffer space for the initial markers (typ. 600-700 bytes) before
+ * calling jpeg_start_compress, and enough space to write the trailing EOI
+ * (a few bytes) before calling jpeg_finish_compress.  Multipass compression
+ * modes are not supported at all with suspension, so those two are the only
+ * points where markers will be written.
+ */
+
+LOCAL(void)
+emit_byte (j_compress_ptr cinfo, int val)
+/* Emit a byte */
+{
+  struct jpeg_destination_mgr * dest = cinfo->dest;
+
+  *(dest->next_output_byte)++ = (JOCTET) val;
+  if (--dest->free_in_buffer == 0) {
+    if (! (*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  }
+}
+
+
+LOCAL(void)
+emit_marker (j_compress_ptr cinfo, JPEG_MARKER mark)
+/* Emit a marker code */
+{
+  emit_byte(cinfo, 0xFF);
+  emit_byte(cinfo, (int) mark);
+}
+
+
+LOCAL(void)
+emit_2bytes (j_compress_ptr cinfo, int value)
+/* Emit a 2-byte integer; these are always MSB first in JPEG files */
+{
+  emit_byte(cinfo, (value >> 8) & 0xFF);
+  emit_byte(cinfo, value & 0xFF);
+}
+
+
+/*
+ * Routines to write specific marker types.
+ */
+
+LOCAL(int)
+emit_dqt (j_compress_ptr cinfo, int index)
+/* Emit a DQT marker */
+/* Returns the precision used (0 = 8bits, 1 = 16bits) for baseline checking */
+{
+  JQUANT_TBL * qtbl = cinfo->quant_tbl_ptrs[index];
+  int prec;
+  int i;
+
+  if (qtbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, index);
+
+  prec = 0;
+  for (i = 0; i < DCTSIZE2; i++) {
+    if (qtbl->quantval[i] > 255)
+      prec = 1;
+  }
+
+  if (! qtbl->sent_table) {
+    emit_marker(cinfo, M_DQT);
+
+    emit_2bytes(cinfo, prec ? DCTSIZE2*2 + 1 + 2 : DCTSIZE2 + 1 + 2);
+
+    emit_byte(cinfo, index + (prec<<4));
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      /* The table entries must be emitted in zigzag order. */
+      unsigned int qval = qtbl->quantval[jpeg_natural_order[i]];
+      if (prec)
+	emit_byte(cinfo, (int) (qval >> 8));
+      emit_byte(cinfo, (int) (qval & 0xFF));
+    }
+
+    qtbl->sent_table = TRUE;
+  }
+
+  return prec;
+}
+
+
+LOCAL(void)
+emit_dht (j_compress_ptr cinfo, int index, int is_ac)
+/* Emit a DHT marker */
+{
+  JHUFF_TBL * htbl;
+  int length, i;
+  
+  if (is_ac) {
+    htbl = cinfo->ac_huff_tbl_ptrs[index];
+    index += 0x10;		/* output index has AC bit set */
+  } else {
+    htbl = cinfo->dc_huff_tbl_ptrs[index];
+  }
+
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, index);
+  
+  if (! htbl->sent_table) {
+    emit_marker(cinfo, M_DHT);
+    
+    length = 0;
+    for (i = 1; i <= 16; i++)
+      length += htbl->bits[i];
+    
+    emit_2bytes(cinfo, length + 2 + 1 + 16);
+    emit_byte(cinfo, index);
+    
+    for (i = 1; i <= 16; i++)
+      emit_byte(cinfo, htbl->bits[i]);
+    
+    for (i = 0; i < length; i++)
+      emit_byte(cinfo, htbl->huffval[i]);
+    
+    htbl->sent_table = TRUE;
+  }
+}
+
+
+LOCAL(void)
+emit_dac (j_compress_ptr )//cinfo)
+/* Emit a DAC marker */
+/* Since the useful info is so small, we want to emit all the tables in */
+/* one DAC marker.  Therefore this routine does its own scan of the table. */
+{
+#ifdef C_ARITH_CODING_SUPPORTED
+  char dc_in_use[NUM_ARITH_TBLS];
+  char ac_in_use[NUM_ARITH_TBLS];
+  int length, i;
+  jpeg_component_info *compptr;
+  
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    dc_in_use[i] = ac_in_use[i] = 0;
+  
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    dc_in_use[compptr->dc_tbl_no] = 1;
+    ac_in_use[compptr->ac_tbl_no] = 1;
+  }
+  
+  length = 0;
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    length += dc_in_use[i] + ac_in_use[i];
+  
+  emit_marker(cinfo, M_DAC);
+  
+  emit_2bytes(cinfo, length*2 + 2);
+  
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    if (dc_in_use[i]) {
+      emit_byte(cinfo, i);
+      emit_byte(cinfo, cinfo->arith_dc_L[i] + (cinfo->arith_dc_U[i]<<4));
+    }
+    if (ac_in_use[i]) {
+      emit_byte(cinfo, i + 0x10);
+      emit_byte(cinfo, cinfo->arith_ac_K[i]);
+    }
+  }
+#endif /* C_ARITH_CODING_SUPPORTED */
+}
+
+
+LOCAL(void)
+emit_dri (j_compress_ptr cinfo)
+/* Emit a DRI marker */
+{
+  emit_marker(cinfo, M_DRI);
+  
+  emit_2bytes(cinfo, 4);	/* fixed length */
+
+  emit_2bytes(cinfo, (int) cinfo->restart_interval);
+}
+
+
+LOCAL(void)
+emit_sof (j_compress_ptr cinfo, JPEG_MARKER code)
+/* Emit a SOF marker */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, code);
+  
+  emit_2bytes(cinfo, 3 * cinfo->num_components + 2 + 5 + 1); /* length */
+
+  /* Make sure image isn't bigger than SOF field can handle */
+  if ((long) cinfo->image_height > 65535L ||
+      (long) cinfo->image_width > 65535L)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) 65535);
+
+  emit_byte(cinfo, cinfo->data_precision);
+  emit_2bytes(cinfo, (int) cinfo->image_height);
+  emit_2bytes(cinfo, (int) cinfo->image_width);
+
+  emit_byte(cinfo, cinfo->num_components);
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    emit_byte(cinfo, compptr->component_id);
+    emit_byte(cinfo, (compptr->h_samp_factor << 4) + compptr->v_samp_factor);
+    emit_byte(cinfo, compptr->quant_tbl_no);
+  }
+}
+
+
+LOCAL(void)
+emit_sos (j_compress_ptr cinfo)
+/* Emit a SOS marker */
+{
+  int i, td, ta;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, M_SOS);
+  
+  emit_2bytes(cinfo, 2 * cinfo->comps_in_scan + 2 + 1 + 3); /* length */
+  
+  emit_byte(cinfo, cinfo->comps_in_scan);
+  
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    emit_byte(cinfo, compptr->component_id);
+    td = compptr->dc_tbl_no;
+    ta = compptr->ac_tbl_no;
+    if (cinfo->progressive_mode) {
+      /* Progressive mode: only DC or only AC tables are used in one scan;
+       * furthermore, Huffman coding of DC refinement uses no table at all.
+       * We emit 0 for unused field(s); this is recommended by the P&M text
+       * but does not seem to be specified in the standard.
+       */
+      if (cinfo->Ss == 0) {
+	ta = 0;			/* DC scan */
+	if (cinfo->Ah != 0 && !cinfo->arith_code)
+	  td = 0;		/* no DC table either */
+      } else {
+	td = 0;			/* AC scan */
+      }
+    }
+    emit_byte(cinfo, (td << 4) + ta);
+  }
+
+  emit_byte(cinfo, cinfo->Ss);
+  emit_byte(cinfo, cinfo->Se);
+  emit_byte(cinfo, (cinfo->Ah << 4) + cinfo->Al);
+}
+
+
+LOCAL(void)
+emit_jfif_app0 (j_compress_ptr cinfo)
+/* Emit a JFIF-compliant APP0 marker */
+{
+  /*
+   * Length of APP0 block	(2 bytes)
+   * Block ID			(4 bytes - ASCII "JFIF")
+   * Zero byte			(1 byte to terminate the ID string)
+   * Version Major, Minor	(2 bytes - major first)
+   * Units			(1 byte - 0x00 = none, 0x01 = inch, 0x02 = cm)
+   * Xdpu			(2 bytes - dots per unit horizontal)
+   * Ydpu			(2 bytes - dots per unit vertical)
+   * Thumbnail X size		(1 byte)
+   * Thumbnail Y size		(1 byte)
+   */
+  
+  emit_marker(cinfo, M_APP0);
+  
+  emit_2bytes(cinfo, 2 + 4 + 1 + 2 + 1 + 2 + 2 + 1 + 1); /* length */
+
+  emit_byte(cinfo, 0x4A);	/* Identifier: ASCII "JFIF" */
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0x49);
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0);
+  emit_byte(cinfo, cinfo->JFIF_major_version); /* Version fields */
+  emit_byte(cinfo, cinfo->JFIF_minor_version);
+  emit_byte(cinfo, cinfo->density_unit); /* Pixel size information */
+  emit_2bytes(cinfo, (int) cinfo->X_density);
+  emit_2bytes(cinfo, (int) cinfo->Y_density);
+  emit_byte(cinfo, 0);		/* No thumbnail image */
+  emit_byte(cinfo, 0);
+}
+
+
+LOCAL(void)
+emit_adobe_app14 (j_compress_ptr cinfo)
+/* Emit an Adobe APP14 marker */
+{
+  /*
+   * Length of APP14 block	(2 bytes)
+   * Block ID			(5 bytes - ASCII "Adobe")
+   * Version Number		(2 bytes - currently 100)
+   * Flags0			(2 bytes - currently 0)
+   * Flags1			(2 bytes - currently 0)
+   * Color transform		(1 byte)
+   *
+   * Although Adobe TN 5116 mentions Version = 101, all the Adobe files
+   * now in circulation seem to use Version = 100, so that's what we write.
+   *
+   * We write the color transform byte as 1 if the JPEG color space is
+   * YCbCr, 2 if it's YCCK, 0 otherwise.  Adobe's definition has to do with
+   * whether the encoder performed a transformation, which is pretty useless.
+   */
+  
+  emit_marker(cinfo, M_APP14);
+  
+  emit_2bytes(cinfo, 2 + 5 + 2 + 2 + 2 + 1); /* length */
+
+  emit_byte(cinfo, 0x41);	/* Identifier: ASCII "Adobe" */
+  emit_byte(cinfo, 0x64);
+  emit_byte(cinfo, 0x6F);
+  emit_byte(cinfo, 0x62);
+  emit_byte(cinfo, 0x65);
+  emit_2bytes(cinfo, 100);	/* Version */
+  emit_2bytes(cinfo, 0);	/* Flags0 */
+  emit_2bytes(cinfo, 0);	/* Flags1 */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_YCbCr:
+    emit_byte(cinfo, 1);	/* Color transform = 1 */
+    break;
+  case JCS_YCCK:
+    emit_byte(cinfo, 2);	/* Color transform = 2 */
+    break;
+  default:
+    emit_byte(cinfo, 0);	/* Color transform = 0 */
+    break;
+  }
+}
+
+
+/*
+ * These routines allow writing an arbitrary marker with parameters.
+ * The only intended use is to emit COM or APPn markers after calling
+ * write_file_header and before calling write_frame_header.
+ * Other uses are not guaranteed to produce desirable results.
+ * Counting the parameter bytes properly is the caller's responsibility.
+ */
+
+METHODDEF(void)
+write_marker_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+/* Emit an arbitrary marker header */
+{
+  if (datalen > (unsigned int) 65533)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  emit_marker(cinfo, (JPEG_MARKER) marker);
+
+  emit_2bytes(cinfo, (int) (datalen + 2));	/* total length */
+}
+
+METHODDEF(void)
+write_marker_byte (j_compress_ptr cinfo, int val)
+/* Emit one byte of marker parameters following write_marker_header */
+{
+  emit_byte(cinfo, val);
+}
+
+
+/*
+ * Write datastream header.
+ * This consists of an SOI and optional APPn markers.
+ * We recommend use of the JFIF marker, but not the Adobe marker,
+ * when using YCbCr or grayscale data.  The JFIF marker should NOT
+ * be used for any other JPEG colorspace.  The Adobe marker is helpful
+ * to distinguish RGB, CMYK, and YCCK colorspaces.
+ * Note that an application can write additional header markers after
+ * jpeg_start_compress returns.
+ */
+
+METHODDEF(void)
+write_file_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  emit_marker(cinfo, M_SOI);	/* first the SOI */
+
+  /* SOI is defined to reset restart interval to 0 */
+  marker->last_restart_interval = 0;
+
+  if (cinfo->write_JFIF_header)	/* next an optional JFIF APP0 */
+    emit_jfif_app0(cinfo);
+  if (cinfo->write_Adobe_marker) /* next an optional Adobe APP14 */
+    emit_adobe_app14(cinfo);
+}
+
+
+/*
+ * Write frame header.
+ * This consists of DQT and SOFn markers.
+ * Note that we do not emit the SOF until we have emitted the DQT(s).
+ * This avoids compatibility problems with incorrect implementations that
+ * try to error-check the quant table numbers as soon as they see the SOF.
+ */
+
+METHODDEF(void)
+write_frame_header (j_compress_ptr cinfo)
+{
+  int ci, prec;
+  int is_baseline;
+  jpeg_component_info *compptr;
+  
+  /* Emit DQT for each quantization table.
+   * Note that emit_dqt() suppresses any duplicate tables.
+   */
+  prec = 0;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    prec += emit_dqt(cinfo, compptr->quant_tbl_no);
+  }
+  /* now prec is nonzero iff there are any 16-bit quant tables. */
+
+  /* Check for a non-baseline specification.
+   * Note we assume that Huffman table numbers won't be changed later.
+   */
+  if (cinfo->arith_code || cinfo->progressive_mode ||
+      cinfo->data_precision != 8) {
+    is_baseline = FALSE;
+  } else {
+    is_baseline = TRUE;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (compptr->dc_tbl_no > 1 || compptr->ac_tbl_no > 1)
+	is_baseline = FALSE;
+    }
+    if (prec && is_baseline) {
+      is_baseline = FALSE;
+      /* If it's baseline except for quantizer size, warn the user */
+      TRACEMS(cinfo, 0, JTRC_16BIT_TABLES);
+    }
+  }
+
+  /* Emit the proper SOF marker */
+  if (cinfo->arith_code) {
+    emit_sof(cinfo, M_SOF9);	/* SOF code for arithmetic coding */
+  } else {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF2);	/* SOF code for progressive Huffman */
+    else if (is_baseline)
+      emit_sof(cinfo, M_SOF0);	/* SOF code for baseline implementation */
+    else
+      emit_sof(cinfo, M_SOF1);	/* SOF code for non-baseline Huffman file */
+  }
+}
+
+
+/*
+ * Write scan header.
+ * This consists of DHT or DAC markers, optional DRI, and SOS.
+ * Compressed data will be written following the SOS.
+ */
+
+METHODDEF(void)
+write_scan_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  int i;
+  jpeg_component_info *compptr;
+
+  if (cinfo->arith_code) {
+    /* Emit arith conditioning info.  We may have some duplication
+     * if the file has multiple scans, but it's so small it's hardly
+     * worth worrying about.
+     */
+    emit_dac(cinfo);
+  } else {
+    /* Emit Huffman tables.
+     * Note that emit_dht() suppresses any duplicate tables.
+     */
+    for (i = 0; i < cinfo->comps_in_scan; i++) {
+      compptr = cinfo->cur_comp_info[i];
+      if (cinfo->progressive_mode) {
+	/* Progressive mode: only DC or only AC tables are used in one scan */
+	if (cinfo->Ss == 0) {
+	  if (cinfo->Ah == 0)	/* DC needs no table for refinement scan */
+	    emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+	} else {
+	  emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+	}
+      } else {
+	/* Sequential mode: need both DC and AC tables */
+	emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+	emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+      }
+    }
+  }
+
+  /* Emit DRI if required --- note that DRI value could change for each scan.
+   * We avoid wasting space with unnecessary DRIs, however.
+   */
+  if (cinfo->restart_interval != marker->last_restart_interval) {
+    emit_dri(cinfo);
+    marker->last_restart_interval = cinfo->restart_interval;
+  }
+
+  emit_sos(cinfo);
+}
+
+
+/*
+ * Write datastream trailer.
+ */
+
+METHODDEF(void)
+write_file_trailer (j_compress_ptr cinfo)
+{
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Write an abbreviated table-specification datastream.
+ * This consists of SOI, DQT and DHT tables, and EOI.
+ * Any table that is defined and not marked sent_table = TRUE will be
+ * emitted.  Note that all tables will be marked sent_table = TRUE at exit.
+ */
+
+METHODDEF(void)
+write_tables_only (j_compress_ptr cinfo)
+{
+  int i;
+
+  emit_marker(cinfo, M_SOI);
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if (cinfo->quant_tbl_ptrs[i] != NULL)
+      (void) emit_dqt(cinfo, i);
+  }
+
+  if (! cinfo->arith_code) {
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      if (cinfo->dc_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, FALSE);
+      if (cinfo->ac_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, TRUE);
+    }
+  }
+
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Initialize the marker writer module.
+ */
+
+GLOBAL(void)
+jinit_marker_writer (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker;
+
+  /* Create the subobject */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_marker_writer));
+  cinfo->marker = (struct jpeg_marker_writer *) marker;
+  /* Initialize method pointers */
+  marker->pub.write_file_header = write_file_header;
+  marker->pub.write_frame_header = write_frame_header;
+  marker->pub.write_scan_header = write_scan_header;
+  marker->pub.write_file_trailer = write_file_trailer;
+  marker->pub.write_tables_only = write_tables_only;
+  marker->pub.write_marker_header = write_marker_header;
+  marker->pub.write_marker_byte = write_marker_byte;
+  /* Initialize private state */
+  marker->last_restart_interval = 0;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcmaster.cpp b/ml/dlib/dlib/external/libjpeg/jcmaster.cpp
new file mode 100644
index 000000000..3e1b1d711
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcmaster.cpp
@@ -0,0 +1,590 @@
+/*
+ * jcmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG compressor.
+ * These routines are concerned with parameter validation, initial setup,
+ * and inter-pass control (determining the number of passes and the work 
+ * to be done in each pass).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef enum {
+	main_pass,		/* input data, also do first output step */
+	huff_opt_pass,		/* Huffman code optimization pass */
+	output_pass		/* data output pass */
+} c_pass_type;
+
+typedef struct {
+  struct jpeg_comp_master pub;	/* public fields */
+
+  c_pass_type pass_type;	/* the type of the current pass */
+
+  int pass_number;		/* # of passes completed */
+  int total_passes;		/* total # of passes needed */
+
+  int scan_number;		/* current index in scan_info[] */
+} my_comp_master;
+
+typedef my_comp_master * my_master_ptr;
+
+
+/*
+ * Support routines that do various essential calculations.
+ */
+
+LOCAL(void)
+initial_setup (j_compress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  /* Sanity check on image dimensions */
+  if (cinfo->image_height <= 0 || cinfo->image_width <= 0
+      || cinfo->num_components <= 0 || cinfo->input_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->image_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->image_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* Width of an input scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->image_width * (long) cinfo->input_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Fill in the correct component_index value; don't rely on application */
+    compptr->component_index = ci;
+    /* For compression, we never do DCT scaling. */
+    compptr->DCT_scaled_size = DCTSIZE;
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) cinfo->max_v_samp_factor);
+    /* Mark component needed (this flag isn't actually used for compression) */
+    compptr->component_needed = TRUE;
+  }
+
+  /* Compute number of fully interleaved MCU rows (number of times that
+   * main controller will call coefficient controller).
+   */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->image_height,
+		  (long) (cinfo->max_v_samp_factor*DCTSIZE));
+}
+
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+
+LOCAL(void)
+validate_script (j_compress_ptr cinfo)
+/* Verify that the scan script in cinfo->scan_info[] is valid; also
+ * determine whether it uses progressive JPEG, and set cinfo->progressive_mode.
+ */
+{
+  const jpeg_scan_info * scanptr;
+  int scanno, ncomps, ci, coefi, thisi;
+  int Ss, Se, Ah, Al;
+  int component_sent[MAX_COMPONENTS];
+#ifdef C_PROGRESSIVE_SUPPORTED
+  int * last_bitpos_ptr;
+  int last_bitpos[MAX_COMPONENTS][DCTSIZE2];
+  /* -1 until that coefficient has been seen; then last Al for it */
+#endif
+
+  if (cinfo->num_scans <= 0)
+    ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, 0);
+
+  /* For sequential JPEG, all scans must have Ss=0, Se=DCTSIZE2-1;
+   * for progressive JPEG, no scan can have this.
+   */
+  scanptr = cinfo->scan_info;
+  if (scanptr->Ss != 0 || scanptr->Se != DCTSIZE2-1) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    cinfo->progressive_mode = TRUE;
+    last_bitpos_ptr = & last_bitpos[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      for (coefi = 0; coefi < DCTSIZE2; coefi++)
+	*last_bitpos_ptr++ = -1;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      component_sent[ci] = FALSE;
+  }
+
+  for (scanno = 1; scanno <= cinfo->num_scans; scanptr++, scanno++) {
+    /* Validate component indexes */
+    ncomps = scanptr->comps_in_scan;
+    if (ncomps <= 0 || ncomps > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, ncomps, MAX_COMPS_IN_SCAN);
+    for (ci = 0; ci < ncomps; ci++) {
+      thisi = scanptr->component_index[ci];
+      if (thisi < 0 || thisi >= cinfo->num_components)
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+      /* Components must appear in SOF order within each scan */
+      if (ci > 0 && thisi <= scanptr->component_index[ci-1])
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+    }
+    /* Validate progression parameters */
+    Ss = scanptr->Ss;
+    Se = scanptr->Se;
+    Ah = scanptr->Ah;
+    Al = scanptr->Al;
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      /* The JPEG spec simply gives the ranges 0..13 for Ah and Al, but that
+       * seems wrong: the upper bound ought to depend on data precision.
+       * Perhaps they really meant 0..N+1 for N-bit precision.
+       * Here we allow 0..10 for 8-bit data; Al larger than 10 results in
+       * out-of-range reconstructed DC values during the first DC scan,
+       * which might cause problems for some decoders.
+       */
+#if BITS_IN_JSAMPLE == 8
+#define MAX_AH_AL 10
+#else
+#define MAX_AH_AL 13
+#endif
+      if (Ss < 0 || Ss >= DCTSIZE2 || Se < Ss || Se >= DCTSIZE2 ||
+	  Ah < 0 || Ah > MAX_AH_AL || Al < 0 || Al > MAX_AH_AL)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      if (Ss == 0) {
+	if (Se != 0)		/* DC and AC together not OK */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      } else {
+	if (ncomps != 1)	/* AC scans must be for only one component */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      }
+      for (ci = 0; ci < ncomps; ci++) {
+	last_bitpos_ptr = & last_bitpos[scanptr->component_index[ci]][0];
+	if (Ss != 0 && last_bitpos_ptr[0] < 0) /* AC without prior DC scan */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	for (coefi = Ss; coefi <= Se; coefi++) {
+	  if (last_bitpos_ptr[coefi] < 0) {
+	    /* first scan of this coefficient */
+	    if (Ah != 0)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  } else {
+	    /* not first scan */
+	    if (Ah != last_bitpos_ptr[coefi] || Al != Ah-1)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  }
+	  last_bitpos_ptr[coefi] = Al;
+	}
+      }
+#endif
+    } else {
+      /* For sequential JPEG, all progression parameters must be these: */
+      if (Ss != 0 || Se != DCTSIZE2-1 || Ah != 0 || Al != 0)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      /* Make sure components are not sent twice */
+      for (ci = 0; ci < ncomps; ci++) {
+	thisi = scanptr->component_index[ci];
+	if (component_sent[thisi])
+	  ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+	component_sent[thisi] = TRUE;
+      }
+    }
+  }
+
+  /* Now verify that everything got sent. */
+  if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    /* For progressive mode, we only check that at least some DC data
+     * got sent for each component; the spec does not require that all bits
+     * of all coefficients be transmitted.  Would it be wiser to enforce
+     * transmission of all coefficient bits??
+     */
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (last_bitpos[ci][0] < 0)
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+#endif
+  } else {
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (! component_sent[ci])
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+  }
+}
+
+#endif /* C_MULTISCAN_FILES_SUPPORTED */
+
+
+LOCAL(void)
+select_scan_parameters (j_compress_ptr cinfo)
+/* Set up the scan parameters for the current scan */
+{
+  int ci;
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+  if (cinfo->scan_info != NULL) {
+    /* Prepare for current scan --- the script is already validated */
+    my_master_ptr master = (my_master_ptr) cinfo->master;
+    const jpeg_scan_info * scanptr = cinfo->scan_info + master->scan_number;
+
+    cinfo->comps_in_scan = scanptr->comps_in_scan;
+    for (ci = 0; ci < scanptr->comps_in_scan; ci++) {
+      cinfo->cur_comp_info[ci] =
+	&cinfo->comp_info[scanptr->component_index[ci]];
+    }
+    cinfo->Ss = scanptr->Ss;
+    cinfo->Se = scanptr->Se;
+    cinfo->Ah = scanptr->Ah;
+    cinfo->Al = scanptr->Al;
+  }
+  else
+#endif
+  {
+    /* Prepare for single sequential-JPEG scan containing all components */
+    if (cinfo->num_components > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPS_IN_SCAN);
+    cinfo->comps_in_scan = cinfo->num_components;
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      cinfo->cur_comp_info[ci] = &cinfo->comp_info[ci];
+    }
+    cinfo->Ss = 0;
+    cinfo->Se = DCTSIZE2-1;
+    cinfo->Ah = 0;
+    cinfo->Al = 0;
+  }
+}
+
+
+LOCAL(void)
+per_scan_setup (j_compress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] are already set */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = DCTSIZE;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width,
+		    (long) (cinfo->max_h_samp_factor*DCTSIZE));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height,
+		    (long) (cinfo->max_v_samp_factor*DCTSIZE));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * DCTSIZE;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > C_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+
+  /* Convert restart specified in rows to actual MCU count. */
+  /* Note that count must fit in 16 bits, so we provide limiting. */
+  if (cinfo->restart_in_rows > 0) {
+    long nominal = (long) cinfo->restart_in_rows * (long) cinfo->MCUs_per_row;
+    cinfo->restart_interval = (unsigned int) MIN(nominal, 65535L);
+  }
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each pass.  We determine which modules
+ * will be active during this pass and give them appropriate start_pass calls.
+ * We also set is_last_pass to indicate whether any more passes will be
+ * required.
+ */
+
+METHODDEF(void)
+prepare_for_pass (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  switch (master->pass_type) {
+  case main_pass:
+    /* Initial pass: will collect input data, and do either Huffman
+     * optimization or data output for the first scan.
+     */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (! cinfo->raw_data_in) {
+      (*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->downsample->start_pass) (cinfo);
+      (*cinfo->prep->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+    (*cinfo->fdct->start_pass) (cinfo);
+    (*cinfo->entropy->start_pass) (cinfo, cinfo->optimize_coding);
+    (*cinfo->coef->start_pass) (cinfo,
+				(master->total_passes > 1 ?
+				 JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+    (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    if (cinfo->optimize_coding) {
+      /* No immediate data output; postpone writing frame/scan headers */
+      master->pub.call_pass_startup = FALSE;
+    } else {
+      /* Will write frame/scan headers at first jpeg_write_scanlines call */
+      master->pub.call_pass_startup = TRUE;
+    }
+    break;
+#ifdef ENTROPY_OPT_SUPPORTED
+  case huff_opt_pass:
+    /* Do Huffman optimization for a scan after the first one. */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (cinfo->Ss != 0 || cinfo->Ah == 0 || cinfo->arith_code) {
+      (*cinfo->entropy->start_pass) (cinfo, TRUE);
+      (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+      master->pub.call_pass_startup = FALSE;
+      break;
+    }
+    /* Special case: Huffman DC refinement scans need no Huffman table
+     * and therefore we can skip the optimization pass for them.
+     */
+    master->pass_type = output_pass;
+    master->pass_number++;
+    /*FALLTHROUGH*/
+#endif
+  case output_pass:
+    /* Do a data-output pass. */
+    /* We need not repeat per-scan setup if prior optimization pass did it. */
+    if (! cinfo->optimize_coding) {
+      select_scan_parameters(cinfo);
+      per_scan_setup(cinfo);
+    }
+    (*cinfo->entropy->start_pass) (cinfo, FALSE);
+    (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+    /* We emit frame/scan headers now */
+    if (master->scan_number == 0)
+      (*cinfo->marker->write_frame_header) (cinfo);
+    (*cinfo->marker->write_scan_header) (cinfo);
+    master->pub.call_pass_startup = FALSE;
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+  }
+
+  master->pub.is_last_pass = (master->pass_number == master->total_passes-1);
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->total_passes;
+  }
+}
+
+
+/*
+ * Special start-of-pass hook.
+ * This is called by jpeg_write_scanlines if call_pass_startup is TRUE.
+ * In single-pass processing, we need this hook because we don't want to
+ * write frame/scan headers during jpeg_start_compress; we want to let the
+ * application write COM markers etc. between jpeg_start_compress and the
+ * jpeg_write_scanlines loop.
+ * In multi-pass processing, this routine is not used.
+ */
+
+METHODDEF(void)
+pass_startup (j_compress_ptr cinfo)
+{
+  cinfo->master->call_pass_startup = FALSE; /* reset flag so call only once */
+
+  (*cinfo->marker->write_frame_header) (cinfo);
+  (*cinfo->marker->write_scan_header) (cinfo);
+}
+
+
+/*
+ * Finish up at end of pass.
+ */
+
+METHODDEF(void)
+finish_pass_master (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* The entropy coder always needs an end-of-pass call,
+   * either to analyze statistics or to flush its output buffer.
+   */
+  (*cinfo->entropy->finish_pass) (cinfo);
+
+  /* Update state for next pass */
+  switch (master->pass_type) {
+  case main_pass:
+    /* next pass is either output of scan 0 (after optimization)
+     * or output of scan 1 (if no optimization).
+     */
+    master->pass_type = output_pass;
+    if (! cinfo->optimize_coding)
+      master->scan_number++;
+    break;
+  case huff_opt_pass:
+    /* next pass is always output of current scan */
+    master->pass_type = output_pass;
+    break;
+  case output_pass:
+    /* next pass is either optimization or output of next scan */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    master->scan_number++;
+    break;
+  }
+
+  master->pass_number++;
+}
+
+
+/*
+ * Initialize master compression control.
+ */
+
+GLOBAL(void)
+jinit_c_master_control (j_compress_ptr cinfo, int transcode_only)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_comp_master));
+  cinfo->master = (struct jpeg_comp_master *) master;
+  master->pub.prepare_for_pass = prepare_for_pass;
+  master->pub.pass_startup = pass_startup;
+  master->pub.finish_pass = finish_pass_master;
+  master->pub.is_last_pass = FALSE;
+
+  /* Validate parameters, determine derived values */
+  initial_setup(cinfo);
+
+  if (cinfo->scan_info != NULL) {
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+    validate_script(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    cinfo->num_scans = 1;
+  }
+
+  if (cinfo->progressive_mode)	/*  TEMPORARY HACK ??? */
+    cinfo->optimize_coding = TRUE; /* assume default tables no good for progressive mode */
+
+  /* Initialize my private state */
+  if (transcode_only) {
+    /* no main pass in transcoding */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    else
+      master->pass_type = output_pass;
+  } else {
+    /* for normal compression, first pass is always this type: */
+    master->pass_type = main_pass;
+  }
+  master->scan_number = 0;
+  master->pass_number = 0;
+  if (cinfo->optimize_coding)
+    master->total_passes = cinfo->num_scans * 2;
+  else
+    master->total_passes = cinfo->num_scans;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcomapi.cpp b/ml/dlib/dlib/external/libjpeg/jcomapi.cpp
new file mode 100644
index 000000000..9b1fa7568
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcomapi.cpp
@@ -0,0 +1,106 @@
+/*
+ * jcomapi.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface routines that are used for both
+ * compression and decompression.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Abort processing of a JPEG compression or decompression operation,
+ * but don't destroy the object itself.
+ *
+ * For this, we merely clean up all the nonpermanent memory pools.
+ * Note that temp files (virtual arrays) are not allowed to belong to
+ * the permanent pool, so we will be able to close all temp files here.
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_abort (j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Do nothing if called on a not-initialized or destroyed JPEG object. */
+  if (cinfo->mem == NULL)
+    return;
+
+  /* Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS-1; pool > JPOOL_PERMANENT; pool--) {
+    (*cinfo->mem->free_pool) (cinfo, pool);
+  }
+
+  /* Reset overall state for possible reuse of object */
+  if (cinfo->is_decompressor) {
+    cinfo->global_state = DSTATE_START;
+    /* Try to keep application from accessing now-deleted marker list.
+     * A bit kludgy to do it here, but this is the most central place.
+     */
+    ((j_decompress_ptr) cinfo)->marker_list = NULL;
+  } else {
+    cinfo->global_state = CSTATE_START;
+  }
+}
+
+
+/*
+ * Destruction of a JPEG object.
+ *
+ * Everything gets deallocated except the master jpeg_compress_struct itself
+ * and the error manager struct.  Both of these are supplied by the application
+ * and must be freed, if necessary, by the application.  (Often they are on
+ * the stack and so don't need to be freed anyway.)
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_destroy (j_common_ptr cinfo)
+{
+  /* We need only tell the memory manager to release everything. */
+  /* NB: mem pointer is NULL if memory mgr failed to initialize. */
+  if (cinfo->mem != NULL)
+    (*cinfo->mem->self_destruct) (cinfo);
+  cinfo->mem = NULL;		/* be safe if jpeg_destroy is called twice */
+  cinfo->global_state = 0;	/* mark it destroyed */
+}
+
+
+/*
+ * Convenience routines for allocating quantization and Huffman tables.
+ * (Would jutils.c be a more reasonable place to put these?)
+ */
+
+GLOBAL(JQUANT_TBL *)
+jpeg_alloc_quant_table (j_common_ptr cinfo)
+{
+  JQUANT_TBL *tbl;
+
+  tbl = (JQUANT_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JQUANT_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}
+
+
+GLOBAL(JHUFF_TBL *)
+jpeg_alloc_huff_table (j_common_ptr cinfo)
+{
+  JHUFF_TBL *tbl;
+
+  tbl = (JHUFF_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JHUFF_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jconfig.h b/ml/dlib/dlib/external/libjpeg/jconfig.h
new file mode 100644
index 000000000..9594ec56b
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jconfig.h
@@ -0,0 +1,45 @@
+/* jconfig.h.  Generated automatically by configure.  */
+/* jconfig.cfg --- source file edited by configure script */
+/* see jconfig.doc for explanations */
+
+#define HAVE_PROTOTYPES 
+#define HAVE_UNSIGNED_CHAR 
+#define HAVE_UNSIGNED_SHORT 
+#undef void
+#undef const
+#undef CHAR_IS_UNSIGNED
+#define HAVE_STDDEF_H 
+#define HAVE_STDLIB_H 
+#undef NEED_BSD_STRINGS
+#undef NEED_SYS_TYPES_H
+#undef NEED_FAR_POINTERS
+#undef NEED_SHORT_EXTERNAL_NAMES
+/* Define this if you get warnings about undefined structures. */
+#undef INCOMPLETE_TYPES_BROKEN
+
+#ifdef JPEG_INTERNALS
+
+#undef RIGHT_SHIFT_IS_UNSIGNED
+#define INLINE __inline__
+/* These are for configuring the JPEG memory manager. */
+#undef DEFAULT_MAX_MEM
+#undef NO_MKTEMP
+
+#endif /* JPEG_INTERNALS */
+
+#ifdef JPEG_CJPEG_DJPEG
+
+#define BMP_SUPPORTED		/* BMP image file format */
+#define GIF_SUPPORTED		/* GIF image file format */
+#define PPM_SUPPORTED		/* PBMPLUS PPM/PGM image file format */
+#undef RLE_SUPPORTED		/* Utah RLE image file format */
+#define TARGA_SUPPORTED		/* Targa image file format */
+
+#undef TWO_FILE_COMMANDLINE
+#undef NEED_SIGNAL_CATCHER
+#undef DONT_USE_B_MODE
+
+/* Define this if you want percent-done progress reports from cjpeg/djpeg. */
+#undef PROGRESS_REPORT
+
+#endif /* JPEG_CJPEG_DJPEG */
diff --git a/ml/dlib/dlib/external/libjpeg/jcparam.cpp b/ml/dlib/dlib/external/libjpeg/jcparam.cpp
new file mode 100644
index 000000000..a87ef2079
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcparam.cpp
@@ -0,0 +1,610 @@
+/*
+ * jcparam.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains optional default-setting code for the JPEG compressor.
+ * Applications do not have to use this file, but those that don't use it
+ * must know a lot more about the innards of the JPEG code.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Quantization table setup routines
+ */
+
+GLOBAL(void)
+jpeg_add_quant_table (j_compress_ptr cinfo, int which_tbl,
+		      const unsigned int *basic_table,
+		      int scale_factor, int force_baseline)
+/* Define a quantization table equal to the basic_table times
+ * a scale factor (given as a percentage).
+ * If force_baseline is TRUE, the computed quantization table entries
+ * are limited to 1..255 for JPEG baseline compatibility.
+ */
+{
+  JQUANT_TBL ** qtblptr;
+  int i;
+  long temp;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (which_tbl < 0 || which_tbl >= NUM_QUANT_TBLS)
+    ERREXIT1(cinfo, JERR_DQT_INDEX, which_tbl);
+
+  qtblptr = & cinfo->quant_tbl_ptrs[which_tbl];
+
+  if (*qtblptr == NULL)
+    *qtblptr = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = ((long) basic_table[i] * scale_factor + 50L) / 100L;
+    /* limit the values to the valid range */
+    if (temp <= 0L) temp = 1L;
+    if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
+    if (force_baseline && temp > 255L)
+      temp = 255L;		/* limit to baseline range if requested */
+    (*qtblptr)->quantval[i] = (unsigned short) temp;
+  }
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*qtblptr)->sent_table = FALSE;
+}
+
+
+GLOBAL(void)
+jpeg_set_linear_quality (j_compress_ptr cinfo, int scale_factor,
+			 int force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and a straight percentage-scaling quality scale.  In most cases it's better
+ * to use jpeg_set_quality (below); this entry point is provided for
+ * applications that insist on a linear percentage scaling.
+ */
+{
+  /* These are the sample quantization tables given in JPEG spec section K.1.
+   * The spec says that the values given produce "good" quality, and
+   * when divided by 2, "very good" quality.
+   */
+  static const unsigned int std_luminance_quant_tbl[DCTSIZE2] = {
+    16,  11,  10,  16,  24,  40,  51,  61,
+    12,  12,  14,  19,  26,  58,  60,  55,
+    14,  13,  16,  24,  40,  57,  69,  56,
+    14,  17,  22,  29,  51,  87,  80,  62,
+    18,  22,  37,  56,  68, 109, 103,  77,
+    24,  35,  55,  64,  81, 104, 113,  92,
+    49,  64,  78,  87, 103, 121, 120, 101,
+    72,  92,  95,  98, 112, 100, 103,  99
+  };
+  static const unsigned int std_chrominance_quant_tbl[DCTSIZE2] = {
+    17,  18,  24,  47,  99,  99,  99,  99,
+    18,  21,  26,  66,  99,  99,  99,  99,
+    24,  26,  56,  99,  99,  99,  99,  99,
+    47,  66,  99,  99,  99,  99,  99,  99,
+    99,  99,  99,  99,  99,  99,  99,  99,
+    99,  99,  99,  99,  99,  99,  99,  99,
+    99,  99,  99,  99,  99,  99,  99,  99,
+    99,  99,  99,  99,  99,  99,  99,  99
+  };
+
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       scale_factor, force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       scale_factor, force_baseline);
+}
+
+
+GLOBAL(int)
+jpeg_quality_scaling (int quality)
+/* Convert a user-specified quality rating to a percentage scaling factor
+ * for an underlying quantization table, using our recommended scaling curve.
+ * The input 'quality' factor should be 0 (terrible) to 100 (very good).
+ */
+{
+  /* Safety limit on quality factor.  Convert 0 to 1 to avoid zero divide. */
+  if (quality <= 0) quality = 1;
+  if (quality > 100) quality = 100;
+
+  /* The basic table is used as-is (scaling 100) for a quality of 50.
+   * Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
+   * note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
+   * to make all the table entries 1 (hence, minimum quantization loss).
+   * Qualities 1..50 are converted to scaling percentage 5000/Q.
+   */
+  if (quality < 50)
+    quality = 5000 / quality;
+  else
+    quality = 200 - quality*2;
+
+  return quality;
+}
+
+
+GLOBAL(void)
+jpeg_set_quality (j_compress_ptr cinfo, int quality, int force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables.
+ * This is the standard quality-adjusting entry point for typical user
+ * interfaces; only those who want detailed control over quantization tables
+ * would use the preceding three routines directly.
+ */
+{
+  /* Convert user 0-100 rating to percentage scaling */
+  quality = jpeg_quality_scaling(quality);
+
+  /* Set up standard quality tables */
+  jpeg_set_linear_quality(cinfo, quality, force_baseline);
+}
+
+
+/*
+ * Huffman table setup routines
+ */
+
+LOCAL(void)
+add_huff_table (j_compress_ptr cinfo,
+		JHUFF_TBL **htblptr, const unsigned char *bits, const unsigned char *val)
+/* Define a Huffman table */
+{
+  int nsymbols, len;
+
+  if (*htblptr == NULL)
+    *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+
+  /* Copy the number-of-symbols-of-each-code-length counts */
+  MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+
+  /* Validate the counts.  We do this here mainly so we can copy the right
+   * number of symbols from the val[] array, without risking marching off
+   * the end of memory.  jchuff.c will do a more thorough test later.
+   */
+  nsymbols = 0;
+  for (len = 1; len <= 16; len++)
+    nsymbols += bits[len];
+  if (nsymbols < 1 || nsymbols > 256)
+    ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+  MEMCOPY((*htblptr)->huffval, val, nsymbols * SIZEOF(unsigned char));
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*htblptr)->sent_table = FALSE;
+}
+
+
+LOCAL(void)
+std_huff_tables (j_compress_ptr cinfo)
+/* Set up the standard Huffman tables (cf. JPEG standard section K.3) */
+/* IMPORTANT: these are only valid for 8-bit data precision! */
+{
+  static const unsigned char bits_dc_luminance[17] =
+    { /* 0-base */ 0, 0, 1, 5, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0 };
+  static const unsigned char val_dc_luminance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const unsigned char bits_dc_chrominance[17] =
+    { /* 0-base */ 0, 0, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
+  static const unsigned char val_dc_chrominance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const unsigned char bits_ac_luminance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 3, 3, 2, 4, 3, 5, 5, 4, 4, 0, 0, 1, 0x7d };
+  static const unsigned char val_ac_luminance[] =
+    { 0x01, 0x02, 0x03, 0x00, 0x04, 0x11, 0x05, 0x12,
+      0x21, 0x31, 0x41, 0x06, 0x13, 0x51, 0x61, 0x07,
+      0x22, 0x71, 0x14, 0x32, 0x81, 0x91, 0xa1, 0x08,
+      0x23, 0x42, 0xb1, 0xc1, 0x15, 0x52, 0xd1, 0xf0,
+      0x24, 0x33, 0x62, 0x72, 0x82, 0x09, 0x0a, 0x16,
+      0x17, 0x18, 0x19, 0x1a, 0x25, 0x26, 0x27, 0x28,
+      0x29, 0x2a, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39,
+      0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x49,
+      0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59,
+      0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69,
+      0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79,
+      0x7a, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89,
+      0x8a, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+      0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7,
+      0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6,
+      0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3, 0xc4, 0xc5,
+      0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2, 0xd3, 0xd4,
+      0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xe1, 0xe2,
+      0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea,
+      0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  static const unsigned char bits_ac_chrominance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 2, 4, 4, 3, 4, 7, 5, 4, 4, 0, 1, 2, 0x77 };
+  static const unsigned char val_ac_chrominance[] =
+    { 0x00, 0x01, 0x02, 0x03, 0x11, 0x04, 0x05, 0x21,
+      0x31, 0x06, 0x12, 0x41, 0x51, 0x07, 0x61, 0x71,
+      0x13, 0x22, 0x32, 0x81, 0x08, 0x14, 0x42, 0x91,
+      0xa1, 0xb1, 0xc1, 0x09, 0x23, 0x33, 0x52, 0xf0,
+      0x15, 0x62, 0x72, 0xd1, 0x0a, 0x16, 0x24, 0x34,
+      0xe1, 0x25, 0xf1, 0x17, 0x18, 0x19, 0x1a, 0x26,
+      0x27, 0x28, 0x29, 0x2a, 0x35, 0x36, 0x37, 0x38,
+      0x39, 0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48,
+      0x49, 0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58,
+      0x59, 0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+      0x69, 0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78,
+      0x79, 0x7a, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+      0x88, 0x89, 0x8a, 0x92, 0x93, 0x94, 0x95, 0x96,
+      0x97, 0x98, 0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5,
+      0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4,
+      0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3,
+      0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2,
+      0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda,
+      0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9,
+      0xea, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[0],
+		 bits_dc_luminance, val_dc_luminance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[0],
+		 bits_ac_luminance, val_ac_luminance);
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[1],
+		 bits_dc_chrominance, val_dc_chrominance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[1],
+		 bits_ac_chrominance, val_ac_chrominance);
+}
+
+
+/*
+ * Default parameter setup for compression.
+ *
+ * Applications that don't choose to use this routine must do their
+ * own setup of all these parameters.  Alternately, you can call this
+ * to establish defaults and then alter parameters selectively.  This
+ * is the recommended approach since, if we add any new parameters,
+ * your code will still work (they'll be set to reasonable defaults).
+ */
+
+GLOBAL(void)
+jpeg_set_defaults (j_compress_ptr cinfo)
+{
+  int i;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Allocate comp_info array large enough for maximum component count.
+   * Array is made permanent in case application wants to compress
+   * multiple images at same param settings.
+   */
+  if (cinfo->comp_info == NULL)
+    cinfo->comp_info = (jpeg_component_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  MAX_COMPONENTS * SIZEOF(jpeg_component_info));
+
+  /* Initialize everything not dependent on the color space */
+
+  cinfo->data_precision = BITS_IN_JSAMPLE;
+  /* Set up two quantization tables using default quality of 75 */
+  jpeg_set_quality(cinfo, 75, TRUE);
+  /* Set up two Huffman tables */
+  std_huff_tables(cinfo);
+
+  /* Initialize default arithmetic coding conditioning */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+
+  /* Default is no multiple-scan output */
+  cinfo->scan_info = NULL;
+  cinfo->num_scans = 0;
+
+  /* Expect normal source image, not raw downsampled data */
+  cinfo->raw_data_in = FALSE;
+
+  /* Use Huffman coding, not arithmetic coding, by default */
+  cinfo->arith_code = FALSE;
+
+  /* By default, don't do extra passes to optimize entropy coding */
+  cinfo->optimize_coding = FALSE;
+  /* The standard Huffman tables are only valid for 8-bit data precision.
+   * If the precision is higher, force optimization on so that usable
+   * tables will be computed.  This test can be removed if default tables
+   * are supplied that are valid for the desired precision.
+   */
+  if (cinfo->data_precision > 8)
+    cinfo->optimize_coding = TRUE;
+
+  /* By default, use the simpler non-cosited sampling alignment */
+  cinfo->CCIR601_sampling = FALSE;
+
+  /* No input smoothing */
+  cinfo->smoothing_factor = 0;
+
+  /* DCT algorithm preference */
+  cinfo->dct_method = JDCT_DEFAULT;
+
+  /* No restart markers */
+  cinfo->restart_interval = 0;
+  cinfo->restart_in_rows = 0;
+
+  /* Fill in default JFIF marker parameters.  Note that whether the marker
+   * will actually be written is determined by jpeg_set_colorspace.
+   *
+   * By default, the library emits JFIF version code 1.01.
+   * An application that wants to emit JFIF 1.02 extension markers should set
+   * JFIF_minor_version to 2.  We could probably get away with just defaulting
+   * to 1.02, but there may still be some decoders in use that will complain
+   * about that; saying 1.01 should minimize compatibility problems.
+   */
+  cinfo->JFIF_major_version = 1; /* Default JFIF version = 1.01 */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;	/* Pixel size is unknown by default */
+  cinfo->X_density = 1;		/* Pixel aspect ratio is square by default */
+  cinfo->Y_density = 1;
+
+  /* Choose JPEG colorspace based on input space, set defaults accordingly */
+
+  jpeg_default_colorspace(cinfo);
+}
+
+
+/*
+ * Select an appropriate JPEG colorspace for in_color_space.
+ */
+
+GLOBAL(void)
+jpeg_default_colorspace (j_compress_ptr cinfo)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    jpeg_set_colorspace(cinfo, JCS_GRAYSCALE);
+    break;
+  case JCS_RGB:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_YCbCr:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_CMYK:
+    jpeg_set_colorspace(cinfo, JCS_CMYK); /* By default, no translation */
+    break;
+  case JCS_YCCK:
+    jpeg_set_colorspace(cinfo, JCS_YCCK);
+    break;
+  case JCS_UNKNOWN:
+    jpeg_set_colorspace(cinfo, JCS_UNKNOWN);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+  }
+}
+
+
+/*
+ * Set the JPEG colorspace, and choose colorspace-dependent default values.
+ */
+
+GLOBAL(void)
+jpeg_set_colorspace (j_compress_ptr cinfo, J_COLOR_SPACE colorspace)
+{
+  jpeg_component_info * compptr;
+  int ci;
+
+#define SET_COMP(index,id,hsamp,vsamp,quant,dctbl,actbl)  \
+  (compptr = &cinfo->comp_info[index], \
+   compptr->component_id = (id), \
+   compptr->h_samp_factor = (hsamp), \
+   compptr->v_samp_factor = (vsamp), \
+   compptr->quant_tbl_no = (quant), \
+   compptr->dc_tbl_no = (dctbl), \
+   compptr->ac_tbl_no = (actbl) )
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* For all colorspaces, we use Q and Huff tables 0 for luminance components,
+   * tables 1 for chrominance components.
+   */
+
+  cinfo->jpeg_color_space = colorspace;
+
+  cinfo->write_JFIF_header = FALSE; /* No marker for non-JFIF colorspaces */
+  cinfo->write_Adobe_marker = FALSE; /* write no Adobe marker by default */
+
+  switch (colorspace) {
+  case JCS_GRAYSCALE:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 1;
+    /* JFIF specifies component ID 1 */
+    SET_COMP(0, 1, 1,1, 0, 0,0);
+    break;
+  case JCS_RGB:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag RGB */
+    cinfo->num_components = 3;
+    SET_COMP(0, 0x52 /* 'R' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x47 /* 'G' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x42 /* 'B' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCbCr:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 3;
+    /* JFIF specifies component IDs 1,2,3 */
+    /* We default to 2x2 subsamples of chrominance */
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    break;
+  case JCS_CMYK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag CMYK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 0x43 /* 'C' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x4D /* 'M' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x59 /* 'Y' */, 1,1, 0, 0,0);
+    SET_COMP(3, 0x4B /* 'K' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCCK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag YCCK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    SET_COMP(3, 4, 2,2, 0, 0,0);
+    break;
+  case JCS_UNKNOWN:
+    cinfo->num_components = cinfo->input_components;
+    if (cinfo->num_components < 1 || cinfo->num_components > MAX_COMPONENTS)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPONENTS);
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      SET_COMP(ci, ci, 1,1, 0, 0,0);
+    }
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+  }
+}
+
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+LOCAL(jpeg_scan_info *)
+fill_a_scan (jpeg_scan_info * scanptr, int ci,
+	     int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for specified component */
+{
+  scanptr->comps_in_scan = 1;
+  scanptr->component_index[0] = ci;
+  scanptr->Ss = Ss;
+  scanptr->Se = Se;
+  scanptr->Ah = Ah;
+  scanptr->Al = Al;
+  scanptr++;
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_scans (jpeg_scan_info * scanptr, int ncomps,
+	    int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for each component */
+{
+  int ci;
+
+  for (ci = 0; ci < ncomps; ci++) {
+    scanptr->comps_in_scan = 1;
+    scanptr->component_index[0] = ci;
+    scanptr->Ss = Ss;
+    scanptr->Se = Se;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  }
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_dc_scans (jpeg_scan_info * scanptr, int ncomps, int Ah, int Al)
+/* Support routine: generate interleaved DC scan if possible, else N scans */
+{
+  int ci;
+
+  if (ncomps <= MAX_COMPS_IN_SCAN) {
+    /* Single interleaved DC scan */
+    scanptr->comps_in_scan = ncomps;
+    for (ci = 0; ci < ncomps; ci++)
+      scanptr->component_index[ci] = ci;
+    scanptr->Ss = scanptr->Se = 0;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  } else {
+    /* Noninterleaved DC scan for each component */
+    scanptr = fill_scans(scanptr, ncomps, 0, 0, Ah, Al);
+  }
+  return scanptr;
+}
+
+
+/*
+ * Create a recommended progressive-JPEG script.
+ * cinfo->num_components and cinfo->jpeg_color_space must be correct.
+ */
+
+GLOBAL(void)
+jpeg_simple_progression (j_compress_ptr cinfo)
+{
+  int ncomps = cinfo->num_components;
+  int nscans;
+  jpeg_scan_info * scanptr;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Figure space needed for script.  Calculation must match code below! */
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    nscans = 10;
+  } else {
+    /* All-purpose script for other color spaces. */
+    if (ncomps > MAX_COMPS_IN_SCAN)
+      nscans = 6 * ncomps;	/* 2 DC + 4 AC scans per component */
+    else
+      nscans = 2 + 4 * ncomps;	/* 2 DC scans; 4 AC scans per component */
+  }
+
+  /* Allocate space for script.
+   * We need to put it in the permanent pool in case the application performs
+   * multiple compressions without changing the settings.  To avoid a memory
+   * leak if jpeg_simple_progression is called repeatedly for the same JPEG
+   * object, we try to re-use previously allocated space, and we allocate
+   * enough space to handle YCbCr even if initially asked for grayscale.
+   */
+  if (cinfo->script_space == NULL || cinfo->script_space_size < nscans) {
+    cinfo->script_space_size = MAX(nscans, 10);
+    cinfo->script_space = (jpeg_scan_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+			cinfo->script_space_size * SIZEOF(jpeg_scan_info));
+  }
+  scanptr = cinfo->script_space;
+  cinfo->scan_info = scanptr;
+  cinfo->num_scans = nscans;
+
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    /* Initial DC scan */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    /* Initial AC scan: get some luma data out in a hurry */
+    scanptr = fill_a_scan(scanptr, 0, 1, 5, 0, 2);
+    /* Chroma data is too small to be worth expending many scans on */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 0, 1);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 0, 1);
+    /* Complete spectral selection for luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 6, 63, 0, 2);
+    /* Refine next bit of luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 2, 1);
+    /* Finish DC successive approximation */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    /* Finish AC successive approximation */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 1, 0);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 1, 0);
+    /* Luma bottom bit comes last since it's usually largest scan */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 1, 0);
+  } else {
+    /* All-purpose script for other color spaces. */
+    /* Successive approximation first pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    scanptr = fill_scans(scanptr, ncomps, 1, 5, 0, 2);
+    scanptr = fill_scans(scanptr, ncomps, 6, 63, 0, 2);
+    /* Successive approximation second pass */
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 2, 1);
+    /* Successive approximation final pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 1, 0);
+  }
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jcphuff.cpp b/ml/dlib/dlib/external/libjpeg/jcphuff.cpp
new file mode 100644
index 000000000..66a85b8c7
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcphuff.cpp
@@ -0,0 +1,833 @@
+/*
+ * jcphuff.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy encoding routines for progressive JPEG.
+ *
+ * We do not support output suspension in this module, since the library
+ * currently does not allow multiple-scan files to be written with output
+ * suspension.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jchuff.h"		/* Declarations shared with jchuff.c */
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+/* Expanded entropy encoder object for progressive Huffman encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  /* Mode flag: TRUE for optimization, FALSE for actual data output */
+  int gather_statistics;
+
+  /* Bit-level coding status.
+   * next_output_byte/free_in_buffer are local copies of cinfo->dest fields.
+   */
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  long put_buffer;		/* current bit-accumulation buffer */
+  int put_bits;			/* # of bits now in it */
+  j_compress_ptr cinfo;		/* link to cinfo (needed for dump_buffer) */
+
+  /* Coding status for DC components */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+
+  /* Coding status for AC components */
+  int ac_tbl_no;		/* the table number of the single component */
+  unsigned int EOBRUN;		/* run length of EOBs */
+  unsigned int BE;		/* # of buffered correction bits before MCU */
+  char * bit_buffer;		/* buffer for correction bits (1 per char) */
+  /* packing correction bits tightly would save some space but cost time... */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan).
+   * Since any one scan codes only DC or only AC, we only need one set
+   * of tables, not one for DC and one for AC.
+   */
+  c_derived_tbl * derived_tbls[NUM_HUFF_TBLS];
+
+  /* Statistics tables for optimization; again, one set is enough */
+  long * count_ptrs[NUM_HUFF_TBLS];
+} phuff_entropy_encoder;
+
+typedef phuff_entropy_encoder * phuff_entropy_ptr;
+
+/* MAX_CORR_BITS is the number of bits the AC refinement correction-bit
+ * buffer can hold.  Larger sizes may slightly improve compression, but
+ * 1000 is already well into the realm of overkill.
+ * The minimum safe size is 64 bits.
+ */
+
+#define MAX_CORR_BITS  1000	/* Max # of correction bits I can buffer */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than long.
+ * We assume that int right shift is unsigned if long right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	int ishift_temp;
+#define IRIGHT_SHIFT(x,shft)  \
+	((ishift_temp = (x)) < 0 ? \
+	 (ishift_temp >> (shft)) | ((~0) << (16-(shft))) : \
+	 (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+/* Forward declarations */
+METHODDEF(int) encode_mcu_DC_first JPP((j_compress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(int) encode_mcu_AC_first JPP((j_compress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(int) encode_mcu_DC_refine JPP((j_compress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(int) encode_mcu_AC_refine JPP((j_compress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(void) finish_pass_phuff JPP((j_compress_ptr cinfo));
+METHODDEF(void) finish_pass_gather_phuff JPP((j_compress_ptr cinfo));
+
+
+/*
+ * Initialize for a Huffman-compressed scan using progressive JPEG.
+ */
+
+METHODDEF(void)
+start_pass_phuff (j_compress_ptr cinfo, int gather_statistics)
+{  
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int is_DC_band;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  entropy->cinfo = cinfo;
+  entropy->gather_statistics = gather_statistics;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* We assume jcmaster.c already validated the scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_first;
+    else
+      entropy->pub.encode_mcu = encode_mcu_AC_first;
+  } else {
+    if (is_DC_band)
+      entropy->pub.encode_mcu = encode_mcu_DC_refine;
+    else {
+      entropy->pub.encode_mcu = encode_mcu_AC_refine;
+      /* AC refinement needs a correction bit buffer */
+      if (entropy->bit_buffer == NULL)
+	entropy->bit_buffer = (char *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      MAX_CORR_BITS * SIZEOF(char));
+    }
+  }
+  if (gather_statistics)
+    entropy->pub.finish_pass = finish_pass_gather_phuff;
+  else
+    entropy->pub.finish_pass = finish_pass_phuff;
+
+  /* Only DC coefficients may be interleaved, so cinfo->comps_in_scan = 1
+   * for AC coefficients.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Initialize DC predictions to 0 */
+    entropy->last_dc_val[ci] = 0;
+    /* Get table index */
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)	/* DC refinement needs no table */
+	continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      entropy->ac_tbl_no = tbl = compptr->ac_tbl_no;
+    }
+    if (gather_statistics) {
+      /* Check for invalid table index */
+      /* (make_c_derived_tbl does this in the other path) */
+      if (tbl < 0 || tbl >= NUM_HUFF_TBLS)
+        ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tbl);
+      /* Allocate and zero the statistics tables */
+      /* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+      if (entropy->count_ptrs[tbl] == NULL)
+	entropy->count_ptrs[tbl] = (long *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      257 * SIZEOF(long));
+      MEMZERO(entropy->count_ptrs[tbl], 257 * SIZEOF(long));
+    } else {
+      /* Compute derived values for Huffman table */
+      /* We may do this more than once for a table, but it's not expensive */
+      jpeg_make_c_derived_tbl(cinfo, is_DC_band, tbl,
+			      & entropy->derived_tbls[tbl]);
+    }
+  }
+
+  /* Initialize AC stuff */
+  entropy->EOBRUN = 0;
+  entropy->BE = 0;
+
+  /* Initialize bit buffer to empty */
+  entropy->put_buffer = 0;
+  entropy->put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/* Outputting bytes to the file.
+ * NB: these must be called only when actually outputting,
+ * that is, entropy->gather_statistics == FALSE.
+ */
+
+/* Emit a byte */
+#define emit_byte(entropy,val)  \
+	{ *(entropy)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(entropy)->free_in_buffer == 0)  \
+	    dump_buffer(entropy); }
+
+
+LOCAL(void)
+dump_buffer (phuff_entropy_ptr entropy)
+/* Empty the output buffer; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr * dest = entropy->cinfo->dest;
+
+  if (! (*dest->empty_output_buffer) (entropy->cinfo))
+    ERREXIT(entropy->cinfo, JERR_CANT_SUSPEND);
+  /* After a successful buffer dump, must reset buffer pointers */
+  entropy->next_output_byte = dest->next_output_byte;
+  entropy->free_in_buffer = dest->free_in_buffer;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+inline
+LOCAL(void)
+emit_bits (phuff_entropy_ptr entropy, unsigned int code, int size)
+/* Emit some bits, unless we are in gather mode */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  long put_buffer = (long) code;
+  int put_bits = entropy->put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+  if (entropy->gather_statistics)
+    return;			/* do nothing if we're only getting stats */
+
+  put_buffer &= (((long) 1)<<size) - 1; /* mask off any extra bits in code */
+  
+  put_bits += size;		/* new number of bits in buffer */
+  
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  put_buffer |= entropy->put_buffer; /* and merge with old buffer contents */
+
+  while (put_bits >= 8) {
+    int c = (int) ((put_buffer >> 16) & 0xFF);
+    
+    emit_byte(entropy, c);
+    if (c == 0xFF) {		/* need to stuff a zero byte? */
+      emit_byte(entropy, 0);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  entropy->put_buffer = put_buffer; /* update variables */
+  entropy->put_bits = put_bits;
+}
+
+
+LOCAL(void)
+flush_bits (phuff_entropy_ptr entropy)
+{
+  emit_bits(entropy, 0x7F, 7); /* fill any partial byte with ones */
+  entropy->put_buffer = 0;     /* and reset bit-buffer to empty */
+  entropy->put_bits = 0;
+}
+
+
+/*
+ * Emit (or just count) a Huffman symbol.
+ */
+
+inline
+LOCAL(void)
+emit_symbol (phuff_entropy_ptr entropy, int tbl_no, int symbol)
+{
+  if (entropy->gather_statistics)
+    entropy->count_ptrs[tbl_no][symbol]++;
+  else {
+    c_derived_tbl * tbl = entropy->derived_tbls[tbl_no];
+    emit_bits(entropy, tbl->ehufco[symbol], tbl->ehufsi[symbol]);
+  }
+}
+
+
+/*
+ * Emit bits from a correction bit buffer.
+ */
+
+LOCAL(void)
+emit_buffered_bits (phuff_entropy_ptr entropy, char * bufstart,
+		    unsigned int nbits)
+{
+  if (entropy->gather_statistics)
+    return;			/* no real work */
+
+  while (nbits > 0) {
+    emit_bits(entropy, (unsigned int) (*bufstart), 1);
+    bufstart++;
+    nbits--;
+  }
+}
+
+
+/*
+ * Emit any pending EOBRUN symbol.
+ */
+
+LOCAL(void)
+emit_eobrun (phuff_entropy_ptr entropy)
+{
+  int temp, nbits;
+
+  if (entropy->EOBRUN > 0) {	/* if there is any pending EOBRUN */
+    temp = entropy->EOBRUN;
+    nbits = 0;
+    while ((temp >>= 1))
+      nbits++;
+    /* safety check: shouldn't happen given limited correction-bit buffer */
+    if (nbits > 14)
+      ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+    emit_symbol(entropy, entropy->ac_tbl_no, nbits << 4);
+    if (nbits)
+      emit_bits(entropy, entropy->EOBRUN, nbits);
+
+    entropy->EOBRUN = 0;
+
+    /* Emit any buffered correction bits */
+    emit_buffered_bits(entropy, entropy->bit_buffer, entropy->BE);
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart (phuff_entropy_ptr entropy, int restart_num)
+{
+  int ci;
+
+  emit_eobrun(entropy);
+
+  if (! entropy->gather_statistics) {
+    flush_bits(entropy);
+    emit_byte(entropy, 0xFF);
+    emit_byte(entropy, JPEG_RST0 + restart_num);
+  }
+
+  if (entropy->cinfo->Ss == 0) {
+    /* Re-initialize DC predictions to 0 */
+    for (ci = 0; ci < entropy->cinfo->comps_in_scan; ci++)
+      entropy->last_dc_val[ci] = 0;
+  } else {
+    /* Re-initialize all AC-related fields to 0 */
+    entropy->EOBRUN = 0;
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(int)
+encode_mcu_DC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int temp, temp2;
+  int nbits;
+  int blkn, ci;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  jpeg_component_info * compptr;
+  ISHIFT_TEMPS
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    temp2 = IRIGHT_SHIFT((int) ((*block)[0]), Al);
+
+    /* DC differences are figured on the point-transformed values. */
+    temp = temp2 - entropy->last_dc_val[ci];
+    entropy->last_dc_val[ci] = temp2;
+
+    /* Encode the DC coefficient difference per section G.1.2.1 */
+    temp2 = temp;
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      /* For a negative input, want temp2 = bitwise complement of abs(input) */
+      /* This code assumes we are on a two's complement machine */
+      temp2--;
+    }
+    
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 0;
+    while (temp) {
+      nbits++;
+      temp >>= 1;
+    }
+    /* Check for out-of-range coefficient values.
+     * Since we're encoding a difference, the range limit is twice as much.
+     */
+    if (nbits > MAX_COEF_BITS+1)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+    
+    /* Count/emit the Huffman-coded symbol for the number of bits */
+    emit_symbol(entropy, compptr->dc_tbl_no, nbits);
+    
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    if (nbits)			/* emit_bits rejects calls with size 0 */
+      emit_bits(entropy, (unsigned int) temp2, nbits);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(int)
+encode_mcu_AC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int temp, temp2;
+  int nbits;
+  int r, k;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* Encode the AC coefficients per section G.1.2.2, fig. G.3 */
+  
+  r = 0;			/* r = run length of zeros */
+   
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = (*block)[jpeg_natural_order[k]]) == 0) {
+      r++;
+      continue;
+    }
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value; so the code is
+     * interwoven with finding the abs value (temp) and output bits (temp2).
+     */
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      temp >>= Al;		/* apply the point transform */
+      /* For a negative coef, want temp2 = bitwise complement of abs(coef) */
+      temp2 = ~temp;
+    } else {
+      temp >>= Al;		/* apply the point transform */
+      temp2 = temp;
+    }
+    /* Watch out for case that nonzero coef is zero after point transform */
+    if (temp == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any pending EOBRUN */
+    if (entropy->EOBRUN > 0)
+      emit_eobrun(entropy);
+    /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+    while (r > 15) {
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+    }
+
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 1;			/* there must be at least one 1 bit */
+    while ((temp >>= 1))
+      nbits++;
+    /* Check for out-of-range coefficient values */
+    if (nbits > MAX_COEF_BITS)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + nbits);
+
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    emit_bits(entropy, (unsigned int) temp2, nbits);
+
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0) {			/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    if (entropy->EOBRUN == 0x7FFF)
+      emit_eobrun(entropy);	/* force it out to avoid overflow */
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(int)
+encode_mcu_DC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int temp;
+  int blkn;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    temp = (*block)[0];
+    emit_bits(entropy, (unsigned int) (temp >> Al), 1);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(int)
+encode_mcu_AC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int temp;
+  int r, k;
+  int EOB;
+  char *BR_buffer;
+  unsigned int BR;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  int absvalues[DCTSIZE2];
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* It is convenient to make a pre-pass to determine the transformed
+   * coefficients' absolute values and the EOB position.
+   */
+  EOB = 0;
+  for (k = cinfo->Ss; k <= Se; k++) {
+    temp = (*block)[jpeg_natural_order[k]];
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if (temp < 0)
+      temp = -temp;		/* temp is abs value of input */
+    temp >>= Al;		/* apply the point transform */
+    absvalues[k] = temp;	/* save abs value for main pass */
+    if (temp == 1)
+      EOB = k;			/* EOB = index of last newly-nonzero coef */
+  }
+
+  /* Encode the AC coefficients per section G.1.2.3, fig. G.7 */
+  
+  r = 0;			/* r = run length of zeros */
+  BR = 0;			/* BR = count of buffered bits added now */
+  BR_buffer = entropy->bit_buffer + entropy->BE; /* Append bits to buffer */
+
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = absvalues[k]) == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any required ZRLs, but not if they can be folded into EOB */
+    while (r > 15 && k <= EOB) {
+      /* emit any pending EOBRUN and the BE correction bits */
+      emit_eobrun(entropy);
+      /* Emit ZRL */
+      emit_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+      /* Emit buffered correction bits that must be associated with ZRL */
+      emit_buffered_bits(entropy, BR_buffer, BR);
+      BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+      BR = 0;
+    }
+
+    /* If the coef was previously nonzero, it only needs a correction bit.
+     * NOTE: a straight translation of the spec's figure G.7 would suggest
+     * that we also need to test r > 15.  But if r > 15, we can only get here
+     * if k > EOB, which implies that this coefficient is not 1.
+     */
+    if (temp > 1) {
+      /* The correction bit is the next bit of the absolute value. */
+      BR_buffer[BR++] = (char) (temp & 1);
+      continue;
+    }
+
+    /* Emit any pending EOBRUN and the BE correction bits */
+    emit_eobrun(entropy);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_symbol(entropy, entropy->ac_tbl_no, (r << 4) + 1);
+
+    /* Emit output bit for newly-nonzero coef */
+    temp = ((*block)[jpeg_natural_order[k]] < 0) ? 0 : 1;
+    emit_bits(entropy, (unsigned int) temp, 1);
+
+    /* Emit buffered correction bits that must be associated with this code */
+    emit_buffered_bits(entropy, BR_buffer, BR);
+    BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+    BR = 0;
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0 || BR > 0) {	/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    entropy->BE += BR;		/* concat my correction bits to older ones */
+    /* We force out the EOB if we risk either:
+     * 1. overflow of the EOB counter;
+     * 2. overflow of the correction bit buffer during the next MCU.
+     */
+    if (entropy->EOBRUN == 0x7FFF || entropy->BE > (MAX_CORR_BITS-DCTSIZE2+1))
+      emit_eobrun(entropy);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed progressive scan.
+ */
+
+METHODDEF(void)
+finish_pass_phuff (j_compress_ptr cinfo)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Flush out any buffered data */
+  emit_eobrun(entropy);
+  flush_bits(entropy);
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather_phuff (j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int is_DC_band;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+  JHUFF_TBL **htblptr;
+  int did[NUM_HUFF_TBLS];
+
+  /* Flush out buffered data (all we care about is counting the EOB symbol) */
+  emit_eobrun(entropy);
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  MEMZERO(did, SIZEOF(did));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (is_DC_band) {
+      if (cinfo->Ah != 0)	/* DC refinement needs no table */
+	continue;
+      tbl = compptr->dc_tbl_no;
+    } else {
+      tbl = compptr->ac_tbl_no;
+    }
+    if (! did[tbl]) {
+      if (is_DC_band)
+        htblptr = & cinfo->dc_huff_tbl_ptrs[tbl];
+      else
+        htblptr = & cinfo->ac_huff_tbl_ptrs[tbl];
+      if (*htblptr == NULL)
+        *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+      jpeg_gen_optimal_table(cinfo, *htblptr, entropy->count_ptrs[tbl]);
+      did[tbl] = TRUE;
+    }
+  }
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_encoder (j_compress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(phuff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass_phuff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+    entropy->count_ptrs[i] = NULL;
+  }
+  entropy->bit_buffer = NULL;	/* needed only in AC refinement scan */
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jcprepct.cpp b/ml/dlib/dlib/external/libjpeg/jcprepct.cpp
new file mode 100644
index 000000000..d1532c273
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcprepct.cpp
@@ -0,0 +1,354 @@
+/*
+ * jcprepct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the compression preprocessing controller.
+ * This controller manages the color conversion, downsampling,
+ * and edge expansion steps.
+ *
+ * Most of the complexity here is associated with buffering input rows
+ * as required by the downsampler.  See the comments at the head of
+ * jcsample.c for the downsampler's needs.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* At present, jcsample.c can request context rows only for smoothing.
+ * In the future, we might also need context rows for CCIR601 sampling
+ * or other more-complex downsampling procedures.  The code to support
+ * context rows should be compiled only if needed.
+ */
+#ifdef INPUT_SMOOTHING_SUPPORTED
+#define CONTEXT_ROWS_SUPPORTED
+#endif
+
+
+/*
+ * For the simple (no-context-row) case, we just need to buffer one
+ * row group's worth of pixels for the downsampling step.  At the bottom of
+ * the image, we pad to a full row group by replicating the last pixel row.
+ * The downsampler's last output row is then replicated if needed to pad
+ * out to a full iMCU row.
+ *
+ * When providing context rows, we must buffer three row groups' worth of
+ * pixels.  Three row groups are physically allocated, but the row pointer
+ * arrays are made five row groups high, with the extra pointers above and
+ * below "wrapping around" to point to the last and first real row groups.
+ * This allows the downsampler to access the proper context rows.
+ * At the top and bottom of the image, we create dummy context rows by
+ * copying the first or last real pixel row.  This copying could be avoided
+ * by pointer hacking as is done in jdmainct.c, but it doesn't seem worth the
+ * trouble on the compression side.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_prep_controller pub; /* public fields */
+
+  /* Downsampling input buffer.  This buffer holds color-converted data
+   * until we have enough to do a downsample step.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  JDIMENSION rows_to_go;	/* counts rows remaining in source image */
+  int next_buf_row;		/* index of next row to store in color_buf */
+
+#ifdef CONTEXT_ROWS_SUPPORTED	/* only needed for context case */
+  int this_row_group;		/* starting row index of group to process */
+  int next_buf_stop;		/* downsample when we reach this index */
+#endif
+} my_prep_controller;
+
+typedef my_prep_controller * my_prep_ptr;
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_prep (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+
+  if (pass_mode != JBUF_PASS_THRU)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Initialize total-height counter for detecting bottom of image */
+  prep->rows_to_go = cinfo->image_height;
+  /* Mark the conversion buffer empty */
+  prep->next_buf_row = 0;
+#ifdef CONTEXT_ROWS_SUPPORTED
+  /* Preset additional state variables for context mode.
+   * These aren't used in non-context mode, so we needn't test which mode.
+   */
+  prep->this_row_group = 0;
+  /* Set next_buf_stop to stop after two row groups have been read in. */
+  prep->next_buf_stop = 2 * cinfo->max_v_samp_factor;
+#endif
+}
+
+
+/*
+ * Expand an image vertically from height input_rows to height output_rows,
+ * by duplicating the bottom row.
+ */
+
+LOCAL(void)
+expand_bottom_edge (JSAMPARRAY image_data, JDIMENSION num_cols,
+		    int input_rows, int output_rows)
+{
+  int row;
+
+  for (row = input_rows; row < output_rows; row++) {
+    jcopy_sample_rows(image_data, input_rows-1, image_data, row,
+		      1, num_cols);
+  }
+}
+
+
+/*
+ * Process some data in the simple no-context case.
+ *
+ * Preprocessor output data is counted in "row groups".  A row group
+ * is defined to be v_samp_factor sample rows of each component.
+ * Downsampling will produce this much data from each max_v_samp_factor
+ * input rows.
+ */
+
+METHODDEF(void)
+pre_process_data (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		  JDIMENSION in_rows_avail,
+		  JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		  JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  JDIMENSION inrows;
+  jpeg_component_info * compptr;
+
+  while (*in_row_ctr < in_rows_avail &&
+	 *out_row_group_ctr < out_row_groups_avail) {
+    /* Do color conversion to fill the conversion buffer. */
+    inrows = in_rows_avail - *in_row_ctr;
+    numrows = cinfo->max_v_samp_factor - prep->next_buf_row;
+    numrows = (int) MIN((JDIMENSION) numrows, inrows);
+    (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+				       prep->color_buf,
+				       (JDIMENSION) prep->next_buf_row,
+				       numrows);
+    *in_row_ctr += numrows;
+    prep->next_buf_row += numrows;
+    prep->rows_to_go -= numrows;
+    /* If at bottom of image, pad to fill the conversion buffer. */
+    if (prep->rows_to_go == 0 &&
+	prep->next_buf_row < cinfo->max_v_samp_factor) {
+      for (ci = 0; ci < cinfo->num_components; ci++) {
+	expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			   prep->next_buf_row, cinfo->max_v_samp_factor);
+      }
+      prep->next_buf_row = cinfo->max_v_samp_factor;
+    }
+    /* If we've filled the conversion buffer, empty it. */
+    if (prep->next_buf_row == cinfo->max_v_samp_factor) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf, (JDIMENSION) 0,
+					output_buf, *out_row_group_ctr);
+      prep->next_buf_row = 0;
+      (*out_row_group_ctr)++;
+    }
+    /* If at bottom of image, pad the output to a full iMCU height.
+     * Note we assume the caller is providing a one-iMCU-height output buffer!
+     */
+    if (prep->rows_to_go == 0 &&
+	*out_row_group_ctr < out_row_groups_avail) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	expand_bottom_edge(output_buf[ci],
+			   compptr->width_in_blocks * DCTSIZE,
+			   (int) (*out_row_group_ctr * compptr->v_samp_factor),
+			   (int) (out_row_groups_avail * compptr->v_samp_factor));
+      }
+      *out_row_group_ctr = out_row_groups_avail;
+      break;			/* can exit outer loop without test */
+    }
+  }
+}
+
+
+#ifdef CONTEXT_ROWS_SUPPORTED
+
+/*
+ * Process some data in the context case.
+ */
+
+METHODDEF(void)
+pre_process_context (j_compress_ptr cinfo,
+		     JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		     JDIMENSION in_rows_avail,
+		     JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		     JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  int buf_height = cinfo->max_v_samp_factor * 3;
+  JDIMENSION inrows;
+
+  while (*out_row_group_ctr < out_row_groups_avail) {
+    if (*in_row_ctr < in_rows_avail) {
+      /* Do color conversion to fill the conversion buffer. */
+      inrows = in_rows_avail - *in_row_ctr;
+      numrows = prep->next_buf_stop - prep->next_buf_row;
+      numrows = (int) MIN((JDIMENSION) numrows, inrows);
+      (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+					 prep->color_buf,
+					 (JDIMENSION) prep->next_buf_row,
+					 numrows);
+      /* Pad at top of image, if first time through */
+      if (prep->rows_to_go == cinfo->image_height) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  int row;
+	  for (row = 1; row <= cinfo->max_v_samp_factor; row++) {
+	    jcopy_sample_rows(prep->color_buf[ci], 0,
+			      prep->color_buf[ci], -row,
+			      1, cinfo->image_width);
+	  }
+	}
+      }
+      *in_row_ctr += numrows;
+      prep->next_buf_row += numrows;
+      prep->rows_to_go -= numrows;
+    } else {
+      /* Return for more data, unless we are at the bottom of the image. */
+      if (prep->rows_to_go != 0)
+	break;
+      /* When at bottom of image, pad to fill the conversion buffer. */
+      if (prep->next_buf_row < prep->next_buf_stop) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			     prep->next_buf_row, prep->next_buf_stop);
+	}
+	prep->next_buf_row = prep->next_buf_stop;
+      }
+    }
+    /* If we've gotten enough data, downsample a row group. */
+    if (prep->next_buf_row == prep->next_buf_stop) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf,
+					(JDIMENSION) prep->this_row_group,
+					output_buf, *out_row_group_ctr);
+      (*out_row_group_ctr)++;
+      /* Advance pointers with wraparound as necessary. */
+      prep->this_row_group += cinfo->max_v_samp_factor;
+      if (prep->this_row_group >= buf_height)
+	prep->this_row_group = 0;
+      if (prep->next_buf_row >= buf_height)
+	prep->next_buf_row = 0;
+      prep->next_buf_stop = prep->next_buf_row + cinfo->max_v_samp_factor;
+    }
+  }
+}
+
+
+/*
+ * Create the wrapped-around downsampling input buffer needed for context mode.
+ */
+
+LOCAL(void)
+create_context_buffer (j_compress_ptr cinfo)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int rgroup_height = cinfo->max_v_samp_factor;
+  int ci, i;
+  jpeg_component_info * compptr;
+  JSAMPARRAY true_buffer, fake_buffer;
+
+  /* Grab enough space for fake row pointers for all the components;
+   * we need five row groups' worth of pointers for each component.
+   */
+  fake_buffer = (JSAMPARRAY)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(cinfo->num_components * 5 * rgroup_height) *
+				SIZEOF(JSAMPROW));
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate the actual buffer space (3 row groups) for this component.
+     * We make the buffer wide enough to allow the downsampler to edge-expand
+     * horizontally within the buffer, if it so chooses.
+     */
+    true_buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) (((long) compptr->width_in_blocks * DCTSIZE *
+		      cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+       (JDIMENSION) (3 * rgroup_height));
+    /* Copy true buffer row pointers into the middle of the fake row array */
+    MEMCOPY(fake_buffer + rgroup_height, true_buffer,
+	    3 * rgroup_height * SIZEOF(JSAMPROW));
+    /* Fill in the above and below wraparound pointers */
+    for (i = 0; i < rgroup_height; i++) {
+      fake_buffer[i] = true_buffer[2 * rgroup_height + i];
+      fake_buffer[4 * rgroup_height + i] = true_buffer[i];
+    }
+    prep->color_buf[ci] = fake_buffer + rgroup_height;
+    fake_buffer += 5 * rgroup_height; /* point to space for next component */
+  }
+}
+
+#endif /* CONTEXT_ROWS_SUPPORTED */
+
+
+/*
+ * Initialize preprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_c_prep_controller (j_compress_ptr cinfo, int need_full_buffer)
+{
+  my_prep_ptr prep;
+  int ci;
+  jpeg_component_info * compptr;
+
+  if (need_full_buffer)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  prep = (my_prep_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_prep_controller));
+  cinfo->prep = (struct jpeg_c_prep_controller *) prep;
+  prep->pub.start_pass = start_pass_prep;
+
+  /* Allocate the color conversion buffer.
+   * We make the buffer wide enough to allow the downsampler to edge-expand
+   * horizontally within the buffer, if it so chooses.
+   */
+  if (cinfo->downsample->need_context_rows) {
+    /* Set up to provide context rows */
+#ifdef CONTEXT_ROWS_SUPPORTED
+    prep->pub.pre_process_data = pre_process_context;
+    create_context_buffer(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* No context, just make it tall enough for one row group */
+    prep->pub.pre_process_data = pre_process_data;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      prep->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) (((long) compptr->width_in_blocks * DCTSIZE *
+			cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jcsample.cpp b/ml/dlib/dlib/external/libjpeg/jcsample.cpp
new file mode 100644
index 000000000..b73270120
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jcsample.cpp
@@ -0,0 +1,519 @@
+/*
+ * jcsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains downsampling routines.
+ *
+ * Downsampling input data is counted in "row groups".  A row group
+ * is defined to be max_v_samp_factor pixel rows of each component,
+ * from which the downsampler produces v_samp_factor sample rows.
+ * A single row group is processed in each call to the downsampler module.
+ *
+ * The downsampler is responsible for edge-expansion of its output data
+ * to fill an integral number of DCT blocks horizontally.  The source buffer
+ * may be modified if it is helpful for this purpose (the source buffer is
+ * allocated wide enough to correspond to the desired output width).
+ * The caller (the prep controller) is responsible for vertical padding.
+ *
+ * The downsampler may request "context rows" by setting need_context_rows
+ * during startup.  In this case, the input arrays will contain at least
+ * one row group's worth of pixels above and below the passed-in data;
+ * the caller will create dummy rows at image top and bottom by replicating
+ * the first or last real pixel row.
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ *
+ * The downsampling algorithm used here is a simple average of the source
+ * pixels covered by the output pixel.  The hi-falutin sampling literature
+ * refers to this as a "box filter".  In general the characteristics of a box
+ * filter are not very good, but for the specific cases we normally use (1:1
+ * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
+ * nearly so bad.  If you intend to use other sampling ratios, you'd be well
+ * advised to improve this code.
+ *
+ * A simple input-smoothing capability is provided.  This is mainly intended
+ * for cleaning up color-dithered GIF input files (if you find it inadequate,
+ * we suggest using an external filtering program such as pnmconvol).  When
+ * enabled, each input pixel P is replaced by a weighted sum of itself and its
+ * eight neighbors.  P's weight is 1-8*SF and each neighbor's weight is SF,
+ * where SF = (smoothing_factor / 1024).
+ * Currently, smoothing is only supported for 2h2v sampling factors.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Pointer to routine to downsample a single component */
+typedef JMETHOD(void, downsample1_ptr,
+		(j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_downsampler pub;	/* public fields */
+
+  /* Downsampling method pointers, one per component */
+  downsample1_ptr methods[MAX_COMPONENTS];
+} my_downsampler;
+
+typedef my_downsampler * my_downsample_ptr;
+
+
+/*
+ * Initialize for a downsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_downsample (j_compress_ptr )//cinfo)
+{
+  /* no work for now */
+}
+
+
+/*
+ * Expand a component horizontally from width input_cols to width output_cols,
+ * by duplicating the rightmost samples.
+ */
+
+LOCAL(void)
+expand_right_edge (JSAMPARRAY image_data, int num_rows,
+		   JDIMENSION input_cols, JDIMENSION output_cols)
+{
+  JSAMPROW ptr;
+  JSAMPLE pixval;
+  int count;
+  int row;
+  int numcols = (int) (output_cols - input_cols);
+
+  if (numcols > 0) {
+    for (row = 0; row < num_rows; row++) {
+      ptr = image_data[row] + input_cols;
+      pixval = ptr[-1];		/* don't need GETJSAMPLE() here */
+      for (count = numcols; count > 0; count--)
+	*ptr++ = pixval;
+    }
+  }
+}
+
+
+/*
+ * Do downsampling for a whole row group (all components).
+ *
+ * In this version we simply downsample each component independently.
+ */
+
+METHODDEF(void)
+sep_downsample (j_compress_ptr cinfo,
+		JSAMPIMAGE input_buf, JDIMENSION in_row_index,
+		JSAMPIMAGE output_buf, JDIMENSION out_row_group_index)
+{
+  my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JSAMPARRAY in_ptr, out_ptr;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    in_ptr = input_buf[ci] + in_row_index;
+    out_ptr = output_buf[ci] + (out_row_group_index * compptr->v_samp_factor);
+    (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * One row group is processed per call.
+ * This version handles arbitrary integral sampling ratios, without smoothing.
+ * Note that this version is not actually used for customary sampling ratios.
+ */
+
+METHODDEF(void)
+int_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
+  JDIMENSION outcol, outcol_h;	/* outcol_h == outcol*h_expand */
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr, outptr;
+  long outvalue;
+
+  h_expand = cinfo->max_h_samp_factor / compptr->h_samp_factor;
+  v_expand = cinfo->max_v_samp_factor / compptr->v_samp_factor;
+  numpix = h_expand * v_expand;
+  numpix2 = numpix/2;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * h_expand);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    for (outcol = 0, outcol_h = 0; outcol < output_cols;
+	 outcol++, outcol_h += h_expand) {
+      outvalue = 0;
+      for (v = 0; v < v_expand; v++) {
+	inptr = input_data[inrow+v] + outcol_h;
+	for (h = 0; h < h_expand; h++) {
+	  outvalue += (long) GETJSAMPLE(*inptr++);
+	}
+      }
+      *outptr++ = (JSAMPLE) ((outvalue + numpix2) / numpix);
+    }
+    inrow += v_expand;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+fullsize_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  /* Copy the data */
+  jcopy_sample_rows(input_data, 0, output_data, 0,
+		    cinfo->max_v_samp_factor, cinfo->image_width);
+  /* Edge-expand */
+  expand_right_edge(output_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, compptr->width_in_blocks * DCTSIZE);
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr, outptr;
+  int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    bias = 0;			/* bias = 0,1,0,1,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1])
+			      + bias) >> 1);
+      bias ^= 1;		/* 0=>1, 1=>0 */
+      inptr += 2;
+    }
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr0, inptr1, outptr;
+  int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    bias = 1;			/* bias = 1,2,1,2,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+			      GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1])
+			      + bias) >> 2);
+      bias ^= 3;		/* 1=>2, 2=>1 */
+      inptr0 += 2; inptr1 += 2;
+    }
+    inrow += 2;
+  }
+}
+
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+			JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
+  long membersum, neighsum, memberscale, neighscale;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols * 2);
+
+  /* We don't bother to form the individual "smoothed" input pixel values;
+   * we can directly compute the output which is the average of the four
+   * smoothed values.  Each of the four member pixels contributes a fraction
+   * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
+   * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
+   * output.  The four corner-adjacent neighbor pixels contribute a fraction
+   * SF to just one smoothed pixel, or SF/4 to the final output; while the
+   * eight edge-adjacent neighbors contribute SF to each of two smoothed
+   * pixels, or SF/2 overall.  In order to use integer arithmetic, these
+   * factors are scaled by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
+  neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
+
+  inrow = 0;
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    above_ptr = input_data[inrow-1];
+    below_ptr = input_data[inrow+2];
+
+    /* Special case for first column: pretend column -1 is same as column 0 */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
+	       GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
+		GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      /* sum of pixels directly mapped to this output element */
+      membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		  GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+      /* sum of edge-neighbor pixels */
+      neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+		 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+		 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
+		 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
+      /* The edge-neighbors count twice as much as corner-neighbors */
+      neighsum += neighsum;
+      /* Add in the corner-neighbors */
+      neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
+		  GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
+      /* form final output scaled up by 2^16 */
+      membersum = membersum * memberscale + neighsum * neighscale;
+      /* round, descale and output it */
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
+	       GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
+		GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+    inrow += 2;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
+			    JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int outrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * DCTSIZE;
+  JSAMPROW inptr, above_ptr, below_ptr, outptr;
+  long membersum, neighsum, memberscale, neighscale;
+  int colsum, lastcolsum, nextcolsum;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols);
+
+  /* Each of the eight neighbor pixels contributes a fraction SF to the
+   * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
+   * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
+  neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
+
+  for (outrow = 0; outrow < compptr->v_samp_factor; outrow++) {
+    outptr = output_data[outrow];
+    inptr = input_data[outrow];
+    above_ptr = input_data[outrow-1];
+    below_ptr = input_data[outrow+1];
+
+    /* Special case for first column */
+    colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
+	     GETJSAMPLE(*inptr);
+    membersum = GETJSAMPLE(*inptr++);
+    nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		 GETJSAMPLE(*inptr);
+    neighsum = colsum + (colsum - membersum) + nextcolsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    lastcolsum = colsum; colsum = nextcolsum;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      membersum = GETJSAMPLE(*inptr++);
+      above_ptr++; below_ptr++;
+      nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		   GETJSAMPLE(*inptr);
+      neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
+      membersum = membersum * memberscale + neighsum * neighscale;
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      lastcolsum = colsum; colsum = nextcolsum;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr);
+    neighsum = lastcolsum + (colsum - membersum) + colsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+  }
+}
+
+#endif /* INPUT_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Module initialization routine for downsampling.
+ * Note that we must select a routine for each component.
+ */
+
+GLOBAL(void)
+jinit_downsampler (j_compress_ptr cinfo)
+{
+  my_downsample_ptr downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  int smoothok = TRUE;
+
+  downsample = (my_downsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_downsampler));
+  cinfo->downsample = (struct jpeg_downsampler *) downsample;
+  downsample->pub.start_pass = start_pass_downsample;
+  downsample->pub.downsample = sep_downsample;
+  downsample->pub.need_context_rows = FALSE;
+
+  if (cinfo->CCIR601_sampling)
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* Verify we can handle the sampling factors, and set up method pointers */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor == cinfo->max_h_samp_factor &&
+	compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = fullsize_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	downsample->methods[ci] = fullsize_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+	       compptr->v_samp_factor == cinfo->max_v_samp_factor) {
+      smoothok = FALSE;
+      downsample->methods[ci] = h2v1_downsample;
+    } else if (compptr->h_samp_factor * 2 == cinfo->max_h_samp_factor &&
+	       compptr->v_samp_factor * 2 == cinfo->max_v_samp_factor) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = h2v2_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	downsample->methods[ci] = h2v2_downsample;
+    } else if ((cinfo->max_h_samp_factor % compptr->h_samp_factor) == 0 &&
+	       (cinfo->max_v_samp_factor % compptr->v_samp_factor) == 0) {
+      smoothok = FALSE;
+      downsample->methods[ci] = int_downsample;
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+  }
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+  if (cinfo->smoothing_factor && !smoothok)
+    TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
+#endif
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdapimin.cpp b/ml/dlib/dlib/external/libjpeg/jdapimin.cpp
new file mode 100644
index 000000000..3ea5bf161
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdapimin.cpp
@@ -0,0 +1,395 @@
+/*
+ * jdapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-decompression case or the
+ * transcoding-only case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jdapistd.c.  But also see jcomapi.c for routines
+ * shared by compression and decompression, and jdtrans.c for the transcoding
+ * case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG decompression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateDecompress (j_decompress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_decompress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_decompress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_decompress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = TRUE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->src = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++)
+    cinfo->quant_tbl_ptrs[i] = NULL;
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  /* Initialize marker processor so application can override methods
+   * for COM, APPn markers before calling jpeg_read_header.
+   */
+  cinfo->marker_list = NULL;
+  jinit_marker_reader(cinfo);
+
+  /* And initialize the overall input controller. */
+  jinit_input_controller(cinfo);
+
+  /* OK, I'm ready */
+  cinfo->global_state = DSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG decompression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG decompression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Set default decompression parameters.
+ */
+
+LOCAL(void)
+default_decompress_parms (j_decompress_ptr cinfo)
+{
+  /* Guess the input colorspace, and set output colorspace accordingly. */
+  /* (Wish JPEG committee had provided a real way to specify this...) */
+  /* Note application may override our guesses. */
+  switch (cinfo->num_components) {
+  case 1:
+    cinfo->jpeg_color_space = JCS_GRAYSCALE;
+    cinfo->out_color_space = JCS_GRAYSCALE;
+    break;
+    
+  case 3:
+    if (cinfo->saw_JFIF_marker) {
+      cinfo->jpeg_color_space = JCS_YCbCr; /* JFIF implies YCbCr */
+    } else if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_RGB;
+	break;
+      case 1:
+	cinfo->jpeg_color_space = JCS_YCbCr;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+	break;
+      }
+    } else {
+      /* Saw no special markers, try to guess from the component IDs */
+      int cid0 = cinfo->comp_info[0].component_id;
+      int cid1 = cinfo->comp_info[1].component_id;
+      int cid2 = cinfo->comp_info[2].component_id;
+
+      if (cid0 == 1 && cid1 == 2 && cid2 == 3)
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume JFIF w/out marker */
+      else if (cid0 == 82 && cid1 == 71 && cid2 == 66)
+	cinfo->jpeg_color_space = JCS_RGB; /* ASCII 'R', 'G', 'B' */
+      else {
+	TRACEMS3(cinfo, 1, JTRC_UNKNOWN_IDS, cid0, cid1, cid2);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+      }
+    }
+    /* Always guess RGB is proper output colorspace. */
+    cinfo->out_color_space = JCS_RGB;
+    break;
+    
+  case 4:
+    if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_CMYK;
+	break;
+      case 2:
+	cinfo->jpeg_color_space = JCS_YCCK;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCCK; /* assume it's YCCK */
+	break;
+      }
+    } else {
+      /* No special markers, assume straight CMYK. */
+      cinfo->jpeg_color_space = JCS_CMYK;
+    }
+    cinfo->out_color_space = JCS_CMYK;
+    break;
+    
+  default:
+    cinfo->jpeg_color_space = JCS_UNKNOWN;
+    cinfo->out_color_space = JCS_UNKNOWN;
+    break;
+  }
+
+  /* Set defaults for other decompression parameters. */
+  cinfo->scale_num = 1;		/* 1:1 scaling */
+  cinfo->scale_denom = 1;
+  cinfo->output_gamma = 1.0;
+  cinfo->buffered_image = FALSE;
+  cinfo->raw_data_out = FALSE;
+  cinfo->dct_method = JDCT_DEFAULT;
+  cinfo->do_fancy_upsampling = TRUE;
+  cinfo->do_block_smoothing = TRUE;
+  cinfo->quantize_colors = FALSE;
+  /* We set these in case application only sets quantize_colors. */
+  cinfo->dither_mode = JDITHER_FS;
+#ifdef QUANT_2PASS_SUPPORTED
+  cinfo->two_pass_quantize = TRUE;
+#else
+  cinfo->two_pass_quantize = FALSE;
+#endif
+  cinfo->desired_number_of_colors = 256;
+  cinfo->colormap = NULL;
+  /* Initialize for no mode change in buffered-image mode. */
+  cinfo->enable_1pass_quant = FALSE;
+  cinfo->enable_external_quant = FALSE;
+  cinfo->enable_2pass_quant = FALSE;
+}
+
+
+/*
+ * Decompression startup: read start of JPEG datastream to see what's there.
+ * Need only initialize JPEG object and supply a data source before calling.
+ *
+ * This routine will read as far as the first SOS marker (ie, actual start of
+ * compressed data), and will save all tables and parameters in the JPEG
+ * object.  It will also initialize the decompression parameters to default
+ * values, and finally return JPEG_HEADER_OK.  On return, the application may
+ * adjust the decompression parameters and then call jpeg_start_decompress.
+ * (Or, if the application only wanted to determine the image parameters,
+ * the data need not be decompressed.  In that case, call jpeg_abort or
+ * jpeg_destroy to release any temporary space.)
+ * If an abbreviated (tables only) datastream is presented, the routine will
+ * return JPEG_HEADER_TABLES_ONLY upon reaching EOI.  The application may then
+ * re-use the JPEG object to read the abbreviated image datastream(s).
+ * It is unnecessary (but OK) to call jpeg_abort in this case.
+ * The JPEG_SUSPENDED return code only occurs if the data source module
+ * requests suspension of the decompressor.  In this case the application
+ * should load more source data and then re-call jpeg_read_header to resume
+ * processing.
+ * If a non-suspending data source is used and require_image is TRUE, then the
+ * return code need not be inspected since only JPEG_HEADER_OK is possible.
+ *
+ * This routine is now just a front end to jpeg_consume_input, with some
+ * extra error checking.
+ */
+
+GLOBAL(int)
+jpeg_read_header (j_decompress_ptr cinfo, int require_image)
+{
+  int retcode;
+
+  if (cinfo->global_state != DSTATE_START &&
+      cinfo->global_state != DSTATE_INHEADER)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  retcode = jpeg_consume_input(cinfo);
+
+  switch (retcode) {
+  case JPEG_REACHED_SOS:
+    retcode = JPEG_HEADER_OK;
+    break;
+  case JPEG_REACHED_EOI:
+    if (require_image)		/* Complain if application wanted an image */
+      ERREXIT(cinfo, JERR_NO_IMAGE);
+    /* Reset to start state; it would be safer to require the application to
+     * call jpeg_abort, but we can't change it now for compatibility reasons.
+     * A side effect is to free any temporary memory (there shouldn't be any).
+     */
+    jpeg_abort((j_common_ptr) cinfo); /* sets state = DSTATE_START */
+    retcode = JPEG_HEADER_TABLES_ONLY;
+    break;
+  case JPEG_SUSPENDED:
+    /* no work */
+    break;
+  }
+
+  return retcode;
+}
+
+
+/*
+ * Consume data in advance of what the decompressor requires.
+ * This can be called at any time once the decompressor object has
+ * been created and a data source has been set up.
+ *
+ * This routine is essentially a state machine that handles a couple
+ * of critical state-transition actions, namely initial setup and
+ * transition from header scanning to ready-for-start_decompress.
+ * All the actual input is done via the input controller's consume_input
+ * method.
+ */
+
+GLOBAL(int)
+jpeg_consume_input (j_decompress_ptr cinfo)
+{
+  int retcode = JPEG_SUSPENDED;
+
+  /* NB: every possible DSTATE value should be listed in this switch */
+  switch (cinfo->global_state) {
+  case DSTATE_START:
+    /* Start-of-datastream actions: reset appropriate modules */
+    (*cinfo->inputctl->reset_input_controller) (cinfo);
+    /* Initialize application's data source module */
+    (*cinfo->src->init_source) (cinfo);
+    cinfo->global_state = DSTATE_INHEADER;
+    /*FALLTHROUGH*/
+  case DSTATE_INHEADER:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    if (retcode == JPEG_REACHED_SOS) { /* Found SOS, prepare to decompress */
+      /* Set up default parameters based on header data */
+      default_decompress_parms(cinfo);
+      /* Set global state: ready for start_decompress */
+      cinfo->global_state = DSTATE_READY;
+    }
+    break;
+  case DSTATE_READY:
+    /* Can't advance past first SOS until start_decompress is called */
+    retcode = JPEG_REACHED_SOS;
+    break;
+  case DSTATE_PRELOAD:
+  case DSTATE_PRESCAN:
+  case DSTATE_SCANNING:
+  case DSTATE_RAW_OK:
+  case DSTATE_BUFIMAGE:
+  case DSTATE_BUFPOST:
+  case DSTATE_STOPPING:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    break;
+  default:
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  return retcode;
+}
+
+
+/*
+ * Have we finished reading the input file?
+ */
+
+GLOBAL(int)
+jpeg_input_complete (j_decompress_ptr cinfo)
+{
+  /* Check for valid jpeg object */
+  if (cinfo->global_state < DSTATE_START ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->eoi_reached;
+}
+
+
+/*
+ * Is there more than one scan?
+ */
+
+GLOBAL(int)
+jpeg_has_multiple_scans (j_decompress_ptr cinfo)
+{
+  /* Only valid after jpeg_read_header completes */
+  if (cinfo->global_state < DSTATE_READY ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->has_multiple_scans;
+}
+
+
+/*
+ * Finish JPEG decompression.
+ *
+ * This will normally just verify the file trailer and release temp storage.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(int)
+jpeg_finish_decompress (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && ! cinfo->buffered_image) {
+    /* Terminate final pass of non-buffered mode */
+    if (cinfo->output_scanline < cinfo->output_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state == DSTATE_BUFIMAGE) {
+    /* Finishing after a buffered-image operation */
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state != DSTATE_STOPPING) {
+    /* STOPPING = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read until EOI */
+  while (! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  /* Do final cleanup */
+  (*cinfo->src->term_source) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+  return TRUE;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdapistd.cpp b/ml/dlib/dlib/external/libjpeg/jdapistd.cpp
new file mode 100644
index 000000000..03d909e00
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdapistd.cpp
@@ -0,0 +1,275 @@
+/*
+ * jdapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-decompression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_decompress, it will end up linking in the entire decompressor.
+ * We thus must separate this file from jdapimin.c to avoid linking the
+ * whole decompression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(int) output_pass_setup JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Decompression initialization.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * If a multipass operating mode was selected, this will do all but the
+ * last pass, and thus may take a great deal of time.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(int)
+jpeg_start_decompress (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize master control, select active modules */
+    jinit_master_decompress(cinfo);
+    if (cinfo->buffered_image) {
+      /* No more work here; expecting jpeg_start_output next */
+      cinfo->global_state = DSTATE_BUFIMAGE;
+      return TRUE;
+    }
+    cinfo->global_state = DSTATE_PRELOAD;
+  }
+  if (cinfo->global_state == DSTATE_PRELOAD) {
+    /* If file has multiple scans, absorb them all into the coef buffer */
+    if (cinfo->inputctl->has_multiple_scans) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+      for (;;) {
+	int retcode;
+	/* Call progress monitor hook if present */
+	if (cinfo->progress != NULL)
+	  (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+	/* Absorb some more input */
+	retcode = (*cinfo->inputctl->consume_input) (cinfo);
+	if (retcode == JPEG_SUSPENDED)
+	  return FALSE;
+	if (retcode == JPEG_REACHED_EOI)
+	  break;
+	/* Advance progress counter if appropriate */
+	if (cinfo->progress != NULL &&
+	    (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+	  if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+	    /* jdmaster underestimated number of scans; ratchet up one scan */
+	    cinfo->progress->pass_limit += (long) cinfo->total_iMCU_rows;
+	  }
+	}
+      }
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+    }
+    cinfo->output_scan_number = cinfo->input_scan_number;
+  } else if (cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any dummy output passes, and set up for the final pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Set up for an output pass, and perform any dummy pass(es) needed.
+ * Common subroutine for jpeg_start_decompress and jpeg_start_output.
+ * Entry: global_state = DSTATE_PRESCAN only if previously suspended.
+ * Exit: If done, returns TRUE and sets global_state for proper output mode.
+ *       If suspended, returns FALSE and sets global_state = DSTATE_PRESCAN.
+ */
+
+LOCAL(int)
+output_pass_setup (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state != DSTATE_PRESCAN) {
+    /* First call: do pass setup */
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+    cinfo->global_state = DSTATE_PRESCAN;
+  }
+  /* Loop over any required dummy passes */
+  while (cinfo->master->is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Crank through the dummy pass */
+    while (cinfo->output_scanline < cinfo->output_height) {
+      JDIMENSION last_scanline;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+	cinfo->progress->pass_limit = (long) cinfo->output_height;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* Process some data */
+      last_scanline = cinfo->output_scanline;
+      (*cinfo->main->process_data) (cinfo, (JSAMPARRAY) NULL,
+				    &cinfo->output_scanline, (JDIMENSION) 0);
+      if (cinfo->output_scanline == last_scanline)
+	return FALSE;		/* No progress made, must suspend */
+    }
+    /* Finish up dummy pass, and set up for another one */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  }
+  /* Ready for application to drive output pass through
+   * jpeg_read_scanlines or jpeg_read_raw_data.
+   */
+  cinfo->global_state = cinfo->raw_data_out ? DSTATE_RAW_OK : DSTATE_SCANNING;
+  return TRUE;
+}
+
+
+/*
+ * Read some scanlines of data from the JPEG decompressor.
+ *
+ * The return value will be the number of lines actually read.
+ * This may be less than the number requested in several cases,
+ * including bottom of image, data source suspension, and operating
+ * modes that emit multiple scanlines at a time.
+ *
+ * Note: we warn about excess calls to jpeg_read_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * an oversize buffer (max_lines > scanlines remaining) is not an error.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_scanlines (j_decompress_ptr cinfo, JSAMPARRAY scanlines,
+		     JDIMENSION max_lines)
+{
+  JDIMENSION row_ctr;
+
+  if (cinfo->global_state != DSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Process some data */
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, max_lines);
+  cinfo->output_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to read raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_raw_data (j_decompress_ptr cinfo, JSAMPIMAGE data,
+		    JDIMENSION max_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != DSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Verify that at least one iMCU row can be returned. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * cinfo->min_DCT_scaled_size;
+  if (max_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Decompress directly into user's buffer. */
+  if (! (*cinfo->coef->decompress_data) (cinfo, data))
+    return 0;			/* suspension forced, can do nothing more */
+
+  /* OK, we processed one iMCU row. */
+  cinfo->output_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
+
+
+/* Additional entry points for buffered-image mode. */
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Initialize for an output pass in buffered-image mode.
+ */
+
+GLOBAL(int)
+jpeg_start_output (j_decompress_ptr cinfo, int scan_number)
+{
+  if (cinfo->global_state != DSTATE_BUFIMAGE &&
+      cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Limit scan number to valid range */
+  if (scan_number <= 0)
+    scan_number = 1;
+  if (cinfo->inputctl->eoi_reached &&
+      scan_number > cinfo->input_scan_number)
+    scan_number = cinfo->input_scan_number;
+  cinfo->output_scan_number = scan_number;
+  /* Perform any dummy output passes, and set up for the real pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Finish up after an output pass in buffered-image mode.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(int)
+jpeg_finish_output (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && cinfo->buffered_image) {
+    /* Terminate this pass. */
+    /* We do not require the whole pass to have been completed. */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_BUFPOST;
+  } else if (cinfo->global_state != DSTATE_BUFPOST) {
+    /* BUFPOST = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read markers looking for SOS or EOI */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  cinfo->global_state = DSTATE_BUFIMAGE;
+  return TRUE;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jdatadst.cpp b/ml/dlib/dlib/external/libjpeg/jdatadst.cpp
new file mode 100644
index 000000000..afa3c83c6
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdatadst.cpp
@@ -0,0 +1,151 @@
+/*
+ * jdatadst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains compression data destination routines for the case of
+ * emitting JPEG data to a file (or any stdio stream).  While these routines
+ * are sufficient for most applications, some will want to use a different
+ * destination manager.
+ * IMPORTANT: we assume that fwrite() will correctly transcribe an array of
+ * JOCTETs into 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data destination object for stdio output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  FILE * outfile;		/* target stream */
+  JOCTET * buffer;		/* start of buffer */
+} my_destination_mgr;
+
+typedef my_destination_mgr * my_dest_ptr;
+
+#define OUTPUT_BUF_SIZE  4096	/* choose an efficiently fwrite'able size */
+
+
+/*
+ * Initialize destination --- called by jpeg_start_compress
+ * before any data is actually written.
+ */
+
+METHODDEF(void)
+init_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  /* Allocate the output buffer --- it will be released when done with image */
+  dest->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  OUTPUT_BUF_SIZE * SIZEOF(JOCTET));
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+}
+
+
+/*
+ * Empty the output buffer --- called whenever buffer fills up.
+ *
+ * In typical applications, this should write the entire output buffer
+ * (ignoring the current state of next_output_byte & free_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been dumped.
+ *
+ * In applications that need to be able to suspend compression due to output
+ * overrun, a FALSE return indicates that the buffer cannot be emptied now.
+ * In this situation, the compressor will return to its caller (possibly with
+ * an indication that it has not accepted all the supplied scanlines).  The
+ * application should resume compression after it has made more room in the
+ * output buffer.  Note that there are substantial restrictions on the use of
+ * suspension --- see the documentation.
+ *
+ * When suspending, the compressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_output_byte & free_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point will be regenerated after resumption, so do not
+ * write it out when emptying the buffer externally.
+ */
+
+METHODDEF(int)
+empty_output_buffer (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  if (JFWRITE(dest->outfile, dest->buffer, OUTPUT_BUF_SIZE) !=
+      (size_t) OUTPUT_BUF_SIZE)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+
+  return TRUE;
+}
+
+
+/*
+ * Terminate destination --- called by jpeg_finish_compress
+ * after all data has been written.  Usually needs to flush buffer.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+  size_t datacount = OUTPUT_BUF_SIZE - dest->pub.free_in_buffer;
+
+  /* Write any data remaining in the buffer */
+  if (datacount > 0) {
+    if (JFWRITE(dest->outfile, dest->buffer, datacount) != datacount)
+      ERREXIT(cinfo, JERR_FILE_WRITE);
+  }
+  fflush(dest->outfile);
+  /* Make sure we wrote the output file OK */
+  if (ferror(dest->outfile))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+
+/*
+ * Prepare for output to a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing compression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_dest (j_compress_ptr cinfo, FILE * outfile)
+{
+  my_dest_ptr dest;
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same file without re-executing jpeg_stdio_dest.
+   * This makes it dangerous to use this manager and a different destination
+   * manager serially with the same JPEG object, because their private object
+   * sizes may be different.  Caveat programmer.
+   */
+  if (cinfo->dest == NULL) {	/* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_destination_mgr));
+  }
+
+  dest = (my_dest_ptr) cinfo->dest;
+  dest->pub.init_destination = init_destination;
+  dest->pub.empty_output_buffer = empty_output_buffer;
+  dest->pub.term_destination = term_destination;
+  dest->outfile = outfile;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdatasrc.cpp b/ml/dlib/dlib/external/libjpeg/jdatasrc.cpp
new file mode 100644
index 000000000..7af097f02
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdatasrc.cpp
@@ -0,0 +1,212 @@
+/*
+ * jdatasrc.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains decompression data source routines for the case of
+ * reading JPEG data from a file (or any stdio stream).  While these routines
+ * are sufficient for most applications, some will want to use a different
+ * source manager.
+ * IMPORTANT: we assume that fread() will correctly transcribe an array of
+ * JOCTETs from 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data source object for stdio input */
+
+typedef struct {
+  struct jpeg_source_mgr pub;	/* public fields */
+
+  FILE * infile;		/* source stream */
+  JOCTET * buffer;		/* start of buffer */
+  int start_of_file;	/* have we gotten any data yet? */
+} my_source_mgr;
+
+typedef my_source_mgr * my_src_ptr;
+
+#define INPUT_BUF_SIZE  4096	/* choose an efficiently fread'able size */
+
+
+/*
+ * Initialize source --- called by jpeg_read_header
+ * before any data is actually read.
+ */
+
+METHODDEF(void)
+init_source (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+
+  /* We reset the empty-input-file flag for each image,
+   * but we don't clear the input buffer.
+   * This is correct behavior for reading a series of images from one source.
+   */
+  src->start_of_file = TRUE;
+}
+
+
+/*
+ * Fill the input buffer --- called whenever buffer is emptied.
+ *
+ * In typical applications, this should read fresh data into the buffer
+ * (ignoring the current state of next_input_byte & bytes_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been reloaded.  It is not necessary to
+ * fill the buffer entirely, only to obtain at least one more byte.
+ *
+ * There is no such thing as an EOF return.  If the end of the file has been
+ * reached, the routine has a choice of ERREXIT() or inserting fake data into
+ * the buffer.  In most cases, generating a warning message and inserting a
+ * fake EOI marker is the best course of action --- this will allow the
+ * decompressor to output however much of the image is there.  However,
+ * the resulting error message is misleading if the real problem is an empty
+ * input file, so we handle that case specially.
+ *
+ * In applications that need to be able to suspend compression due to input
+ * not being available yet, a FALSE return indicates that no more data can be
+ * obtained right now, but more may be forthcoming later.  In this situation,
+ * the decompressor will return to its caller (with an indication of the
+ * number of scanlines it has read, if any).  The application should resume
+ * decompression after it has loaded more data into the input buffer.  Note
+ * that there are substantial restrictions on the use of suspension --- see
+ * the documentation.
+ *
+ * When suspending, the decompressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_input_byte & bytes_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point must be rescanned after resumption, so move it to
+ * the front of the buffer rather than discarding it.
+ */
+
+METHODDEF(int)
+fill_input_buffer (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+  size_t nbytes;
+
+  nbytes = JFREAD(src->infile, src->buffer, INPUT_BUF_SIZE);
+
+  if (nbytes <= 0) {
+    if (src->start_of_file)	/* Treat empty input file as fatal error */
+      ERREXIT(cinfo, JERR_INPUT_EMPTY);
+    WARNMS(cinfo, JWRN_JPEG_EOF);
+    /* Insert a fake EOI marker */
+    src->buffer[0] = (JOCTET) 0xFF;
+    src->buffer[1] = (JOCTET) JPEG_EOI;
+    nbytes = 2;
+  }
+
+  src->pub.next_input_byte = src->buffer;
+  src->pub.bytes_in_buffer = nbytes;
+  src->start_of_file = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Skip data --- used to skip over a potentially large amount of
+ * uninteresting data (such as an APPn marker).
+ *
+ * Writers of suspendable-input applications must note that skip_input_data
+ * is not granted the right to give a suspension return.  If the skip extends
+ * beyond the data currently in the buffer, the buffer can be marked empty so
+ * that the next read will cause a fill_input_buffer call that can suspend.
+ * Arranging for additional bytes to be discarded before reloading the input
+ * buffer is the application writer's problem.
+ */
+
+METHODDEF(void)
+skip_input_data (j_decompress_ptr cinfo, long num_bytes)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+
+  /* Just a dumb implementation for now.  Could use fseek() except
+   * it doesn't work on pipes.  Not clear that being smart is worth
+   * any trouble anyway --- large skips are infrequent.
+   */
+  if (num_bytes > 0) {
+    while (num_bytes > (long) src->pub.bytes_in_buffer) {
+      num_bytes -= (long) src->pub.bytes_in_buffer;
+      (void) fill_input_buffer(cinfo);
+      /* note we assume that fill_input_buffer will never return FALSE,
+       * so suspension need not be handled.
+       */
+    }
+    src->pub.next_input_byte += (size_t) num_bytes;
+    src->pub.bytes_in_buffer -= (size_t) num_bytes;
+  }
+}
+
+
+/*
+ * An additional method that can be provided by data source modules is the
+ * resync_to_restart method for error recovery in the presence of RST markers.
+ * For the moment, this source module just uses the default resync method
+ * provided by the JPEG library.  That method assumes that no backtracking
+ * is possible.
+ */
+
+
+/*
+ * Terminate source --- called by jpeg_finish_decompress
+ * after all data has been read.  Often a no-op.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_source (j_decompress_ptr )
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Prepare for input from a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing decompression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_src (j_decompress_ptr cinfo, FILE * infile)
+{
+  my_src_ptr src;
+
+  /* The source object and input buffer are made permanent so that a series
+   * of JPEG images can be read from the same file by calling jpeg_stdio_src
+   * only before the first one.  (If we discarded the buffer at the end of
+   * one image, we'd likely lose the start of the next one.)
+   * This makes it unsafe to use this manager and a different source
+   * manager serially with the same JPEG object.  Caveat programmer.
+   */
+  if (cinfo->src == NULL) {	/* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_source_mgr));
+    src = (my_src_ptr) cinfo->src;
+    src->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  INPUT_BUF_SIZE * SIZEOF(JOCTET));
+  }
+
+  src = (my_src_ptr) cinfo->src;
+  src->pub.init_source = init_source;
+  src->pub.fill_input_buffer = fill_input_buffer;
+  src->pub.skip_input_data = skip_input_data;
+  src->pub.resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->pub.term_source = term_source;
+  src->infile = infile;
+  src->pub.bytes_in_buffer = 0; /* forces fill_input_buffer on first read */
+  src->pub.next_input_byte = NULL; /* until buffer loaded */
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdcoefct.cpp b/ml/dlib/dlib/external/libjpeg/jdcoefct.cpp
new file mode 100644
index 000000000..11b618920
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdcoefct.cpp
@@ -0,0 +1,736 @@
+/*
+ * jdcoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for decompression.
+ * This controller is the top level of the JPEG decompressor proper.
+ * The coefficient buffer lies between entropy decoding and inverse-DCT steps.
+ *
+ * In buffered-image mode, this controller is the interface between
+ * input-oriented processing and output-oriented processing.
+ * Also, the input side (only) is used when reading a file for transcoding.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+/* Block smoothing is only applicable for progressive JPEG, so: */
+#ifndef D_PROGRESSIVE_SUPPORTED
+#undef BLOCK_SMOOTHING_SUPPORTED
+#endif
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_coef_controller pub; /* public fields */
+
+  /* These variables keep track of the current location of the input side. */
+  /* cinfo->input_iMCU_row is also used for this. */
+  JDIMENSION MCU_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* The output side's location is represented by cinfo->output_iMCU_row. */
+
+  /* In single-pass modes, it's sufficient to buffer just one MCU.
+   * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
+   * and let the entropy decoder write into that workspace each time.
+   * (On 80x86, the workspace is FAR even though it's not really very big;
+   * this is to keep the module interfaces unchanged when a large coefficient
+   * buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays; it is used only by the input side.
+   */
+  JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+#endif
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  /* When doing block smoothing, we latch coefficient Al values here */
+  int * coef_bits_latch;
+#define SAVED_COEFS  6		/* we save coef_bits[0..5] */
+#endif
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+/* Forward declarations */
+METHODDEF(int) decompress_onepass
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+METHODDEF(int) decompress_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+LOCAL(int) smoothing_ok JPP((j_decompress_ptr cinfo));
+METHODDEF(int) decompress_smooth_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_decompress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row (input side) */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->MCU_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for an input processing pass.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->input_iMCU_row = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Initialize for an output processing pass.
+ */
+
+METHODDEF(void)
+start_output_pass (j_decompress_ptr cinfo)
+{
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* If multipass, check to see whether to use block smoothing on this pass */
+  if (coef->pub.coef_arrays != NULL) {
+    if (cinfo->do_block_smoothing && smoothing_ok(cinfo))
+      coef->pub.decompress_data = decompress_smooth_data;
+    else
+      coef->pub.decompress_data = decompress_data;
+  }
+#endif
+  cinfo->output_iMCU_row = 0;
+}
+
+
+/*
+ * Decompress and return some data in the single-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Input and output must run in lockstep since we have only a one-MCU buffer.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(int)
+decompress_onepass (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, useful_width;
+  JSAMPARRAY output_ptr;
+  JDIMENSION start_col, output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Loop to process as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Try to fetch an MCU.  Entropy decoder expects buffer to be zeroed. */
+      jzero_far((void FAR *) coef->MCU_buffer[0],
+		(size_t) (cinfo->blocks_in_MCU * SIZEOF(JBLOCK)));
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+      /* Determine where data should go in output_buf and do the IDCT thing.
+       * We skip dummy blocks at the right and bottom edges (but blkn gets
+       * incremented past them!).  Note the inner loop relies on having
+       * allocated the MCU_buffer[] blocks sequentially.
+       */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	/* Don't bother to IDCT an uninteresting component. */
+	if (! compptr->component_needed) {
+	  blkn += compptr->MCU_blocks;
+	  continue;
+	}
+	inverse_DCT = cinfo->idct->inverse_DCT[compptr->component_index];
+	useful_width = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						    : compptr->last_col_width;
+	output_ptr = output_buf[compptr->component_index] +
+	  yoffset * compptr->DCT_scaled_size;
+	start_col = MCU_col_num * compptr->MCU_sample_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (cinfo->input_iMCU_row < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    output_col = start_col;
+	    for (xindex = 0; xindex < useful_width; xindex++) {
+	      (*inverse_DCT) (cinfo, compptr,
+			      (JCOEFPTR) coef->MCU_buffer[blkn+xindex],
+			      output_ptr, output_col);
+	      output_col += compptr->DCT_scaled_size;
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  output_ptr += compptr->DCT_scaled_size;
+	}
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  cinfo->output_iMCU_row++;
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Dummy consume-input routine for single-pass operation.
+ */
+
+METHODDEF(int)
+dummy_consume_data (j_decompress_ptr )
+{
+  return JPEG_SUSPENDED;	/* Always indicate nothing was done */
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Consume input data and store it in the full-image coefficient buffer.
+ * We read as much as one fully interleaved MCU row ("iMCU" row) per call,
+ * ie, v_samp_factor block rows for each component in the scan.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ */
+
+METHODDEF(int)
+consume_data (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       cinfo->input_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Note: entropy decoder expects buffer to be zeroed,
+     * but this is handled automatically by the memory manager
+     * because we requested a pre-zeroed array.
+     */
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to fetch the MCU. */
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Decompress and return some data in the multi-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image.
+ */
+
+METHODDEF(int)
+decompress_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num;
+  int ci, block_row, block_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number < cinfo->output_scan_number ||
+	 (cinfo->input_scan_number == cinfo->output_scan_number &&
+	  cinfo->input_iMCU_row <= cinfo->output_iMCU_row)) {
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       cinfo->output_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      output_col = 0;
+      for (block_num = 0; block_num < compptr->width_in_blocks; block_num++) {
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) buffer_ptr,
+			output_ptr, output_col);
+	buffer_ptr++;
+	output_col += compptr->DCT_scaled_size;
+      }
+      output_ptr += compptr->DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+
+/*
+ * This code applies interblock smoothing as described by section K.8
+ * of the JPEG standard: the first 5 AC coefficients are estimated from
+ * the DC values of a DCT block and its 8 neighboring blocks.
+ * We apply smoothing only for progressive JPEG decoding, and only if
+ * the coefficients it can estimate are not yet known to full precision.
+ */
+
+/* Natural-order array positions of the first 5 zigzag-order coefficients */
+#define Q01_POS  1
+#define Q10_POS  8
+#define Q20_POS  16
+#define Q11_POS  9
+#define Q02_POS  2
+
+/*
+ * Determine whether block smoothing is applicable and safe.
+ * We also latch the current states of the coef_bits[] entries for the
+ * AC coefficients; otherwise, if the input side of the decompressor
+ * advances into a new scan, we might think the coefficients are known
+ * more accurately than they really are.
+ */
+
+LOCAL(int)
+smoothing_ok (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  int smoothing_useful = FALSE;
+  int ci, coefi;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtable;
+  int * coef_bits;
+  int * coef_bits_latch;
+
+  if (! cinfo->progressive_mode || cinfo->coef_bits == NULL)
+    return FALSE;
+
+  /* Allocate latch area if not already done */
+  if (coef->coef_bits_latch == NULL)
+    coef->coef_bits_latch = (int *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components *
+				  (SAVED_COEFS * SIZEOF(int)));
+  coef_bits_latch = coef->coef_bits_latch;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* All components' quantization values must already be latched. */
+    if ((qtable = compptr->quant_table) == NULL)
+      return FALSE;
+    /* Verify DC & first 5 AC quantizers are nonzero to avoid zero-divide. */
+    if (qtable->quantval[0] == 0 ||
+	qtable->quantval[Q01_POS] == 0 ||
+	qtable->quantval[Q10_POS] == 0 ||
+	qtable->quantval[Q20_POS] == 0 ||
+	qtable->quantval[Q11_POS] == 0 ||
+	qtable->quantval[Q02_POS] == 0)
+      return FALSE;
+    /* DC values must be at least partly known for all components. */
+    coef_bits = cinfo->coef_bits[ci];
+    if (coef_bits[0] < 0)
+      return FALSE;
+    /* Block smoothing is helpful if some AC coefficients remain inaccurate. */
+    for (coefi = 1; coefi <= 5; coefi++) {
+      coef_bits_latch[coefi] = coef_bits[coefi];
+      if (coef_bits[coefi] != 0)
+	smoothing_useful = TRUE;
+    }
+    coef_bits_latch += SAVED_COEFS;
+  }
+
+  return smoothing_useful;
+}
+
+
+/*
+ * Variant of decompress_data for use when doing block smoothing.
+ */
+
+METHODDEF(int)
+decompress_smooth_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num, last_block_column;
+  int ci, block_row, block_rows, access_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr, prev_block_row, next_block_row;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+  int first_row, last_row;
+  JBLOCK workspace;
+  int *coef_bits;
+  JQUANT_TBL *quanttbl;
+  long Q00,Q01,Q02,Q10,Q11,Q20, num;
+  int DC1,DC2,DC3,DC4,DC5,DC6,DC7,DC8,DC9;
+  int Al, pred;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if (cinfo->input_scan_number == cinfo->output_scan_number) {
+      /* If input is working on current scan, we ordinarily want it to
+       * have completed the current row.  But if input scan is DC,
+       * we want it to keep one row ahead so that next block row's DC
+       * values are up to date.
+       */
+      JDIMENSION delta = (cinfo->Ss == 0) ? 1 : 0;
+      if (cinfo->input_iMCU_row > cinfo->output_iMCU_row+delta)
+	break;
+    }
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row) {
+      block_rows = compptr->v_samp_factor;
+      access_rows = block_rows * 2; /* this and next iMCU row */
+      last_row = FALSE;
+    } else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+      access_rows = block_rows; /* this iMCU row only */
+      last_row = TRUE;
+    }
+    /* Align the virtual buffer for this component. */
+    if (cinfo->output_iMCU_row > 0) {
+      access_rows += compptr->v_samp_factor; /* prior iMCU row too */
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
+	 (JDIMENSION) access_rows, FALSE);
+      buffer += compptr->v_samp_factor;	/* point to current iMCU row */
+      first_row = FALSE;
+    } else {
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (JDIMENSION) 0, (JDIMENSION) access_rows, FALSE);
+      first_row = TRUE;
+    }
+    /* Fetch component-dependent info */
+    coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+    quanttbl = compptr->quant_table;
+    Q00 = quanttbl->quantval[0];
+    Q01 = quanttbl->quantval[Q01_POS];
+    Q10 = quanttbl->quantval[Q10_POS];
+    Q20 = quanttbl->quantval[Q20_POS];
+    Q11 = quanttbl->quantval[Q11_POS];
+    Q02 = quanttbl->quantval[Q02_POS];
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      if (first_row && block_row == 0)
+	prev_block_row = buffer_ptr;
+      else
+	prev_block_row = buffer[block_row-1];
+      if (last_row && block_row == block_rows-1)
+	next_block_row = buffer_ptr;
+      else
+	next_block_row = buffer[block_row+1];
+      /* We fetch the surrounding DC values using a sliding-register approach.
+       * Initialize all nine here so as to do the right thing on narrow pics.
+       */
+      DC1 = DC2 = DC3 = (int) prev_block_row[0][0];
+      DC4 = DC5 = DC6 = (int) buffer_ptr[0][0];
+      DC7 = DC8 = DC9 = (int) next_block_row[0][0];
+      output_col = 0;
+      last_block_column = compptr->width_in_blocks - 1;
+      for (block_num = 0; block_num <= last_block_column; block_num++) {
+	/* Fetch current DCT block into workspace so we can modify it. */
+	jcopy_block_row(buffer_ptr, (JBLOCKROW) workspace, (JDIMENSION) 1);
+	/* Update DC values */
+	if (block_num < last_block_column) {
+	  DC3 = (int) prev_block_row[1][0];
+	  DC6 = (int) buffer_ptr[1][0];
+	  DC9 = (int) next_block_row[1][0];
+	}
+	/* Compute coefficient estimates per K.8.
+	 * An estimate is applied only if coefficient is still zero,
+	 * and is not known to be fully accurate.
+	 */
+	/* AC01 */
+	if ((Al=coef_bits[1]) != 0 && workspace[1] == 0) {
+	  num = 36 * Q00 * (DC4 - DC6);
+	  if (num >= 0) {
+	    pred = (int) (((Q01<<7) + num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q01<<7) - num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[1] = (JCOEF) pred;
+	}
+	/* AC10 */
+	if ((Al=coef_bits[2]) != 0 && workspace[8] == 0) {
+	  num = 36 * Q00 * (DC2 - DC8);
+	  if (num >= 0) {
+	    pred = (int) (((Q10<<7) + num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q10<<7) - num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[8] = (JCOEF) pred;
+	}
+	/* AC20 */
+	if ((Al=coef_bits[3]) != 0 && workspace[16] == 0) {
+	  num = 9 * Q00 * (DC2 + DC8 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q20<<7) + num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q20<<7) - num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[16] = (JCOEF) pred;
+	}
+	/* AC11 */
+	if ((Al=coef_bits[4]) != 0 && workspace[9] == 0) {
+	  num = 5 * Q00 * (DC1 - DC3 - DC7 + DC9);
+	  if (num >= 0) {
+	    pred = (int) (((Q11<<7) + num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q11<<7) - num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[9] = (JCOEF) pred;
+	}
+	/* AC02 */
+	if ((Al=coef_bits[5]) != 0 && workspace[2] == 0) {
+	  num = 9 * Q00 * (DC4 + DC6 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q02<<7) + num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q02<<7) - num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[2] = (JCOEF) pred;
+	}
+	/* OK, do the IDCT */
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) workspace,
+			output_ptr, output_col);
+	/* Advance for next column */
+	DC1 = DC2; DC2 = DC3;
+	DC4 = DC5; DC5 = DC6;
+	DC7 = DC8; DC8 = DC9;
+	buffer_ptr++, prev_block_row++, next_block_row++;
+	output_col += compptr->DCT_scaled_size;
+      }
+      output_ptr += compptr->DCT_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* BLOCK_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_coef_controller (j_decompress_ptr cinfo, int need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_d_coef_controller *) coef;
+  coef->pub.start_input_pass = start_input_pass;
+  coef->pub.start_output_pass = start_output_pass;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  coef->coef_bits_latch = NULL;
+#endif
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    /* Note we ask for a pre-zeroed array. */
+    int ci, access_rows;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      access_rows = compptr->v_samp_factor;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+      /* If block smoothing could be used, need a bigger window */
+      if (cinfo->progressive_mode)
+	access_rows *= 3;
+#endif
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, TRUE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) access_rows);
+    }
+    coef->pub.consume_data = consume_data;
+    coef->pub.decompress_data = decompress_data;
+    coef->pub.coef_arrays = coef->whole_image; /* link to virtual arrays */
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  D_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < D_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->pub.consume_data = dummy_consume_data;
+    coef->pub.decompress_data = decompress_onepass;
+    coef->pub.coef_arrays = NULL; /* flag for no virtual arrays */
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdcolor.cpp b/ml/dlib/dlib/external/libjpeg/jdcolor.cpp
new file mode 100644
index 000000000..8dd88bfd3
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdcolor.cpp
@@ -0,0 +1,396 @@
+/*
+ * jdcolor.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_deconverter pub; /* public fields */
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  long * Cr_g_tab;		/* => table for Cr to G conversion */
+  long * Cb_g_tab;		/* => table for Cb to G conversion */
+} my_color_deconverter;
+
+typedef my_color_deconverter * my_cconvert_ptr;
+
+
+/**************** YCbCr -> RGB conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	R = Y                + 1.40200 * Cr
+ *	G = Y - 0.34414 * Cb - 0.71414 * Cr
+ *	B = Y + 1.77200 * Cb
+ * where Cb and Cr represent the incoming values less CENTERJSAMPLE.
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ * Notice that Y, being an integral input, does not contribute any fraction
+ * so it need not participate in the rounding.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times Cb and Cr for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The Cr=>R and Cb=>B values can be rounded to integers in advance; the
+ * values for the G calculation are left scaled up, since we must add them
+ * together before rounding.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((long) 1 << (SCALEBITS-1))
+#define FIX(x)		((long) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int i;
+  long x;
+  SHIFT_TEMPS
+
+  cconvert->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cr_g_tab = (long *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(long));
+  cconvert->Cb_g_tab = (long *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(long));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    cconvert->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    cconvert->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    cconvert->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    cconvert->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the output colorspace.
+ *
+ * Note that we change from noninterleaved, one-plane-per-component format
+ * to interleaved-pixel format.  The output buffer is therefore three times
+ * as wide as the input buffer.
+ * A starting row offset is provided only for the input buffer.  The caller
+ * can easily adjust the passed output_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+ycc_rgb_convert (j_decompress_ptr cinfo,
+		 JSAMPIMAGE input_buf, JDIMENSION input_row,
+		 JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int y, cb, cr;
+  JSAMPROW outptr;
+  JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = cconvert->Cr_r_tab;
+  int * Cbbtab = cconvert->Cb_b_tab;
+  long * Crgtab = cconvert->Cr_g_tab;
+  long * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[RGB_RED] =   range_limit[y + Crrtab[cr]];
+      outptr[RGB_GREEN] = range_limit[y +
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS))];
+      outptr[RGB_BLUE] =  range_limit[y + Cbbtab[cb]];
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}
+
+
+/**************** Cases other than YCbCr -> RGB **************/
+
+
+/*
+ * Color conversion for no colorspace change: just copy the data,
+ * converting from separate-planes to interleaved representation.
+ */
+
+METHODDEF(void)
+null_convert (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION input_row,
+	      JSAMPARRAY output_buf, int num_rows)
+{
+  JSAMPROW inptr, outptr;
+  JDIMENSION count;
+  int num_components = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->output_width;
+  int ci;
+
+  while (--num_rows >= 0) {
+    for (ci = 0; ci < num_components; ci++) {
+      inptr = input_buf[ci][input_row];
+      outptr = output_buf[0] + ci;
+      for (count = num_cols; count > 0; count--) {
+	*outptr = *inptr++;	/* needn't bother with GETJSAMPLE() here */
+	outptr += num_components;
+      }
+    }
+    input_row++;
+    output_buf++;
+  }
+}
+
+
+/*
+ * Color conversion for grayscale: just copy the data.
+ * This also works for YCbCr -> grayscale conversion, in which
+ * we just copy the Y (luminance) component and ignore chrominance.
+ */
+
+METHODDEF(void)
+grayscale_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  jcopy_sample_rows(input_buf[0], (int) input_row, output_buf, 0,
+		    num_rows, cinfo->output_width);
+}
+
+
+/*
+ * Convert grayscale to RGB: just duplicate the graylevel three times.
+ * This is provided to support applications that don't want to cope
+ * with grayscale as a separate case.
+ */
+
+METHODDEF(void)
+gray_rgb_convert (j_decompress_ptr cinfo,
+		  JSAMPIMAGE input_buf, JDIMENSION input_row,
+		  JSAMPARRAY output_buf, int num_rows)
+{
+  JSAMPROW inptr, outptr;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    inptr = input_buf[0][input_row++];
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      /* We can dispense with GETJSAMPLE() here */
+      outptr[RGB_RED] = outptr[RGB_GREEN] = outptr[RGB_BLUE] = inptr[col];
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}
+
+
+/*
+ * Adobe-style YCCK->CMYK conversion.
+ * We convert YCbCr to R=1-C, G=1-M, and B=1-Y using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume build_ycc_rgb_table has been called.
+ */
+
+METHODDEF(void)
+ycck_cmyk_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int y, cb, cr;
+  JSAMPROW outptr;
+  JSAMPROW inptr0, inptr1, inptr2, inptr3;
+  JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = cconvert->Cr_r_tab;
+  int * Cbbtab = cconvert->Cb_b_tab;
+  long * Crgtab = cconvert->Cr_g_tab;
+  long * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    inptr3 = input_buf[3][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[0] = range_limit[MAXJSAMPLE - (y + Crrtab[cr])];	/* red */
+      outptr[1] = range_limit[MAXJSAMPLE - (y +			/* green */
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS)))];
+      outptr[2] = range_limit[MAXJSAMPLE - (y + Cbbtab[cb])];	/* blue */
+      /* K passes through unchanged */
+      outptr[3] = inptr3[col];	/* don't need GETJSAMPLE here */
+      outptr += 4;
+    }
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+start_pass_dcolor (j_decompress_ptr )
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for output colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_deconverter (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+  int ci;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_deconverter));
+  cinfo->cconvert = (struct jpeg_color_deconverter *) cconvert;
+  cconvert->pub.start_pass = start_pass_dcolor;
+
+  /* Make sure num_components agrees with jpeg_color_space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->num_components < 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+  }
+
+  /* Set out_color_components and conversion method based on requested space.
+   * Also clear the component_needed flags for any unused components,
+   * so that earlier pipeline stages can avoid useless computation.
+   */
+
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    if (cinfo->jpeg_color_space == JCS_GRAYSCALE ||
+	cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub.color_convert = grayscale_convert;
+      /* For color->grayscale conversion, only the Y (0) component is needed */
+      for (ci = 1; ci < cinfo->num_components; ci++)
+	cinfo->comp_info[ci].component_needed = FALSE;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+    cinfo->out_color_components = RGB_PIXELSIZE;
+    if (cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub.color_convert = ycc_rgb_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+      cconvert->pub.color_convert = gray_rgb_convert;
+    } else if (cinfo->jpeg_color_space == JCS_RGB && RGB_PIXELSIZE == 3) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    cinfo->out_color_components = 4;
+    if (cinfo->jpeg_color_space == JCS_YCCK) {
+      cconvert->pub.color_convert = ycck_cmyk_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_CMYK) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:
+    /* Permit null conversion to same output space */
+    if (cinfo->out_color_space == cinfo->jpeg_color_space) {
+      cinfo->out_color_components = cinfo->num_components;
+      cconvert->pub.color_convert = null_convert;
+    } else			/* unsupported non-null conversion */
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+  }
+
+  if (cinfo->quantize_colors)
+    cinfo->output_components = 1; /* single colormapped output component */
+  else
+    cinfo->output_components = cinfo->out_color_components;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdct.h b/ml/dlib/dlib/external/libjpeg/jdct.h
new file mode 100644
index 000000000..a89c9550b
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdct.h
@@ -0,0 +1,176 @@
+/*
+ * jdct.h
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This include file contains common declarations for the forward and
+ * inverse DCT modules.  These declarations are private to the DCT managers
+ * (jcdctmgr.c, jddctmgr.c) and the individual DCT algorithms.
+ * The individual DCT algorithms are kept in separate files to ease 
+ * machine-dependent tuning (e.g., assembly coding).
+ */
+
+
+/*
+ * A forward DCT routine is given a pointer to a work area of type DCTELEM[];
+ * the DCT is to be performed in-place in that buffer.  Type DCTELEM is int
+ * for 8-bit samples, long for 12-bit samples.  (NOTE: Floating-point DCT
+ * implementations use an array of type FAST_FLOAT, instead.)
+ * The DCT inputs are expected to be signed (range +-CENTERJSAMPLE).
+ * The DCT outputs are returned scaled up by a factor of 8; they therefore
+ * have a range of +-8K for 8-bit data, +-128K for 12-bit data.  This
+ * convention improves accuracy in integer implementations and saves some
+ * work in floating-point ones.
+ * Quantization of the output coefficients is done by jcdctmgr.c.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef int DCTELEM;		/* 16 or 32 bits is fine */
+#else
+typedef long DCTELEM;		/* must have 32 bits */
+#endif
+
+typedef JMETHOD(void, forward_DCT_method_ptr, (DCTELEM * data));
+typedef JMETHOD(void, float_DCT_method_ptr, (FAST_FLOAT * data));
+
+
+/*
+ * An inverse DCT routine is given a pointer to the input JBLOCK and a pointer
+ * to an output sample array.  The routine must dequantize the input data as
+ * well as perform the IDCT; for dequantization, it uses the multiplier table
+ * pointed to by compptr->dct_table.  The output data is to be placed into the
+ * sample array starting at a specified column.  (Any row offset needed will
+ * be applied to the array pointer before it is passed to the IDCT code.)
+ * Note that the number of samples emitted by the IDCT routine is
+ * DCT_scaled_size * DCT_scaled_size.
+ */
+
+/* typedef inverse_DCT_method_ptr is declared in jpegint.h */
+
+/*
+ * Each IDCT routine has its own ideas about the best dct_table element type.
+ */
+
+typedef MULTIPLIER ISLOW_MULT_TYPE; /* short or int, whichever is faster */
+#if BITS_IN_JSAMPLE == 8
+typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
+#define IFAST_SCALE_BITS  2	/* fractional bits in scale factors */
+#else
+typedef long IFAST_MULT_TYPE;	/* need 32 bits for scaled quantizers */
+#define IFAST_SCALE_BITS  13	/* fractional bits in scale factors */
+#endif
+typedef FAST_FLOAT FLOAT_MULT_TYPE; /* preferred floating type */
+
+
+/*
+ * Each IDCT routine is responsible for range-limiting its results and
+ * converting them to unsigned form (0..MAXJSAMPLE).  The raw outputs could
+ * be quite far out of range if the input data is corrupt, so a bulletproof
+ * range-limiting step is required.  We use a mask-and-table-lookup method
+ * to do the combined operations quickly.  See the comments with
+ * prepare_range_limit_table (in jdmaster.c) for more info.
+ */
+
+#define IDCT_range_limit(cinfo)  ((cinfo)->sample_range_limit + CENTERJSAMPLE)
+
+#define RANGE_MASK  (MAXJSAMPLE * 4 + 3) /* 2 bits wider than legal samples */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_fdct_islow		jFDislow
+#define jpeg_fdct_ifast		jFDifast
+#define jpeg_fdct_float		jFDfloat
+#define jpeg_idct_islow		jRDislow
+#define jpeg_idct_ifast		jRDifast
+#define jpeg_idct_float		jRDfloat
+#define jpeg_idct_4x4		jRD4x4
+#define jpeg_idct_2x2		jRD2x2
+#define jpeg_idct_1x1		jRD1x1
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Extern declarations for the forward and inverse DCT routines. */
+
+EXTERN(void) jpeg_fdct_islow JPP((DCTELEM * data));
+EXTERN(void) jpeg_fdct_ifast JPP((DCTELEM * data));
+EXTERN(void) jpeg_fdct_float JPP((FAST_FLOAT * data));
+
+EXTERN(void) jpeg_idct_islow
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_ifast
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_float
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_4x4
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_2x2
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_1x1
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+
+
+/*
+ * Macros for handling fixed-point arithmetic; these are used by many
+ * but not all of the DCT/IDCT modules.
+ *
+ * All values are expected to be of type long.
+ * Fractional constants are scaled left by CONST_BITS bits.
+ * CONST_BITS is defined within each module using these macros,
+ * and may differ from one module to the next.
+ */
+
+#define ONE	((long) 1)
+#define CONST_SCALE (ONE << CONST_BITS)
+
+/* Convert a positive real constant to an integer scaled by CONST_SCALE.
+ * Caution: some C compilers fail to reduce "FIX(constant)" at compile time,
+ * thus causing a lot of useless floating-point operations at run time.
+ */
+
+#define FIX(x)	((long) ((x) * CONST_SCALE + 0.5))
+
+/* Descale and correctly round an long value that's scaled by N bits.
+ * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
+ * the fudge factor is correct for either sign of X.
+ */
+
+#define DESCALE(x,n)  RIGHT_SHIFT((x) + (ONE << ((n)-1)), n)
+
+/* Multiply an long variable by an long constant to yield an long result.
+ * This macro is used only when the two inputs will actually be no more than
+ * 16 bits wide, so that a 16x16->32 bit multiply can be used instead of a
+ * full 32x32 multiply.  This provides a useful speedup on many machines.
+ * Unfortunately there is no way to specify a 16x16->32 multiply portably
+ * in C, but some C compilers will do the right thing if you provide the
+ * correct combination of casts.
+ */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16C16(var,const)  (((short) (var)) * ((short) (const)))
+#endif
+#ifdef SHORTxLCONST_32		/* known to work with Microsoft C 6.0 */
+#define MULTIPLY16C16(var,const)  (((short) (var)) * ((long) (const)))
+#endif
+
+#ifndef MULTIPLY16C16		/* default definition */
+#define MULTIPLY16C16(var,const)  ((var) * (const))
+#endif
+
+/* Same except both inputs are variables. */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16V16(var1,var2)  (((short) (var1)) * ((short) (var2)))
+#endif
+
+#ifndef MULTIPLY16V16		/* default definition */
+#define MULTIPLY16V16(var1,var2)  ((var1) * (var2))
+#endif
diff --git a/ml/dlib/dlib/external/libjpeg/jddctmgr.cpp b/ml/dlib/dlib/external/libjpeg/jddctmgr.cpp
new file mode 100644
index 000000000..620da686d
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jddctmgr.cpp
@@ -0,0 +1,269 @@
+/*
+ * jddctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the inverse-DCT management logic.
+ * This code selects a particular IDCT implementation to be used,
+ * and it performs related housekeeping chores.  No code in this file
+ * is executed per IDCT step, only during output pass setup.
+ *
+ * Note that the IDCT routines are responsible for performing coefficient
+ * dequantization as well as the IDCT proper.  This module sets up the
+ * dequantization multiplier table needed by the IDCT routine.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+
+/*
+ * The decompressor input side (jdinput.c) saves away the appropriate
+ * quantization table for each component at the start of the first scan
+ * involving that component.  (This is necessary in order to correctly
+ * decode files that reuse Q-table slots.)
+ * When we are ready to make an output pass, the saved Q-table is converted
+ * to a multiplier table that will actually be used by the IDCT routine.
+ * The multiplier table contents are IDCT-method-dependent.  To support
+ * application changes in IDCT method between scans, we can remake the
+ * multiplier tables if necessary.
+ * In buffered-image mode, the first output pass may occur before any data
+ * has been seen for some components, and thus before their Q-tables have
+ * been saved away.  To handle this case, multiplier tables are preset
+ * to zeroes; the result of the IDCT will be a neutral gray level.
+ */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_inverse_dct pub;	/* public fields */
+
+  /* This array contains the IDCT method code that each multiplier table
+   * is currently set up for, or -1 if it's not yet set up.
+   * The actual multiplier tables are pointed to by dct_table in the
+   * per-component comp_info structures.
+   */
+  int cur_method[MAX_COMPONENTS];
+} my_idct_controller;
+
+typedef my_idct_controller * my_idct_ptr;
+
+
+/* Allocated multiplier tables: big enough for any supported variant */
+
+typedef union {
+  ISLOW_MULT_TYPE islow_array[DCTSIZE2];
+#ifdef DCT_IFAST_SUPPORTED
+  IFAST_MULT_TYPE ifast_array[DCTSIZE2];
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  FLOAT_MULT_TYPE float_array[DCTSIZE2];
+#endif
+} multiplier_table;
+
+
+/* The current scaled-IDCT routines require ISLOW-style multiplier tables,
+ * so be sure to compile that code if either ISLOW or SCALING is requested.
+ */
+#ifdef DCT_ISLOW_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#else
+#ifdef IDCT_SCALING_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#endif
+#endif
+
+
+/*
+ * Prepare for an output pass.
+ * Here we select the proper IDCT routine for each component and build
+ * a matching multiplier table.
+ */
+
+METHODDEF(void)
+start_pass (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct = (my_idct_ptr) cinfo->idct;
+  int ci, i;
+  jpeg_component_info *compptr;
+  int method = 0;
+  inverse_DCT_method_ptr method_ptr = NULL;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Select the proper IDCT routine for this component's scaling */
+    switch (compptr->DCT_scaled_size) {
+#ifdef IDCT_SCALING_SUPPORTED
+    case 1:
+      method_ptr = jpeg_idct_1x1;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+    case 2:
+      method_ptr = jpeg_idct_2x2;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+    case 4:
+      method_ptr = jpeg_idct_4x4;
+      method = JDCT_ISLOW;	/* jidctred uses islow-style table */
+      break;
+#endif
+    case DCTSIZE:
+      switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+      case JDCT_ISLOW:
+	method_ptr = jpeg_idct_islow;
+	method = JDCT_ISLOW;
+	break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+      case JDCT_IFAST:
+	method_ptr = jpeg_idct_ifast;
+	method = JDCT_IFAST;
+	break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+      case JDCT_FLOAT:
+	method_ptr = jpeg_idct_float;
+	method = JDCT_FLOAT;
+	break;
+#endif
+      default:
+	ERREXIT(cinfo, JERR_NOT_COMPILED);
+	break;
+      }
+      break;
+    default:
+      ERREXIT1(cinfo, JERR_BAD_DCTSIZE, compptr->DCT_scaled_size);
+      break;
+    }
+    idct->pub.inverse_DCT[ci] = method_ptr;
+    /* Create multiplier table from quant table.
+     * However, we can skip this if the component is uninteresting
+     * or if we already built the table.  Also, if no quant table
+     * has yet been saved for the component, we leave the
+     * multiplier table all-zero; we'll be reading zeroes from the
+     * coefficient controller's buffer anyway.
+     */
+    if (! compptr->component_needed || idct->cur_method[ci] == method)
+      continue;
+    qtbl = compptr->quant_table;
+    if (qtbl == NULL)		/* happens if no data yet for component */
+      continue;
+    idct->cur_method[ci] = method;
+    switch (method) {
+#ifdef PROVIDE_ISLOW_TABLES
+    case JDCT_ISLOW:
+      {
+	/* For LL&M IDCT method, multipliers are equal to raw quantization
+	 * coefficients, but are stored as ints to ensure access efficiency.
+	 */
+	ISLOW_MULT_TYPE * ismtbl = (ISLOW_MULT_TYPE *) compptr->dct_table;
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ismtbl[i] = (ISLOW_MULT_TYPE) qtbl->quantval[i];
+	}
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * For integer operation, the multiplier table is to be scaled by
+	 * IFAST_SCALE_BITS.
+	 */
+	IFAST_MULT_TYPE * ifmtbl = (IFAST_MULT_TYPE *) compptr->dct_table;
+#define CONST_BITS 14
+	static const short aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ifmtbl[i] = (IFAST_MULT_TYPE)
+	    DESCALE(MULTIPLY16V16((long) qtbl->quantval[i],
+				  (long) aanscales[i]),
+		    CONST_BITS-IFAST_SCALE_BITS);
+	}
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 */
+	FLOAT_MULT_TYPE * fmtbl = (FLOAT_MULT_TYPE *) compptr->dct_table;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fmtbl[i] = (FLOAT_MULT_TYPE)
+	      ((double) qtbl->quantval[i] *
+	       aanscalefactor[row] * aanscalefactor[col]);
+	    i++;
+	  }
+	}
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Initialize IDCT manager.
+ */
+
+GLOBAL(void)
+jinit_inverse_dct (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct;
+  int ci;
+  jpeg_component_info *compptr;
+
+  idct = (my_idct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_idct_controller));
+  cinfo->idct = (struct jpeg_inverse_dct *) idct;
+  idct->pub.start_pass = start_pass;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate and pre-zero a multiplier table for each component */
+    compptr->dct_table =
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(multiplier_table));
+    MEMZERO(compptr->dct_table, SIZEOF(multiplier_table));
+    /* Mark multiplier table not yet set up for any method */
+    idct->cur_method[ci] = -1;
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdhuff.cpp b/ml/dlib/dlib/external/libjpeg/jdhuff.cpp
new file mode 100644
index 000000000..26a2a36f2
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdhuff.cpp
@@ -0,0 +1,654 @@
+/*
+ * jdhuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy decoding routines.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"		/* Declarations shared with jdphuff.c */
+
+#ifdef __GNUC__
+#pragma GCC diagnostic ignored "-Wshift-negative-value"
+#endif
+
+/*
+ * Expanded entropy decoder object for Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;	/* Bit buffer at start of MCU */
+  savable_state saved;		/* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  d_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+  /* Precalculated info set up by start_pass for use in decode_mcu: */
+
+  /* Pointers to derived tables to be used for each block within an MCU */
+  d_derived_tbl * dc_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  d_derived_tbl * ac_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  /* Whether we care about the DC and AC coefficient values for each block */
+  int dc_needed[D_MAX_BLOCKS_IN_MCU];
+  int ac_needed[D_MAX_BLOCKS_IN_MCU];
+} huff_entropy_decoder;
+
+typedef huff_entropy_decoder * huff_entropy_ptr;
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, blkn, dctbl, actbl;
+  jpeg_component_info * compptr;
+
+  /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+   * This ought to be an error condition, but we make it a warning because
+   * there are some baseline files out there with all zeroes in these bytes.
+   */
+  if (cinfo->Ss != 0 || cinfo->Se != DCTSIZE2-1 ||
+      cinfo->Ah != 0 || cinfo->Al != 0)
+    WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    dctbl = compptr->dc_tbl_no;
+    actbl = compptr->ac_tbl_no;
+    /* Compute derived values for Huffman tables */
+    /* We may do this more than once for a table, but it's not expensive */
+    jpeg_make_d_derived_tbl(cinfo, TRUE, dctbl,
+			    & entropy->dc_derived_tbls[dctbl]);
+    jpeg_make_d_derived_tbl(cinfo, FALSE, actbl,
+			    & entropy->ac_derived_tbls[actbl]);
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Precalculate decoding info for each block in an MCU of this scan */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    /* Precalculate which table to use for each block */
+    entropy->dc_cur_tbls[blkn] = entropy->dc_derived_tbls[compptr->dc_tbl_no];
+    entropy->ac_cur_tbls[blkn] = entropy->ac_derived_tbls[compptr->ac_tbl_no];
+    /* Decide whether we really care about the coefficient values */
+    if (compptr->component_needed) {
+      entropy->dc_needed[blkn] = TRUE;
+      /* we don't need the ACs if producing a 1/8th-size image */
+      entropy->ac_needed[blkn] = (compptr->DCT_scaled_size > 1);
+    } else {
+      entropy->dc_needed[blkn] = entropy->ac_needed[blkn] = FALSE;
+    }
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ *
+ * Note this is also used by jdphuff.c.
+ */
+
+GLOBAL(void)
+jpeg_make_d_derived_tbl (j_decompress_ptr cinfo, int isDC, int tblno,
+			 d_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  d_derived_tbl *dtbl;
+  int p, i, l, si, numsymbols;
+  int lookbits, ctr;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (d_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(d_derived_tbl));
+  dtbl = *pdtbl;
+  dtbl->pub = htbl;		/* fill in back link */
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  numsymbols = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+  
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((long) code) >= (((long) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+
+  /* Figure F.15: generate decoding tables for bit-sequential decoding */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    if (htbl->bits[l]) {
+      /* valoffset[l] = huffval[] index of 1st symbol of code length l,
+       * minus the minimum code of length l
+       */
+      dtbl->valoffset[l] = (long) p - (long) huffcode[p];
+      p += htbl->bits[l];
+      dtbl->maxcode[l] = huffcode[p-1]; /* maximum code of length l */
+    } else {
+      dtbl->maxcode[l] = -1;	/* -1 if no codes of this length */
+    }
+  }
+  dtbl->maxcode[17] = 0xFFFFFL; /* ensures jpeg_huff_decode terminates */
+
+  /* Compute lookahead tables to speed up decoding.
+   * First we set all the table entries to 0, indicating "too long";
+   * then we iterate through the Huffman codes that are short enough and
+   * fill in all the entries that correspond to bit sequences starting
+   * with that code.
+   */
+
+  MEMZERO(dtbl->look_nbits, SIZEOF(dtbl->look_nbits));
+
+  p = 0;
+  for (l = 1; l <= HUFF_LOOKAHEAD; l++) {
+    for (i = 1; i <= (int) htbl->bits[l]; i++, p++) {
+      /* l = current code's length, p = its index in huffcode[] & huffval[]. */
+      /* Generate left-justified code followed by all possible bit sequences */
+      lookbits = huffcode[p] << (HUFF_LOOKAHEAD-l);
+      for (ctr = 1 << (HUFF_LOOKAHEAD-l); ctr > 0; ctr--) {
+	dtbl->look_nbits[lookbits] = l;
+	dtbl->look_sym[lookbits] = htbl->huffval[p];
+	lookbits++;
+      }
+    }
+  }
+
+  /* Validate symbols as being reasonable.
+   * For AC tables, we make no check, but accept all byte values 0..255.
+   * For DC tables, we require the symbols to be in range 0..15.
+   * (Tighter bounds could be applied depending on the data depth and mode,
+   * but this is sufficient to ensure safe decoding.)
+   */
+  if (isDC) {
+    for (i = 0; i < numsymbols; i++) {
+      int sym = htbl->huffval[i];
+      if (sym < 0 || sym > 15)
+	ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    }
+  }
+}
+
+
+/*
+ * Out-of-line code for bit fetching (shared with jdphuff.c).
+ * See jdhuff.h for info about usage.
+ * Note: current values of get_buffer and bits_left are passed as parameters,
+ * but are returned in the corresponding fields of the state struct.
+ *
+ * On most machines MIN_GET_BITS should be 25 to allow the full 32-bit width
+ * of get_buffer to be used.  (On machines with wider words, an even larger
+ * buffer could be used.)  However, on some machines 32-bit shifts are
+ * quite slow and take time proportional to the number of places shifted.
+ * (This is true with most PC compilers, for instance.)  In this case it may
+ * be a win to set MIN_GET_BITS to the minimum value of 15.  This reduces the
+ * average shift distance at the cost of more calls to jpeg_fill_bit_buffer.
+ */
+
+#ifdef SLOW_SHIFT_32
+#define MIN_GET_BITS  15	/* minimum allowable value */
+#else
+#define MIN_GET_BITS  (BIT_BUF_SIZE-7)
+#endif
+
+
+GLOBAL(int)
+jpeg_fill_bit_buffer (bitread_working_state * state,
+		      bit_buf_type get_buffer, register int bits_left,
+		      int nbits)
+/* Load up the bit buffer to a depth of at least nbits */
+{
+  /* Copy heavily used state fields into locals (hopefully registers) */
+  const JOCTET * next_input_byte = state->next_input_byte;
+  size_t bytes_in_buffer = state->bytes_in_buffer;
+  j_decompress_ptr cinfo = state->cinfo;
+
+  /* Attempt to load at least MIN_GET_BITS bits into get_buffer. */
+  /* (It is assumed that no request will be for more than that many bits.) */
+  /* We fail to do so only if we hit a marker or are forced to suspend. */
+
+  if (cinfo->unread_marker == 0) {	/* cannot advance past a marker */
+    while (bits_left < MIN_GET_BITS) {
+      int c;
+
+      /* Attempt to read a byte */
+      if (bytes_in_buffer == 0) {
+	if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	  return FALSE;
+	next_input_byte = cinfo->src->next_input_byte;
+	bytes_in_buffer = cinfo->src->bytes_in_buffer;
+      }
+      bytes_in_buffer--;
+      c = GETJOCTET(*next_input_byte++);
+
+      /* If it's 0xFF, check and discard stuffed zero byte */
+      if (c == 0xFF) {
+	/* Loop here to discard any padding FF's on terminating marker,
+	 * so that we can save a valid unread_marker value.  NOTE: we will
+	 * accept multiple FF's followed by a 0 as meaning a single FF data
+	 * byte.  This data pattern is not valid according to the standard.
+	 */
+	do {
+	  if (bytes_in_buffer == 0) {
+	    if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	      return FALSE;
+	    next_input_byte = cinfo->src->next_input_byte;
+	    bytes_in_buffer = cinfo->src->bytes_in_buffer;
+	  }
+	  bytes_in_buffer--;
+	  c = GETJOCTET(*next_input_byte++);
+	} while (c == 0xFF);
+
+	if (c == 0) {
+	  /* Found FF/00, which represents an FF data byte */
+	  c = 0xFF;
+	} else {
+	  /* Oops, it's actually a marker indicating end of compressed data.
+	   * Save the marker code for later use.
+	   * Fine point: it might appear that we should save the marker into
+	   * bitread working state, not straight into permanent state.  But
+	   * once we have hit a marker, we cannot need to suspend within the
+	   * current MCU, because we will read no more bytes from the data
+	   * source.  So it is OK to update permanent state right away.
+	   */
+	  cinfo->unread_marker = c;
+	  /* See if we need to insert some fake zero bits. */
+	  goto no_more_bytes;
+	}
+      }
+
+      /* OK, load c into get_buffer */
+      get_buffer = (get_buffer << 8) | c;
+      bits_left += 8;
+    } /* end while */
+  } else {
+  no_more_bytes:
+    /* We get here if we've read the marker that terminates the compressed
+     * data segment.  There should be enough bits in the buffer register
+     * to satisfy the request; if so, no problem.
+     */
+    if (nbits > bits_left) {
+      /* Uh-oh.  Report corrupted data to user and stuff zeroes into
+       * the data stream, so that we can produce some kind of image.
+       * We use a nonvolatile flag to ensure that only one warning message
+       * appears per data segment.
+       */
+      if (! cinfo->entropy->insufficient_data) {
+	WARNMS(cinfo, JWRN_HIT_MARKER);
+	cinfo->entropy->insufficient_data = TRUE;
+      }
+      /* Fill the buffer with zero bits */
+      get_buffer <<= MIN_GET_BITS - bits_left;
+      bits_left = MIN_GET_BITS;
+    }
+  }
+
+  /* Unload the local registers */
+  state->next_input_byte = next_input_byte;
+  state->bytes_in_buffer = bytes_in_buffer;
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  return TRUE;
+}
+
+
+/*
+ * Out-of-line code for Huffman code decoding.
+ * See jdhuff.h for info about usage.
+ */
+
+GLOBAL(int)
+jpeg_huff_decode (bitread_working_state * state,
+		  bit_buf_type get_buffer, register int bits_left,
+		  d_derived_tbl * htbl, int min_bits)
+{
+  int l = min_bits;
+  long code;
+
+  /* HUFF_DECODE has determined that the code is at least min_bits */
+  /* bits long, so fetch that many bits in one swoop. */
+
+  CHECK_BIT_BUFFER(*state, l, return -1);
+  code = GET_BITS(l);
+
+  /* Collect the rest of the Huffman code one bit at a time. */
+  /* This is per Figure F.16 in the JPEG spec. */
+
+  while (code > htbl->maxcode[l]) {
+    code <<= 1;
+    CHECK_BIT_BUFFER(*state, 1, return -1);
+    code |= GET_BITS(1);
+    l++;
+  }
+
+  /* Unload the local registers */
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  /* With garbage input we may reach the sentinel value l = 17. */
+
+  if (l > 16) {
+    WARNMS(state->cinfo, JWRN_HUFF_BAD_CODE);
+    return 0;			/* fake a zero as the safest result */
+  }
+
+  return htbl->pub->huffval[ (int) (code + htbl->valoffset[l]) ];
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#ifdef AVOID_TABLES
+
+#define HUFF_EXTEND(x,s)  ((x) < (1<<((s)-1)) ? (x) + (((-1)<<(s)) + 1) : (x))
+
+#else
+
+#define HUFF_EXTEND(x,s)  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] =   /* entry n is 2**(n-1) */
+  { 0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+    0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000 };
+
+static const int extend_offset[16] = /* entry n is (-1 << n) + 1 */
+  { 0, ((-1)<<1) + 1, ((-1)<<2) + 1, ((-1)<<3) + 1, ((-1)<<4) + 1,
+    ((-1)<<5) + 1, ((-1)<<6) + 1, ((-1)<<7) + 1, ((-1)<<8) + 1,
+    ((-1)<<9) + 1, ((-1)<<10) + 1, ((-1)<<11) + 1, ((-1)<<12) + 1,
+    ((-1)<<13) + 1, ((-1)<<14) + 1, ((-1)<<15) + 1 };
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(int)
+process_restart (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Decode and return one MCU's worth of Huffman-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA HAS BEEN ZEROED BY THE CALLER.
+ * (Wholesale zeroing is usually a little faster than retail...)
+ *
+ * Returns FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * this module, since we'll just re-assign them on the next call.)
+ */
+
+METHODDEF(int)
+decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int blkn;
+  BITREAD_STATE_VARS;
+  savable_state state;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      JBLOCKROW block = MCU_data[blkn];
+      d_derived_tbl * dctbl = entropy->dc_cur_tbls[blkn];
+      d_derived_tbl * actbl = entropy->ac_cur_tbls[blkn];
+      int s, k, r;
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      HUFF_DECODE(s, br_state, dctbl, return FALSE, label1);
+      if (s) {
+	CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	r = GET_BITS(s);
+	s = HUFF_EXTEND(r, s);
+      }
+
+      if (entropy->dc_needed[blkn]) {
+	/* Convert DC difference to actual value, update last_dc_val */
+	int ci = cinfo->MCU_membership[blkn];
+	s += state.last_dc_val[ci];
+	state.last_dc_val[ci] = s;
+	/* Output the DC coefficient (assumes jpeg_natural_order[0] = 0) */
+	(*block)[0] = (JCOEF) s;
+      }
+
+      if (entropy->ac_needed[blkn]) {
+
+	/* Section F.2.2.2: decode the AC coefficients */
+	/* Since zeroes are skipped, output area must be cleared beforehand */
+	for (k = 1; k < DCTSIZE2; k++) {
+	  HUFF_DECODE(s, br_state, actbl, return FALSE, label2);
+      
+	  r = s >> 4;
+	  s &= 15;
+      
+	  if (s) {
+	    k += r;
+	    CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	    r = GET_BITS(s);
+	    s = HUFF_EXTEND(r, s);
+	    /* Output coefficient in natural (dezigzagged) order.
+	     * Note: the extra entries in jpeg_natural_order[] will save us
+	     * if k >= DCTSIZE2, which could happen if the data is corrupted.
+	     */
+	    (*block)[jpeg_natural_order[k]] = (JCOEF) s;
+	  } else {
+	    if (r != 15)
+	      break;
+	    k += 15;
+	  }
+	}
+
+      } else {
+
+	/* Section F.2.2.2: decode the AC coefficients */
+	/* In this path we just discard the values */
+	for (k = 1; k < DCTSIZE2; k++) {
+	  HUFF_DECODE(s, br_state, actbl, return FALSE, label3);
+      
+	  r = s >> 4;
+	  s &= 15;
+      
+	  if (s) {
+	    k += r;
+	    CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	    DROP_BITS(s);
+	  } else {
+	    if (r != 15)
+	      break;
+	    k += 15;
+	  }
+	}
+
+      }
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * Module initialization routine for Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff_decoder;
+  entropy->pub.decode_mcu = decode_mcu;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdhuff.h b/ml/dlib/dlib/external/libjpeg/jdhuff.h
new file mode 100644
index 000000000..6a0e939af
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdhuff.h
@@ -0,0 +1,201 @@
+/*
+ * jdhuff.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains declarations for Huffman entropy decoding routines
+ * that are shared between the sequential decoder (jdhuff.c) and the
+ * progressive decoder (jdphuff.c).  No other modules need to see these.
+ */
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_make_d_derived_tbl	jMkDDerived
+#define jpeg_fill_bit_buffer	jFilBitBuf
+#define jpeg_huff_decode	jHufDecode
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Derived data constructed for each Huffman table */
+
+#define HUFF_LOOKAHEAD	8	/* # of bits of lookahead */
+
+typedef struct {
+  /* Basic tables: (element [0] of each array is unused) */
+  long maxcode[18];		/* largest code of length k (-1 if none) */
+  /* (maxcode[17] is a sentinel to ensure jpeg_huff_decode terminates) */
+  long valoffset[17];		/* huffval[] offset for codes of length k */
+  /* valoffset[k] = huffval[] index of 1st symbol of code length k, less
+   * the smallest code of length k; so given a code of length k, the
+   * corresponding symbol is huffval[code + valoffset[k]]
+   */
+
+  /* Link to public Huffman table (needed only in jpeg_huff_decode) */
+  JHUFF_TBL *pub;
+
+  /* Lookahead tables: indexed by the next HUFF_LOOKAHEAD bits of
+   * the input data stream.  If the next Huffman code is no more
+   * than HUFF_LOOKAHEAD bits long, we can obtain its length and
+   * the corresponding symbol directly from these tables.
+   */
+  int look_nbits[1<<HUFF_LOOKAHEAD]; /* # bits, or 0 if too long */
+  unsigned char look_sym[1<<HUFF_LOOKAHEAD]; /* symbol, or unused */
+} d_derived_tbl;
+
+/* Expand a Huffman table definition into the derived format */
+EXTERN(void) jpeg_make_d_derived_tbl
+	JPP((j_decompress_ptr cinfo, int isDC, int tblno,
+	     d_derived_tbl ** pdtbl));
+
+
+/*
+ * Fetching the next N bits from the input stream is a time-critical operation
+ * for the Huffman decoders.  We implement it with a combination of inline
+ * macros and out-of-line subroutines.  Note that N (the number of bits
+ * demanded at one time) never exceeds 15 for JPEG use.
+ *
+ * We read source bytes into get_buffer and dole out bits as needed.
+ * If get_buffer already contains enough bits, they are fetched in-line
+ * by the macros CHECK_BIT_BUFFER and GET_BITS.  When there aren't enough
+ * bits, jpeg_fill_bit_buffer is called; it will attempt to fill get_buffer
+ * as full as possible (not just to the number of bits needed; this
+ * prefetching reduces the overhead cost of calling jpeg_fill_bit_buffer).
+ * Note that jpeg_fill_bit_buffer may return FALSE to indicate suspension.
+ * On TRUE return, jpeg_fill_bit_buffer guarantees that get_buffer contains
+ * at least the requested number of bits --- dummy zeroes are inserted if
+ * necessary.
+ */
+
+typedef long bit_buf_type;	/* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  32	/* size of buffer in bits */
+
+/* If long is > 32 bits on your machine, and shifting/masking longs is
+ * reasonably fast, making bit_buf_type be long and setting BIT_BUF_SIZE
+ * appropriately should be a win.  Unfortunately we can't define the size
+ * with something like  #define BIT_BUF_SIZE (sizeof(bit_buf_type)*8)
+ * because not all machines measure sizeof in 8-bit bytes.
+ */
+
+typedef struct {		/* Bitreading state saved across MCUs */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+} bitread_perm_state;
+
+typedef struct {		/* Bitreading working state within an MCU */
+  /* Current data source location */
+  /* We need a copy, rather than munging the original, in case of suspension */
+  const JOCTET * next_input_byte; /* => next byte to read from source */
+  size_t bytes_in_buffer;	/* # of bytes remaining in source buffer */
+  /* Bit input buffer --- note these values are kept in variables,
+   * not in this struct, inside the inner loops.
+   */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+  /* Pointer needed by jpeg_fill_bit_buffer. */
+  j_decompress_ptr cinfo;	/* back link to decompress master record */
+} bitread_working_state;
+
+/* Macros to declare and load/save bitread local variables. */
+#define BITREAD_STATE_VARS  \
+	bit_buf_type get_buffer;  \
+	int bits_left;  \
+	bitread_working_state br_state
+
+#define BITREAD_LOAD_STATE(cinfop,permstate)  \
+	br_state.cinfo = cinfop; \
+	br_state.next_input_byte = cinfop->src->next_input_byte; \
+	br_state.bytes_in_buffer = cinfop->src->bytes_in_buffer; \
+	get_buffer = permstate.get_buffer; \
+	bits_left = permstate.bits_left;
+
+#define BITREAD_SAVE_STATE(cinfop,permstate)  \
+	cinfop->src->next_input_byte = br_state.next_input_byte; \
+	cinfop->src->bytes_in_buffer = br_state.bytes_in_buffer; \
+	permstate.get_buffer = get_buffer; \
+	permstate.bits_left = bits_left
+
+/*
+ * These macros provide the in-line portion of bit fetching.
+ * Use CHECK_BIT_BUFFER to ensure there are N bits in get_buffer
+ * before using GET_BITS, PEEK_BITS, or DROP_BITS.
+ * The variables get_buffer and bits_left are assumed to be locals,
+ * but the state struct might not be (jpeg_huff_decode needs this).
+ *	CHECK_BIT_BUFFER(state,n,action);
+ *		Ensure there are N bits in get_buffer; if suspend, take action.
+ *      val = GET_BITS(n);
+ *		Fetch next N bits.
+ *      val = PEEK_BITS(n);
+ *		Fetch next N bits without removing them from the buffer.
+ *	DROP_BITS(n);
+ *		Discard next N bits.
+ * The value N should be a simple variable, not an expression, because it
+ * is evaluated multiple times.
+ */
+
+#define CHECK_BIT_BUFFER(state,nbits,action) \
+	{ if (bits_left < (nbits)) {  \
+	    if (! jpeg_fill_bit_buffer(&(state),get_buffer,bits_left,nbits))  \
+	      { action; }  \
+	    get_buffer = (state).get_buffer; bits_left = (state).bits_left; } }
+
+#define GET_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -= (nbits)))) & ((1<<(nbits))-1))
+
+#define PEEK_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -  (nbits)))) & ((1<<(nbits))-1))
+
+#define DROP_BITS(nbits) \
+	(bits_left -= (nbits))
+
+/* Load up the bit buffer to a depth of at least nbits */
+EXTERN(int) jpeg_fill_bit_buffer
+	JPP((bitread_working_state * state, bit_buf_type get_buffer,
+	     int bits_left, int nbits));
+
+
+/*
+ * Code for extracting next Huffman-coded symbol from input bit stream.
+ * Again, this is time-critical and we make the main paths be macros.
+ *
+ * We use a lookahead table to process codes of up to HUFF_LOOKAHEAD bits
+ * without looping.  Usually, more than 95% of the Huffman codes will be 8
+ * or fewer bits long.  The few overlength codes are handled with a loop,
+ * which need not be inline code.
+ *
+ * Notes about the HUFF_DECODE macro:
+ * 1. Near the end of the data segment, we may fail to get enough bits
+ *    for a lookahead.  In that case, we do it the hard way.
+ * 2. If the lookahead table contains no entry, the next code must be
+ *    more than HUFF_LOOKAHEAD bits long.
+ * 3. jpeg_huff_decode returns -1 if forced to suspend.
+ */
+
+#define HUFF_DECODE(result,state,htbl,failaction,slowlabel) \
+{ int nb, look; \
+  if (bits_left < HUFF_LOOKAHEAD) { \
+    if (! jpeg_fill_bit_buffer(&state,get_buffer,bits_left, 0)) {failaction;} \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+    if (bits_left < HUFF_LOOKAHEAD) { \
+      nb = 1; goto slowlabel; \
+    } \
+  } \
+  look = PEEK_BITS(HUFF_LOOKAHEAD); \
+  if ((nb = htbl->look_nbits[look]) != 0) { \
+    DROP_BITS(nb); \
+    result = htbl->look_sym[look]; \
+  } else { \
+    nb = HUFF_LOOKAHEAD+1; \
+slowlabel: \
+    if ((result=jpeg_huff_decode(&state,get_buffer,bits_left,htbl,nb)) < 0) \
+	{ failaction; } \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+  } \
+}
+
+/* Out-of-line case for Huffman code fetching */
+EXTERN(int) jpeg_huff_decode
+	JPP((bitread_working_state * state, bit_buf_type get_buffer,
+	     int bits_left, d_derived_tbl * htbl, int min_bits));
diff --git a/ml/dlib/dlib/external/libjpeg/jdinput.cpp b/ml/dlib/dlib/external/libjpeg/jdinput.cpp
new file mode 100644
index 000000000..42f79977d
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdinput.cpp
@@ -0,0 +1,381 @@
+/*
+ * jdinput.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input control logic for the JPEG decompressor.
+ * These routines are concerned with controlling the decompressor's input
+ * processing (marker reading and coefficient decoding).  The actual input
+ * reading is done in jdmarker.c, jdhuff.c, and jdphuff.c.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_input_controller pub; /* public fields */
+
+  int inheaders;		/* TRUE until first SOS is reached */
+} my_input_controller;
+
+typedef my_input_controller * my_inputctl_ptr;
+
+
+/* Forward declarations */
+METHODDEF(int) consume_markers JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Routines to calculate various quantities related to the size of the image.
+ */
+
+LOCAL(void)
+initial_setup (j_decompress_ptr cinfo)
+/* Called once, when first SOS marker is reached */
+{
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->image_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->image_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+  /* We initialize DCT_scaled_size and min_DCT_scaled_size to DCTSIZE.
+   * In the full decompressor, this will be overridden by jdmaster.c;
+   * but in the transcoder, jdmaster.c is not used, so we must do it here.
+   */
+  cinfo->min_DCT_scaled_size = DCTSIZE;
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->DCT_scaled_size = DCTSIZE;
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+    /* downsampled_width and downsampled_height will also be overridden by
+     * jdmaster.c if we are doing full decompression.  The transcoder library
+     * doesn't use these values, but the calling application might.
+     */
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) cinfo->max_v_samp_factor);
+    /* Mark component needed, until color conversion says otherwise */
+    compptr->component_needed = TRUE;
+    /* Mark no quantization table yet saved for component */
+    compptr->quant_table = NULL;
+  }
+
+  /* Compute number of fully interleaved MCU rows. */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->image_height,
+		  (long) (cinfo->max_v_samp_factor*DCTSIZE));
+
+  /* Decide whether file contains multiple scans */
+  if (cinfo->comps_in_scan < cinfo->num_components || cinfo->progressive_mode)
+    cinfo->inputctl->has_multiple_scans = TRUE;
+  else
+    cinfo->inputctl->has_multiple_scans = FALSE;
+}
+
+
+LOCAL(void)
+per_scan_setup (j_decompress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] were set from SOS marker */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = compptr->DCT_scaled_size;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width,
+		    (long) (cinfo->max_h_samp_factor*DCTSIZE));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height,
+		    (long) (cinfo->max_v_samp_factor*DCTSIZE));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * compptr->DCT_scaled_size;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > D_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+}
+
+
+/*
+ * Save away a copy of the Q-table referenced by each component present
+ * in the current scan, unless already saved during a prior scan.
+ *
+ * In a multiple-scan JPEG file, the encoder could assign different components
+ * the same Q-table slot number, but change table definitions between scans
+ * so that each component uses a different Q-table.  (The IJG encoder is not
+ * currently capable of doing this, but other encoders might.)  Since we want
+ * to be able to dequantize all the components at the end of the file, this
+ * means that we have to save away the table actually used for each component.
+ * We do this by copying the table at the start of the first scan containing
+ * the component.
+ * The JPEG spec prohibits the encoder from changing the contents of a Q-table
+ * slot between scans of a component using that slot.  If the encoder does so
+ * anyway, this decoder will simply use the Q-table values that were current
+ * at the start of the first scan for the component.
+ *
+ * The decompressor output side looks only at the saved quant tables,
+ * not at the current Q-table slots.
+ */
+
+LOCAL(void)
+latch_quant_tables (j_decompress_ptr cinfo)
+{
+  int ci, qtblno;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* No work if we already saved Q-table for this component */
+    if (compptr->quant_table != NULL)
+      continue;
+    /* Make sure specified quantization table is present */
+    qtblno = compptr->quant_tbl_no;
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    /* OK, save away the quantization table */
+    qtbl = (JQUANT_TBL *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(JQUANT_TBL));
+    MEMCOPY(qtbl, cinfo->quant_tbl_ptrs[qtblno], SIZEOF(JQUANT_TBL));
+    compptr->quant_table = qtbl;
+  }
+}
+
+
+/*
+ * Initialize the input modules to read a scan of compressed data.
+ * The first call to this is done by jdmaster.c after initializing
+ * the entire decompressor (during jpeg_start_decompress).
+ * Subsequent calls come from consume_markers, below.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  per_scan_setup(cinfo);
+  latch_quant_tables(cinfo);
+  (*cinfo->entropy->start_pass) (cinfo);
+  (*cinfo->coef->start_input_pass) (cinfo);
+  cinfo->inputctl->consume_input = cinfo->coef->consume_data;
+}
+
+
+/*
+ * Finish up after inputting a compressed-data scan.
+ * This is called by the coefficient controller after it's read all
+ * the expected data of the scan.
+ */
+
+METHODDEF(void)
+finish_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->inputctl->consume_input = consume_markers;
+}
+
+
+/*
+ * Read JPEG markers before, between, or after compressed-data scans.
+ * Change state as necessary when a new scan is reached.
+ * Return value is JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ *
+ * The consume_input method pointer points either here or to the
+ * coefficient controller's consume_data routine, depending on whether
+ * we are reading a compressed data segment or inter-segment markers.
+ */
+
+METHODDEF(int)
+consume_markers (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+  int val;
+
+  if (inputctl->pub.eoi_reached) /* After hitting EOI, read no further */
+    return JPEG_REACHED_EOI;
+
+  val = (*cinfo->marker->read_markers) (cinfo);
+
+  switch (val) {
+  case JPEG_REACHED_SOS:	/* Found SOS */
+    if (inputctl->inheaders) {	/* 1st SOS */
+      initial_setup(cinfo);
+      inputctl->inheaders = FALSE;
+      /* Note: start_input_pass must be called by jdmaster.c
+       * before any more input can be consumed.  jdapimin.c is
+       * responsible for enforcing this sequencing.
+       */
+    } else {			/* 2nd or later SOS marker */
+      if (! inputctl->pub.has_multiple_scans)
+	ERREXIT(cinfo, JERR_EOI_EXPECTED); /* Oops, I wasn't expecting this! */
+      start_input_pass(cinfo);
+    }
+    break;
+  case JPEG_REACHED_EOI:	/* Found EOI */
+    inputctl->pub.eoi_reached = TRUE;
+    if (inputctl->inheaders) {	/* Tables-only datastream, apparently */
+      if (cinfo->marker->saw_SOF)
+	ERREXIT(cinfo, JERR_SOF_NO_SOS);
+    } else {
+      /* Prevent infinite loop in coef ctlr's decompress_data routine
+       * if user set output_scan_number larger than number of scans.
+       */
+      if (cinfo->output_scan_number > cinfo->input_scan_number)
+	cinfo->output_scan_number = cinfo->input_scan_number;
+    }
+    break;
+  case JPEG_SUSPENDED:
+    break;
+  }
+
+  return val;
+}
+
+
+/*
+ * Reset state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+  /* Reset other modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->marker->reset_marker_reader) (cinfo);
+  /* Reset progression state -- would be cleaner if entropy decoder did this */
+  cinfo->coef_bits = NULL;
+}
+
+
+/*
+ * Initialize the input controller module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl;
+
+  /* Create subobject in permanent pool */
+  inputctl = (my_inputctl_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_input_controller));
+  cinfo->inputctl = (struct jpeg_input_controller *) inputctl;
+  /* Initialize method pointers */
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.reset_input_controller = reset_input_controller;
+  inputctl->pub.start_input_pass = start_input_pass;
+  inputctl->pub.finish_input_pass = finish_input_pass;
+  /* Initialize state: can't use reset_input_controller since we don't
+   * want to try to reset other modules yet.
+   */
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = TRUE;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdmainct.cpp b/ml/dlib/dlib/external/libjpeg/jdmainct.cpp
new file mode 100644
index 000000000..bc2c378aa
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdmainct.cpp
@@ -0,0 +1,512 @@
+/*
+ * jdmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for decompression.
+ * The main buffer lies between the JPEG decompressor proper and the
+ * post-processor; it holds downsampled data in the JPEG colorspace.
+ *
+ * Note that this code is bypassed in raw-data mode, since the application
+ * supplies the equivalent of the main buffer in that case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * In the current system design, the main buffer need never be a full-image
+ * buffer; any full-height buffers will be found inside the coefficient or
+ * postprocessing controllers.  Nonetheless, the main controller is not
+ * trivial.  Its responsibility is to provide context rows for upsampling/
+ * rescaling, and doing this in an efficient fashion is a bit tricky.
+ *
+ * Postprocessor input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  (We require DCT_scaled_size values to be
+ * chosen such that these numbers are integers.  In practice DCT_scaled_size
+ * values will likely be powers of two, so we actually have the stronger
+ * condition that DCT_scaled_size / min_DCT_scaled_size is an integer.)
+ * Upsampling will typically produce max_v_samp_factor pixel rows from each
+ * row group (times any additional scale factor that the upsampler is
+ * applying).
+ *
+ * The coefficient controller will deliver data to us one iMCU row at a time;
+ * each iMCU row contains v_samp_factor * DCT_scaled_size sample rows, or
+ * exactly min_DCT_scaled_size row groups.  (This amount of data corresponds
+ * to one row of MCUs when the image is fully interleaved.)  Note that the
+ * number of sample rows varies across components, but the number of row
+ * groups does not.  Some garbage sample rows may be included in the last iMCU
+ * row at the bottom of the image.
+ *
+ * Depending on the vertical scaling algorithm used, the upsampler may need
+ * access to the sample row(s) above and below its current input row group.
+ * The upsampler is required to set need_context_rows TRUE at global selection
+ * time if so.  When need_context_rows is FALSE, this controller can simply
+ * obtain one iMCU row at a time from the coefficient controller and dole it
+ * out as row groups to the postprocessor.
+ *
+ * When need_context_rows is TRUE, this controller guarantees that the buffer
+ * passed to postprocessing contains at least one row group's worth of samples
+ * above and below the row group(s) being processed.  Note that the context
+ * rows "above" the first passed row group appear at negative row offsets in
+ * the passed buffer.  At the top and bottom of the image, the required
+ * context rows are manufactured by duplicating the first or last real sample
+ * row; this avoids having special cases in the upsampling inner loops.
+ *
+ * The amount of context is fixed at one row group just because that's a
+ * convenient number for this controller to work with.  The existing
+ * upsamplers really only need one sample row of context.  An upsampler
+ * supporting arbitrary output rescaling might wish for more than one row
+ * group of context when shrinking the image; tough, we don't handle that.
+ * (This is justified by the assumption that downsizing will be handled mostly
+ * by adjusting the DCT_scaled_size values, so that the actual scale factor at
+ * the upsample step needn't be much less than one.)
+ *
+ * To provide the desired context, we have to retain the last two row groups
+ * of one iMCU row while reading in the next iMCU row.  (The last row group
+ * can't be processed until we have another row group for its below-context,
+ * and so we have to save the next-to-last group too for its above-context.)
+ * We could do this most simply by copying data around in our buffer, but
+ * that'd be very slow.  We can avoid copying any data by creating a rather
+ * strange pointer structure.  Here's how it works.  We allocate a workspace
+ * consisting of M+2 row groups (where M = min_DCT_scaled_size is the number
+ * of row groups per iMCU row).  We create two sets of redundant pointers to
+ * the workspace.  Labeling the physical row groups 0 to M+1, the synthesized
+ * pointer lists look like this:
+ *                   M+1                          M-1
+ * master pointer --> 0         master pointer --> 0
+ *                    1                            1
+ *                   ...                          ...
+ *                   M-3                          M-3
+ *                   M-2                           M
+ *                   M-1                          M+1
+ *                    M                           M-2
+ *                   M+1                          M-1
+ *                    0                            0
+ * We read alternate iMCU rows using each master pointer; thus the last two
+ * row groups of the previous iMCU row remain un-overwritten in the workspace.
+ * The pointer lists are set up so that the required context rows appear to
+ * be adjacent to the proper places when we pass the pointer lists to the
+ * upsampler.
+ *
+ * The above pictures describe the normal state of the pointer lists.
+ * At top and bottom of the image, we diddle the pointer lists to duplicate
+ * the first or last sample row as necessary (this is cheaper than copying
+ * sample rows around).
+ *
+ * This scheme breaks down if M < 2, ie, min_DCT_scaled_size is 1.  In that
+ * situation each iMCU row provides only one row group so the buffering logic
+ * must be different (eg, we must read two iMCU rows before we can emit the
+ * first row group).  For now, we simply do not support providing context
+ * rows when min_DCT_scaled_size is 1.  That combination seems unlikely to
+ * be worth providing --- if someone wants a 1/8th-size preview, they probably
+ * want it quick and dirty, so a context-free upsampler is sufficient.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_main_controller pub; /* public fields */
+
+  /* Pointer to allocated workspace (M or M+2 row groups). */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+  int buffer_full;		/* Have we gotten an iMCU row from decoder? */
+  JDIMENSION rowgroup_ctr;	/* counts row groups output to postprocessor */
+
+  /* Remaining fields are only used in the context case. */
+
+  /* These are the master pointers to the funny-order pointer lists. */
+  JSAMPIMAGE xbuffer[2];	/* pointers to weird pointer lists */
+
+  int whichptr;			/* indicates which pointer set is now in use */
+  int context_state;		/* process_data state machine status */
+  JDIMENSION rowgroups_avail;	/* row groups available to postprocessor */
+  JDIMENSION iMCU_row_ctr;	/* counts iMCU rows to detect image top/bot */
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+/* context_state values: */
+#define CTX_PREPARE_FOR_IMCU	0	/* need to prepare for MCU row */
+#define CTX_PROCESS_IMCU	1	/* feeding iMCU to postprocessor */
+#define CTX_POSTPONED_ROW	2	/* feeding postponed row group */
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+METHODDEF(void) process_data_context_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) process_data_crank_post
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#endif
+
+
+LOCAL(void)
+alloc_funny_pointers (j_decompress_ptr cinfo)
+/* Allocate space for the funny pointer lists.
+ * This is done only once, not once per pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, rgroup;
+  int M = cinfo->min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  /* Get top-level space for component array pointers.
+   * We alloc both arrays with one call to save a few cycles.
+   */
+  main->xbuffer[0] = (JSAMPIMAGE)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				cinfo->num_components * 2 * SIZEOF(JSAMPARRAY));
+  main->xbuffer[1] = main->xbuffer[0] + cinfo->num_components;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_scaled_size) /
+      cinfo->min_DCT_scaled_size; /* height of a row group of component */
+    /* Get space for pointer lists --- M+4 row groups in each list.
+     * We alloc both pointer lists with one call to save a few cycles.
+     */
+    xbuf = (JSAMPARRAY)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  2 * (rgroup * (M + 4)) * SIZEOF(JSAMPROW));
+    xbuf += rgroup;		/* want one row group at negative offsets */
+    main->xbuffer[0][ci] = xbuf;
+    xbuf += rgroup * (M + 4);
+    main->xbuffer[1][ci] = xbuf;
+  }
+}
+
+
+LOCAL(void)
+make_funny_pointers (j_decompress_ptr cinfo)
+/* Create the funny pointer lists discussed in the comments above.
+ * The actual workspace is already allocated (in main->buffer),
+ * and the space for the pointer lists is allocated too.
+ * This routine just fills in the curiously ordered lists.
+ * This will be repeated at the beginning of each pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY buf, xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_scaled_size) /
+      cinfo->min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    /* First copy the workspace pointers as-is */
+    buf = main->buffer[ci];
+    for (i = 0; i < rgroup * (M + 2); i++) {
+      xbuf0[i] = xbuf1[i] = buf[i];
+    }
+    /* In the second list, put the last four row groups in swapped order */
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf1[rgroup*(M-2) + i] = buf[rgroup*M + i];
+      xbuf1[rgroup*M + i] = buf[rgroup*(M-2) + i];
+    }
+    /* The wraparound pointers at top and bottom will be filled later
+     * (see set_wraparound_pointers, below).  Initially we want the "above"
+     * pointers to duplicate the first actual data line.  This only needs
+     * to happen in xbuffer[0].
+     */
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[0];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_wraparound_pointers (j_decompress_ptr cinfo)
+/* Set up the "wraparound" pointers at top and bottom of the pointer lists.
+ * This changes the pointer list state from top-of-image to the normal state.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->min_DCT_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_scaled_size) /
+      cinfo->min_DCT_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[rgroup*(M+1) + i];
+      xbuf1[i - rgroup] = xbuf1[rgroup*(M+1) + i];
+      xbuf0[rgroup*(M+2) + i] = xbuf0[i];
+      xbuf1[rgroup*(M+2) + i] = xbuf1[i];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_bottom_pointers (j_decompress_ptr cinfo)
+/* Change the pointer lists to duplicate the last sample row at the bottom
+ * of the image.  whichptr indicates which xbuffer holds the final iMCU row.
+ * Also sets rowgroups_avail to indicate number of nondummy row groups in row.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup, iMCUheight, rows_left;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Count sample rows in one iMCU row and in one row group */
+    iMCUheight = compptr->v_samp_factor * compptr->DCT_scaled_size;
+    rgroup = iMCUheight / cinfo->min_DCT_scaled_size;
+    /* Count nondummy sample rows remaining for this component */
+    rows_left = (int) (compptr->downsampled_height % (JDIMENSION) iMCUheight);
+    if (rows_left == 0) rows_left = iMCUheight;
+    /* Count nondummy row groups.  Should get same answer for each component,
+     * so we need only do it once.
+     */
+    if (ci == 0) {
+      main->rowgroups_avail = (JDIMENSION) ((rows_left-1) / rgroup + 1);
+    }
+    /* Duplicate the last real sample row rgroup*2 times; this pads out the
+     * last partial rowgroup and ensures at least one full rowgroup of context.
+     */
+    xbuf = main->xbuffer[main->whichptr][ci];
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf[rows_left + i] = xbuf[rows_left-1];
+    }
+  }
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->upsample->need_context_rows) {
+      main->pub.process_data = process_data_context_main;
+      make_funny_pointers(cinfo); /* Create the xbuffer[] lists */
+      main->whichptr = 0;	/* Read first iMCU row into xbuffer[0] */
+      main->context_state = CTX_PREPARE_FOR_IMCU;
+      main->iMCU_row_ctr = 0;
+    } else {
+      /* Simple case with no context needed */
+      main->pub.process_data = process_data_simple_main;
+    }
+    main->buffer_full = FALSE;	/* Mark buffer empty */
+    main->rowgroup_ctr = 0;
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_CRANK_DEST:
+    /* For last pass of 2-pass quantization, just crank the postprocessor */
+    main->pub.process_data = process_data_crank_post;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the simple case where no context is required.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_decompress_ptr cinfo,
+			  JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			  JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  JDIMENSION rowgroups_avail;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo, main->buffer))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+  }
+
+  /* There are always min_DCT_scaled_size row groups in an iMCU row. */
+  rowgroups_avail = (JDIMENSION) cinfo->min_DCT_scaled_size;
+  /* Note: at the bottom of the image, we may pass extra garbage row groups
+   * to the postprocessor.  The postprocessor has to check for bottom
+   * of image anyway (at row resolution), so no point in us doing it too.
+   */
+
+  /* Feed the postprocessor */
+  (*cinfo->post->post_process_data) (cinfo, main->buffer,
+				     &main->rowgroup_ctr, rowgroups_avail,
+				     output_buf, out_row_ctr, out_rows_avail);
+
+  /* Has postprocessor consumed all the data yet? If so, mark buffer empty */
+  if (main->rowgroup_ctr >= rowgroups_avail) {
+    main->buffer_full = FALSE;
+    main->rowgroup_ctr = 0;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the case where context rows must be provided.
+ */
+
+METHODDEF(void)
+process_data_context_main (j_decompress_ptr cinfo,
+			   JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			   JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo,
+					   main->xbuffer[main->whichptr]))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+    main->iMCU_row_ctr++;	/* count rows received */
+  }
+
+  /* Postprocessor typically will not swallow all the input data it is handed
+   * in one call (due to filling the output buffer first).  Must be prepared
+   * to exit and restart.  This switch lets us keep track of how far we got.
+   * Note that each case falls through to the next on successful completion.
+   */
+  switch (main->context_state) {
+  case CTX_POSTPONED_ROW:
+    /* Call postprocessor using previously set pointers for postponed row */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    main->context_state = CTX_PREPARE_FOR_IMCU;
+    if (*out_row_ctr >= out_rows_avail)
+      return;			/* Postprocessor exactly filled output buf */
+    /*FALLTHROUGH*/
+  case CTX_PREPARE_FOR_IMCU:
+    /* Prepare to process first M-1 row groups of this iMCU row */
+    main->rowgroup_ctr = 0;
+    main->rowgroups_avail = (JDIMENSION) (cinfo->min_DCT_scaled_size - 1);
+    /* Check for bottom of image: if so, tweak pointers to "duplicate"
+     * the last sample row, and adjust rowgroups_avail to ignore padding rows.
+     */
+    if (main->iMCU_row_ctr == cinfo->total_iMCU_rows)
+      set_bottom_pointers(cinfo);
+    main->context_state = CTX_PROCESS_IMCU;
+    /*FALLTHROUGH*/
+  case CTX_PROCESS_IMCU:
+    /* Call postprocessor using previously set pointers */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    /* After the first iMCU, change wraparound pointers to normal state */
+    if (main->iMCU_row_ctr == 1)
+      set_wraparound_pointers(cinfo);
+    /* Prepare to load new iMCU row using other xbuffer list */
+    main->whichptr ^= 1;	/* 0=>1 or 1=>0 */
+    main->buffer_full = FALSE;
+    /* Still need to process last row group of this iMCU row, */
+    /* which is saved at index M+1 of the other xbuffer */
+    main->rowgroup_ctr = (JDIMENSION) (cinfo->min_DCT_scaled_size + 1);
+    main->rowgroups_avail = (JDIMENSION) (cinfo->min_DCT_scaled_size + 2);
+    main->context_state = CTX_POSTPONED_ROW;
+  }
+}
+
+
+/*
+ * Process some data.
+ * Final pass of two-pass quantization: just call the postprocessor.
+ * Source data will be the postprocessor controller's internal buffer.
+ */
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+METHODDEF(void)
+process_data_crank_post (j_decompress_ptr cinfo,
+			 JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			 JDIMENSION out_rows_avail)
+{
+  (*cinfo->post->post_process_data) (cinfo, (JSAMPIMAGE) NULL,
+				     (JDIMENSION *) NULL, (JDIMENSION) 0,
+				     output_buf, out_row_ctr, out_rows_avail);
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_main_controller (j_decompress_ptr cinfo, int need_full_buffer)
+{
+  my_main_ptr main;
+  int ci, rgroup, ngroups;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_d_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  if (need_full_buffer)		/* shouldn't happen */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Allocate the workspace.
+   * ngroups is the number of row groups we need.
+   */
+  if (cinfo->upsample->need_context_rows) {
+    if (cinfo->min_DCT_scaled_size < 2) /* unsupported, see comments above */
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    alloc_funny_pointers(cinfo); /* Alloc space for xbuffer[] lists */
+    ngroups = cinfo->min_DCT_scaled_size + 2;
+  } else {
+    ngroups = cinfo->min_DCT_scaled_size;
+  }
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_scaled_size) /
+      cinfo->min_DCT_scaled_size; /* height of a row group of component */
+    main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 compptr->width_in_blocks * compptr->DCT_scaled_size,
+			 (JDIMENSION) (rgroup * ngroups));
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdmarker.cpp b/ml/dlib/dlib/external/libjpeg/jdmarker.cpp
new file mode 100644
index 000000000..c8c9f8e1c
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdmarker.cpp
@@ -0,0 +1,1360 @@
+/*
+ * jdmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to decode JPEG datastream markers.
+ * Most of the complexity arises from our desire to support input
+ * suspension: if not all of the data for a marker is available,
+ * we must exit back to the application.  On resumption, we reprocess
+ * the marker.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_reader pub; /* public fields */
+
+  /* Application-overridable marker processing methods */
+  jpeg_marker_parser_method process_COM;
+  jpeg_marker_parser_method process_APPn[16];
+
+  /* Limit on marker data length to save for each marker type */
+  unsigned int length_limit_COM;
+  unsigned int length_limit_APPn[16];
+
+  /* Status of COM/APPn marker saving */
+  jpeg_saved_marker_ptr cur_marker;	/* NULL if not processing a marker */
+  unsigned int bytes_read;		/* data bytes read so far in marker */
+  /* Note: cur_marker is not linked into marker_list until it's all read. */
+} my_marker_reader;
+
+typedef my_marker_reader * my_marker_ptr;
+
+
+/*
+ * Macros for fetching data from the data source module.
+ *
+ * At all times, cinfo->src->next_input_byte and ->bytes_in_buffer reflect
+ * the current restart point; we update them only when we have reached a
+ * suitable place to restart if a suspension occurs.
+ */
+
+/* Declare and initialize local copies of input pointer/count */
+#define INPUT_VARS(cinfo)  \
+	struct jpeg_source_mgr * datasrc = (cinfo)->src;  \
+	const JOCTET * next_input_byte = datasrc->next_input_byte;  \
+	size_t bytes_in_buffer = datasrc->bytes_in_buffer
+
+/* Unload the local copies --- do this only at a restart boundary */
+#define INPUT_SYNC(cinfo)  \
+	( datasrc->next_input_byte = next_input_byte,  \
+	  datasrc->bytes_in_buffer = bytes_in_buffer )
+
+/* Reload the local copies --- used only in MAKE_BYTE_AVAIL */
+#define INPUT_RELOAD(cinfo)  \
+	( next_input_byte = datasrc->next_input_byte,  \
+	  bytes_in_buffer = datasrc->bytes_in_buffer )
+
+/* Internal macro for INPUT_BYTE and INPUT_2BYTES: make a byte available.
+ * Note we do *not* do INPUT_SYNC before calling fill_input_buffer,
+ * but we must reload the local copies after a successful fill.
+ */
+#define MAKE_BYTE_AVAIL(cinfo,action)  \
+	if (bytes_in_buffer == 0) {  \
+	  if (! (*datasrc->fill_input_buffer) (cinfo))  \
+	    { action; }  \
+	  INPUT_RELOAD(cinfo);  \
+	}
+
+/* Read a byte into variable V.
+ * If must suspend, take the specified action (typically "return FALSE").
+ */
+#define INPUT_BYTE(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = GETJOCTET(*next_input_byte++); )
+
+/* As above, but read two bytes interpreted as an unsigned 16-bit integer.
+ * V should be declared unsigned int or perhaps long.
+ */
+#define INPUT_2BYTES(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = ((unsigned int) GETJOCTET(*next_input_byte++)) << 8; \
+		  MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V += GETJOCTET(*next_input_byte++); )
+
+
+/*
+ * Routines to process JPEG markers.
+ *
+ * Entry condition: JPEG marker itself has been read and its code saved
+ *   in cinfo->unread_marker; input restart point is just after the marker.
+ *
+ * Exit: if return TRUE, have read and processed any parameters, and have
+ *   updated the restart point to point after the parameters.
+ *   If return FALSE, was forced to suspend before reaching end of
+ *   marker parameters; restart point has not been moved.  Same routine
+ *   will be called again after application supplies more input data.
+ *
+ * This approach to suspension assumes that all of a marker's parameters
+ * can fit into a single input bufferload.  This should hold for "normal"
+ * markers.  Some COM/APPn markers might have large parameter segments
+ * that might not fit.  If we are simply dropping such a marker, we use
+ * skip_input_data to get past it, and thereby put the problem on the
+ * source manager's shoulders.  If we are saving the marker's contents
+ * into memory, we use a slightly different convention: when forced to
+ * suspend, the marker processor updates the restart point to the end of
+ * what it's consumed (ie, the end of the buffer) before returning FALSE.
+ * On resumption, cinfo->unread_marker still contains the marker code,
+ * but the data source will point to the next chunk of marker data.
+ * The marker processor must retain internal state to deal with this.
+ *
+ * Note that we don't bother to avoid duplicate trace messages if a
+ * suspension occurs within marker parameters.  Other side effects
+ * require more care.
+ */
+
+
+LOCAL(int)
+get_soi (j_decompress_ptr cinfo)
+/* Process an SOI marker */
+{
+  int i;
+  
+  TRACEMS(cinfo, 1, JTRC_SOI);
+
+  if (cinfo->marker->saw_SOI)
+    ERREXIT(cinfo, JERR_SOI_DUPLICATE);
+
+  /* Reset all parameters that are defined to be reset by SOI */
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+  cinfo->restart_interval = 0;
+
+  /* Set initial assumptions for colorspace etc */
+
+  cinfo->jpeg_color_space = JCS_UNKNOWN;
+  cinfo->CCIR601_sampling = FALSE; /* Assume non-CCIR sampling??? */
+
+  cinfo->saw_JFIF_marker = FALSE;
+  cinfo->JFIF_major_version = 1; /* set default JFIF APP0 values */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;
+  cinfo->X_density = 1;
+  cinfo->Y_density = 1;
+  cinfo->saw_Adobe_marker = FALSE;
+  cinfo->Adobe_transform = 0;
+
+  cinfo->marker->saw_SOI = TRUE;
+
+  return TRUE;
+}
+
+
+LOCAL(int)
+get_sof (j_decompress_ptr cinfo, int is_prog, int is_arith)
+/* Process a SOFn marker */
+{
+  long length;
+  int c, ci;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  cinfo->progressive_mode = is_prog;
+  cinfo->arith_code = is_arith;
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, cinfo->data_precision, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_height, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_width, return FALSE);
+  INPUT_BYTE(cinfo, cinfo->num_components, return FALSE);
+
+  length -= 8;
+
+  TRACEMS4(cinfo, 1, JTRC_SOF, cinfo->unread_marker,
+	   (int) cinfo->image_width, (int) cinfo->image_height,
+	   cinfo->num_components);
+
+  if (cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOF_DUPLICATE);
+
+  /* We don't support files in which the image height is initially specified */
+  /* as 0 and is later redefined by DNL.  As long as we have to check that,  */
+  /* might as well have a general sanity check. */
+  if (cinfo->image_height <= 0 || cinfo->image_width <= 0
+      || cinfo->num_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  if (length != (cinfo->num_components * 3))
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  if (cinfo->comp_info == NULL)	/* do only once, even if suspend */
+    cinfo->comp_info = (jpeg_component_info *) (*cinfo->mem->alloc_small)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 cinfo->num_components * SIZEOF(jpeg_component_info));
+  
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->component_index = ci;
+    INPUT_BYTE(cinfo, compptr->component_id, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    compptr->h_samp_factor = (c >> 4) & 15;
+    compptr->v_samp_factor = (c     ) & 15;
+    INPUT_BYTE(cinfo, compptr->quant_tbl_no, return FALSE);
+
+    TRACEMS4(cinfo, 1, JTRC_SOF_COMPONENT,
+	     compptr->component_id, compptr->h_samp_factor,
+	     compptr->v_samp_factor, compptr->quant_tbl_no);
+  }
+
+  cinfo->marker->saw_SOF = TRUE;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(int)
+get_sos (j_decompress_ptr cinfo)
+/* Process a SOS marker */
+{
+  long length;
+  int i, ci, n, c, cc;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  if (! cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOS_NO_SOF);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, n, return FALSE); /* Number of components */
+
+  TRACEMS1(cinfo, 1, JTRC_SOS, n);
+
+  if (length != (n * 2 + 6) || n < 1 || n > MAX_COMPS_IN_SCAN)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  cinfo->comps_in_scan = n;
+
+  /* Collect the component-spec parameters */
+
+  for (i = 0; i < n; i++) {
+    INPUT_BYTE(cinfo, cc, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (cc == compptr->component_id)
+	goto id_found;
+    }
+
+    ERREXIT1(cinfo, JERR_BAD_COMPONENT_ID, cc);
+
+  id_found:
+
+    cinfo->cur_comp_info[i] = compptr;
+    compptr->dc_tbl_no = (c >> 4) & 15;
+    compptr->ac_tbl_no = (c     ) & 15;
+    
+    TRACEMS3(cinfo, 1, JTRC_SOS_COMPONENT, cc,
+	     compptr->dc_tbl_no, compptr->ac_tbl_no);
+  }
+
+  /* Collect the additional scan parameters Ss, Se, Ah/Al. */
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ss = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Se = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ah = (c >> 4) & 15;
+  cinfo->Al = (c     ) & 15;
+
+  TRACEMS4(cinfo, 1, JTRC_SOS_PARAMS, cinfo->Ss, cinfo->Se,
+	   cinfo->Ah, cinfo->Al);
+
+  /* Prepare to scan data & restart markers */
+  cinfo->marker->next_restart_num = 0;
+
+  /* Count another SOS marker */
+  cinfo->input_scan_number++;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+#ifdef D_ARITH_CODING_SUPPORTED
+
+LOCAL(int)
+get_dac (j_decompress_ptr cinfo)
+/* Process a DAC marker */
+{
+  long length;
+  int index, val;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 0) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+    INPUT_BYTE(cinfo, val, return FALSE);
+
+    length -= 2;
+
+    TRACEMS2(cinfo, 1, JTRC_DAC, index, val);
+
+    if (index < 0 || index >= (2*NUM_ARITH_TBLS))
+      ERREXIT1(cinfo, JERR_DAC_INDEX, index);
+
+    if (index >= NUM_ARITH_TBLS) { /* define AC table */
+      cinfo->arith_ac_K[index-NUM_ARITH_TBLS] = (unsigned char) val;
+    } else {			/* define DC table */
+      cinfo->arith_dc_L[index] = (unsigned char) (val & 0x0F);
+      cinfo->arith_dc_U[index] = (unsigned char) (val >> 4);
+      if (cinfo->arith_dc_L[index] > cinfo->arith_dc_U[index])
+	ERREXIT1(cinfo, JERR_DAC_VALUE, val);
+    }
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+#else /* ! D_ARITH_CODING_SUPPORTED */
+
+#define get_dac(cinfo)  skip_variable(cinfo)
+
+#endif /* D_ARITH_CODING_SUPPORTED */
+
+
+LOCAL(int)
+get_dht (j_decompress_ptr cinfo)
+/* Process a DHT marker */
+{
+  long length;
+  unsigned char bits[17];
+  unsigned char huffval[256];
+  int i, index, count;
+  JHUFF_TBL **htblptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 16) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+
+    TRACEMS1(cinfo, 1, JTRC_DHT, index);
+      
+    bits[0] = 0;
+    count = 0;
+    for (i = 1; i <= 16; i++) {
+      INPUT_BYTE(cinfo, bits[i], return FALSE);
+      count += bits[i];
+    }
+
+    length -= 1 + 16;
+
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[1], bits[2], bits[3], bits[4],
+	     bits[5], bits[6], bits[7], bits[8]);
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[9], bits[10], bits[11], bits[12],
+	     bits[13], bits[14], bits[15], bits[16]);
+
+    /* Here we just do minimal validation of the counts to avoid walking
+     * off the end of our table space.  jdhuff.c will check more carefully.
+     */
+    if (count > 256 || ((long) count) > length)
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+    for (i = 0; i < count; i++)
+      INPUT_BYTE(cinfo, huffval[i], return FALSE);
+
+    length -= count;
+
+    if (index & 0x10) {		/* AC table definition */
+      index -= 0x10;
+      htblptr = &cinfo->ac_huff_tbl_ptrs[index];
+    } else {			/* DC table definition */
+      htblptr = &cinfo->dc_huff_tbl_ptrs[index];
+    }
+
+    if (index < 0 || index >= NUM_HUFF_TBLS)
+      ERREXIT1(cinfo, JERR_DHT_INDEX, index);
+
+    if (*htblptr == NULL)
+      *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+  
+    MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+    MEMCOPY((*htblptr)->huffval, huffval, SIZEOF((*htblptr)->huffval));
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(int)
+get_dqt (j_decompress_ptr cinfo)
+/* Process a DQT marker */
+{
+  long length;
+  int n, i, prec;
+  unsigned int tmp;
+  JQUANT_TBL *quant_ptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 0) {
+    INPUT_BYTE(cinfo, n, return FALSE);
+    prec = n >> 4;
+    n &= 0x0F;
+
+    TRACEMS2(cinfo, 1, JTRC_DQT, n, prec);
+
+    if (n >= NUM_QUANT_TBLS)
+      ERREXIT1(cinfo, JERR_DQT_INDEX, n);
+      
+    if (cinfo->quant_tbl_ptrs[n] == NULL)
+      cinfo->quant_tbl_ptrs[n] = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+    quant_ptr = cinfo->quant_tbl_ptrs[n];
+
+    for (i = 0; i < DCTSIZE2; i++) {
+      if (prec)
+	INPUT_2BYTES(cinfo, tmp, return FALSE);
+      else
+	INPUT_BYTE(cinfo, tmp, return FALSE);
+      /* We convert the zigzag-order table to natural array order. */
+      quant_ptr->quantval[jpeg_natural_order[i]] = (unsigned short) tmp;
+    }
+
+    if (cinfo->err->trace_level >= 2) {
+      for (i = 0; i < DCTSIZE2; i += 8) {
+	TRACEMS8(cinfo, 2, JTRC_QUANTVALS,
+		 quant_ptr->quantval[i],   quant_ptr->quantval[i+1],
+		 quant_ptr->quantval[i+2], quant_ptr->quantval[i+3],
+		 quant_ptr->quantval[i+4], quant_ptr->quantval[i+5],
+		 quant_ptr->quantval[i+6], quant_ptr->quantval[i+7]);
+      }
+    }
+
+    length -= DCTSIZE2+1;
+    if (prec) length -= DCTSIZE2;
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(int)
+get_dri (j_decompress_ptr cinfo)
+/* Process a DRI marker */
+{
+  long length;
+  unsigned int tmp;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  
+  if (length != 4)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_2BYTES(cinfo, tmp, return FALSE);
+
+  TRACEMS1(cinfo, 1, JTRC_DRI, tmp);
+
+  cinfo->restart_interval = tmp;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Routines for processing APPn and COM markers.
+ * These are either saved in memory or discarded, per application request.
+ * APP0 and APP14 are specially checked to see if they are
+ * JFIF and Adobe markers, respectively.
+ */
+
+#define APP0_DATA_LEN	14	/* Length of interesting data in APP0 */
+#define APP14_DATA_LEN	12	/* Length of interesting data in APP14 */
+#define APPN_DATA_LEN	14	/* Must be the largest of the above!! */
+
+
+LOCAL(void)
+examine_app0 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	      unsigned int datalen, long remaining)
+/* Examine first few bytes from an APP0.
+ * Take appropriate action if it is a JFIF marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  long totallen = (long) datalen + remaining;
+
+  if (datalen >= APP0_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x49 &&
+      GETJOCTET(data[3]) == 0x46 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF APP0 marker: save info */
+    cinfo->saw_JFIF_marker = TRUE;
+    cinfo->JFIF_major_version = GETJOCTET(data[5]);
+    cinfo->JFIF_minor_version = GETJOCTET(data[6]);
+    cinfo->density_unit = GETJOCTET(data[7]);
+    cinfo->X_density = (GETJOCTET(data[8]) << 8) + GETJOCTET(data[9]);
+    cinfo->Y_density = (GETJOCTET(data[10]) << 8) + GETJOCTET(data[11]);
+    /* Check version.
+     * Major version must be 1, anything else signals an incompatible change.
+     * (We used to treat this as an error, but now it's a nonfatal warning,
+     * because some bozo at Hijaak couldn't read the spec.)
+     * Minor version should be 0..2, but process anyway if newer.
+     */
+    if (cinfo->JFIF_major_version != 1)
+      WARNMS2(cinfo, JWRN_JFIF_MAJOR,
+	      cinfo->JFIF_major_version, cinfo->JFIF_minor_version);
+    /* Generate trace messages */
+    TRACEMS5(cinfo, 1, JTRC_JFIF,
+	     cinfo->JFIF_major_version, cinfo->JFIF_minor_version,
+	     cinfo->X_density, cinfo->Y_density, cinfo->density_unit);
+    /* Validate thumbnail dimensions and issue appropriate messages */
+    if (GETJOCTET(data[12]) | GETJOCTET(data[13]))
+      TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL,
+	       GETJOCTET(data[12]), GETJOCTET(data[13]));
+    totallen -= APP0_DATA_LEN;
+    if (totallen !=
+	((long)GETJOCTET(data[12]) * (long)GETJOCTET(data[13]) * (long) 3))
+      TRACEMS1(cinfo, 1, JTRC_JFIF_BADTHUMBNAILSIZE, (int) totallen);
+  } else if (datalen >= 6 &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x58 &&
+      GETJOCTET(data[3]) == 0x58 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF "JFXX" extension APP0 marker */
+    /* The library doesn't actually do anything with these,
+     * but we try to produce a helpful trace message.
+     */
+    switch (GETJOCTET(data[5])) {
+    case 0x10:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_JPEG, (int) totallen);
+      break;
+    case 0x11:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_PALETTE, (int) totallen);
+      break;
+    case 0x13:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_RGB, (int) totallen);
+      break;
+    default:
+      TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION,
+	       GETJOCTET(data[5]), (int) totallen);
+      break;
+    }
+  } else {
+    /* Start of APP0 does not match "JFIF" or "JFXX", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP0, (int) totallen);
+  }
+}
+
+
+LOCAL(void)
+examine_app14 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	       unsigned int datalen, long remaining)
+/* Examine first few bytes from an APP14.
+ * Take appropriate action if it is an Adobe marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  unsigned int version, flags0, flags1, transform;
+
+  if (datalen >= APP14_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x41 &&
+      GETJOCTET(data[1]) == 0x64 &&
+      GETJOCTET(data[2]) == 0x6F &&
+      GETJOCTET(data[3]) == 0x62 &&
+      GETJOCTET(data[4]) == 0x65) {
+    /* Found Adobe APP14 marker */
+    version = (GETJOCTET(data[5]) << 8) + GETJOCTET(data[6]);
+    flags0 = (GETJOCTET(data[7]) << 8) + GETJOCTET(data[8]);
+    flags1 = (GETJOCTET(data[9]) << 8) + GETJOCTET(data[10]);
+    transform = GETJOCTET(data[11]);
+    TRACEMS4(cinfo, 1, JTRC_ADOBE, version, flags0, flags1, transform);
+    cinfo->saw_Adobe_marker = TRUE;
+    cinfo->Adobe_transform = (unsigned char) transform;
+  } else {
+    /* Start of APP14 does not match "Adobe", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP14, (int) (datalen + remaining));
+  }
+}
+
+
+METHODDEF(int)
+get_interesting_appn (j_decompress_ptr cinfo)
+/* Process an APP0 or APP14 marker without saving it */
+{
+  long length;
+  JOCTET b[APPN_DATA_LEN];
+  unsigned int i, numtoread;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  /* get the interesting part of the marker data */
+  if (length >= APPN_DATA_LEN)
+    numtoread = APPN_DATA_LEN;
+  else if (length > 0)
+    numtoread = (unsigned int) length;
+  else
+    numtoread = 0;
+  for (i = 0; i < numtoread; i++)
+    INPUT_BYTE(cinfo, b[i], return FALSE);
+  length -= numtoread;
+
+  /* process it */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  default:
+    /* can't get here unless jpeg_save_markers chooses wrong processor */
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+METHODDEF(int)
+save_marker (j_decompress_ptr cinfo)
+/* Save an APPn or COM marker into the marker list */
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  jpeg_saved_marker_ptr cur_marker = marker->cur_marker;
+  unsigned int bytes_read, data_length;
+  JOCTET FAR * data;
+  long length = 0;
+  INPUT_VARS(cinfo);
+
+  if (cur_marker == NULL) {
+    /* begin reading a marker */
+    INPUT_2BYTES(cinfo, length, return FALSE);
+    length -= 2;
+    if (length >= 0) {		/* watch out for bogus length word */
+      /* figure out how much we want to save */
+      unsigned int limit;
+      if (cinfo->unread_marker == (int) M_COM)
+	limit = marker->length_limit_COM;
+      else
+	limit = marker->length_limit_APPn[cinfo->unread_marker - (int) M_APP0];
+      if ((unsigned int) length < limit)
+	limit = (unsigned int) length;
+      /* allocate and initialize the marker item */
+      cur_marker = (jpeg_saved_marker_ptr)
+	(*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				    SIZEOF(struct jpeg_marker_struct) + limit);
+      cur_marker->next = NULL;
+      cur_marker->marker = (unsigned char) cinfo->unread_marker;
+      cur_marker->original_length = (unsigned int) length;
+      cur_marker->data_length = limit;
+      /* data area is just beyond the jpeg_marker_struct */
+      data = cur_marker->data = (JOCTET FAR *) (cur_marker + 1);
+      marker->cur_marker = cur_marker;
+      marker->bytes_read = 0;
+      bytes_read = 0;
+      data_length = limit;
+    } else {
+      /* deal with bogus length word */
+      bytes_read = data_length = 0;
+      data = NULL;
+    }
+  } else {
+    /* resume reading a marker */
+    bytes_read = marker->bytes_read;
+    data_length = cur_marker->data_length;
+    data = cur_marker->data + bytes_read;
+  }
+
+  while (bytes_read < data_length) {
+    INPUT_SYNC(cinfo);		/* move the restart point to here */
+    marker->bytes_read = bytes_read;
+    /* If there's not at least one byte in buffer, suspend */
+    MAKE_BYTE_AVAIL(cinfo, return FALSE);
+    /* Copy bytes with reasonable rapidity */
+    while (bytes_read < data_length && bytes_in_buffer > 0) {
+      *data++ = *next_input_byte++;
+      bytes_in_buffer--;
+      bytes_read++;
+    }
+  }
+
+  /* Done reading what we want to read */
+  if (cur_marker != NULL) {	/* will be NULL if bogus length word */
+    /* Add new marker to end of list */
+    if (cinfo->marker_list == NULL) {
+      cinfo->marker_list = cur_marker;
+    } else {
+      jpeg_saved_marker_ptr prev = cinfo->marker_list;
+      while (prev->next != NULL)
+	prev = prev->next;
+      prev->next = cur_marker;
+    }
+    /* Reset pointer & calc remaining data length */
+    data = cur_marker->data;
+    length = cur_marker->original_length - data_length;
+  }
+  /* Reset to initial state for next marker */
+  marker->cur_marker = NULL;
+
+  /* Process the marker if interesting; else just make a generic trace msg */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, data, data_length, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, data, data_length, length);
+    break;
+  default:
+    TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker,
+	     (int) (data_length + length));
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+METHODDEF(int)
+skip_variable (j_decompress_ptr cinfo)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  long length;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker, (int) length);
+
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+/*
+ * Find the next JPEG marker, save it in cinfo->unread_marker.
+ * Returns FALSE if had to suspend before reaching a marker;
+ * in that case cinfo->unread_marker is unchanged.
+ *
+ * Note that the result might not be a valid marker code,
+ * but it will never be 0 or FF.
+ */
+
+LOCAL(int)
+next_marker (j_decompress_ptr cinfo)
+{
+  int c;
+  INPUT_VARS(cinfo);
+
+  for (;;) {
+    INPUT_BYTE(cinfo, c, return FALSE);
+    /* Skip any non-FF bytes.
+     * This may look a bit inefficient, but it will not occur in a valid file.
+     * We sync after each discarded byte so that a suspending data source
+     * can discard the byte from its buffer.
+     */
+    while (c != 0xFF) {
+      cinfo->marker->discarded_bytes++;
+      INPUT_SYNC(cinfo);
+      INPUT_BYTE(cinfo, c, return FALSE);
+    }
+    /* This loop swallows any duplicate FF bytes.  Extra FFs are legal as
+     * pad bytes, so don't count them in discarded_bytes.  We assume there
+     * will not be so many consecutive FF bytes as to overflow a suspending
+     * data source's input buffer.
+     */
+    do {
+      INPUT_BYTE(cinfo, c, return FALSE);
+    } while (c == 0xFF);
+    if (c != 0)
+      break;			/* found a valid marker, exit loop */
+    /* Reach here if we found a stuffed-zero data sequence (FF/00).
+     * Discard it and loop back to try again.
+     */
+    cinfo->marker->discarded_bytes += 2;
+    INPUT_SYNC(cinfo);
+  }
+
+  if (cinfo->marker->discarded_bytes != 0) {
+    WARNMS2(cinfo, JWRN_EXTRANEOUS_DATA, cinfo->marker->discarded_bytes, c);
+    cinfo->marker->discarded_bytes = 0;
+  }
+
+  cinfo->unread_marker = c;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(int)
+first_marker (j_decompress_ptr cinfo)
+/* Like next_marker, but used to obtain the initial SOI marker. */
+/* For this marker, we do not allow preceding garbage or fill; otherwise,
+ * we might well scan an entire input file before realizing it ain't JPEG.
+ * If an application wants to process non-JFIF files, it must seek to the
+ * SOI before calling the JPEG library.
+ */
+{
+  int c, c2;
+  INPUT_VARS(cinfo);
+
+  INPUT_BYTE(cinfo, c, return FALSE);
+  INPUT_BYTE(cinfo, c2, return FALSE);
+  if (c != 0xFF || c2 != (int) M_SOI)
+    ERREXIT2(cinfo, JERR_NO_SOI, c, c2);
+
+  cinfo->unread_marker = c2;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Read markers until SOS or EOI.
+ *
+ * Returns same codes as are defined for jpeg_consume_input:
+ * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ */
+
+METHODDEF(int)
+read_markers (j_decompress_ptr cinfo)
+{
+  /* Outer loop repeats once for each marker. */
+  for (;;) {
+    /* Collect the marker proper, unless we already did. */
+    /* NB: first_marker() enforces the requirement that SOI appear first. */
+    if (cinfo->unread_marker == 0) {
+      if (! cinfo->marker->saw_SOI) {
+	if (! first_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      } else {
+	if (! next_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      }
+    }
+    /* At this point cinfo->unread_marker contains the marker code and the
+     * input point is just past the marker proper, but before any parameters.
+     * A suspension will cause us to return with this state still true.
+     */
+    switch (cinfo->unread_marker) {
+    case M_SOI:
+      if (! get_soi(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF0:		/* Baseline */
+    case M_SOF1:		/* Extended sequential, Huffman */
+      if (! get_sof(cinfo, FALSE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF2:		/* Progressive, Huffman */
+      if (! get_sof(cinfo, TRUE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF9:		/* Extended sequential, arithmetic */
+      if (! get_sof(cinfo, FALSE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF10:		/* Progressive, arithmetic */
+      if (! get_sof(cinfo, TRUE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    /* Currently unsupported SOFn types */
+    case M_SOF3:		/* Lossless, Huffman */
+    case M_SOF5:		/* Differential sequential, Huffman */
+    case M_SOF6:		/* Differential progressive, Huffman */
+    case M_SOF7:		/* Differential lossless, Huffman */
+    case M_JPG:			/* Reserved for JPEG extensions */
+    case M_SOF11:		/* Lossless, arithmetic */
+    case M_SOF13:		/* Differential sequential, arithmetic */
+    case M_SOF14:		/* Differential progressive, arithmetic */
+    case M_SOF15:		/* Differential lossless, arithmetic */
+      ERREXIT1(cinfo, JERR_SOF_UNSUPPORTED, cinfo->unread_marker);
+      break;
+
+    case M_SOS:
+      if (! get_sos(cinfo))
+	return JPEG_SUSPENDED;
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_SOS;
+    
+    case M_EOI:
+      TRACEMS(cinfo, 1, JTRC_EOI);
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_EOI;
+      
+    case M_DAC:
+      if (! get_dac(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DHT:
+      if (! get_dht(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DQT:
+      if (! get_dqt(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DRI:
+      if (! get_dri(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_APP0:
+    case M_APP1:
+    case M_APP2:
+    case M_APP3:
+    case M_APP4:
+    case M_APP5:
+    case M_APP6:
+    case M_APP7:
+    case M_APP8:
+    case M_APP9:
+    case M_APP10:
+    case M_APP11:
+    case M_APP12:
+    case M_APP13:
+    case M_APP14:
+    case M_APP15:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_APPn[
+		cinfo->unread_marker - (int) M_APP0]) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_COM:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_COM) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_RST0:		/* these are all parameterless */
+    case M_RST1:
+    case M_RST2:
+    case M_RST3:
+    case M_RST4:
+    case M_RST5:
+    case M_RST6:
+    case M_RST7:
+    case M_TEM:
+      TRACEMS1(cinfo, 1, JTRC_PARMLESS_MARKER, cinfo->unread_marker);
+      break;
+
+    case M_DNL:			/* Ignore DNL ... perhaps the wrong thing */
+      if (! skip_variable(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    default:			/* must be DHP, EXP, JPGn, or RESn */
+      /* For now, we treat the reserved markers as fatal errors since they are
+       * likely to be used to signal incompatible JPEG Part 3 extensions.
+       * Once the JPEG 3 version-number marker is well defined, this code
+       * ought to change!
+       */
+      ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+      break;
+    }
+    /* Successfully processed marker, so reset state variable */
+    cinfo->unread_marker = 0;
+  } /* end loop */
+}
+
+
+/*
+ * Read a restart marker, which is expected to appear next in the datastream;
+ * if the marker is not there, take appropriate recovery action.
+ * Returns FALSE if suspension is required.
+ *
+ * This is called by the entropy decoder after it has read an appropriate
+ * number of MCUs.  cinfo->unread_marker may be nonzero if the entropy decoder
+ * has already read a marker from the data source.  Under normal conditions
+ * cinfo->unread_marker will be reset to 0 before returning; if not reset,
+ * it holds a marker which the decoder will be unable to read past.
+ */
+
+METHODDEF(int)
+read_restart_marker (j_decompress_ptr cinfo)
+{
+  /* Obtain a marker unless we already did. */
+  /* Note that next_marker will complain if it skips any data. */
+  if (cinfo->unread_marker == 0) {
+    if (! next_marker(cinfo))
+      return FALSE;
+  }
+
+  if (cinfo->unread_marker ==
+      ((int) M_RST0 + cinfo->marker->next_restart_num)) {
+    /* Normal case --- swallow the marker and let entropy decoder continue */
+    TRACEMS1(cinfo, 3, JTRC_RST, cinfo->marker->next_restart_num);
+    cinfo->unread_marker = 0;
+  } else {
+    /* Uh-oh, the restart markers have been messed up. */
+    /* Let the data source manager determine how to resync. */
+    if (! (*cinfo->src->resync_to_restart) (cinfo,
+					    cinfo->marker->next_restart_num))
+      return FALSE;
+  }
+
+  /* Update next-restart state */
+  cinfo->marker->next_restart_num = (cinfo->marker->next_restart_num + 1) & 7;
+
+  return TRUE;
+}
+
+
+/*
+ * This is the default resync_to_restart method for data source managers
+ * to use if they don't have any better approach.  Some data source managers
+ * may be able to back up, or may have additional knowledge about the data
+ * which permits a more intelligent recovery strategy; such managers would
+ * presumably supply their own resync method.
+ *
+ * read_restart_marker calls resync_to_restart if it finds a marker other than
+ * the restart marker it was expecting.  (This code is *not* used unless
+ * a nonzero restart interval has been declared.)  cinfo->unread_marker is
+ * the marker code actually found (might be anything, except 0 or FF).
+ * The desired restart marker number (0..7) is passed as a parameter.
+ * This routine is supposed to apply whatever error recovery strategy seems
+ * appropriate in order to position the input stream to the next data segment.
+ * Note that cinfo->unread_marker is treated as a marker appearing before
+ * the current data-source input point; usually it should be reset to zero
+ * before returning.
+ * Returns FALSE if suspension is required.
+ *
+ * This implementation is substantially constrained by wanting to treat the
+ * input as a data stream; this means we can't back up.  Therefore, we have
+ * only the following actions to work with:
+ *   1. Simply discard the marker and let the entropy decoder resume at next
+ *      byte of file.
+ *   2. Read forward until we find another marker, discarding intervening
+ *      data.  (In theory we could look ahead within the current bufferload,
+ *      without having to discard data if we don't find the desired marker.
+ *      This idea is not implemented here, in part because it makes behavior
+ *      dependent on buffer size and chance buffer-boundary positions.)
+ *   3. Leave the marker unread (by failing to zero cinfo->unread_marker).
+ *      This will cause the entropy decoder to process an empty data segment,
+ *      inserting dummy zeroes, and then we will reprocess the marker.
+ *
+ * #2 is appropriate if we think the desired marker lies ahead, while #3 is
+ * appropriate if the found marker is a future restart marker (indicating
+ * that we have missed the desired restart marker, probably because it got
+ * corrupted).
+ * We apply #2 or #3 if the found marker is a restart marker no more than
+ * two counts behind or ahead of the expected one.  We also apply #2 if the
+ * found marker is not a legal JPEG marker code (it's certainly bogus data).
+ * If the found marker is a restart marker more than 2 counts away, we do #1
+ * (too much risk that the marker is erroneous; with luck we will be able to
+ * resync at some future point).
+ * For any valid non-restart JPEG marker, we apply #3.  This keeps us from
+ * overrunning the end of a scan.  An implementation limited to single-scan
+ * files might find it better to apply #2 for markers other than EOI, since
+ * any other marker would have to be bogus data in that case.
+ */
+
+GLOBAL(int)
+jpeg_resync_to_restart (j_decompress_ptr cinfo, int desired)
+{
+  int marker = cinfo->unread_marker;
+  int action = 1;
+  
+  /* Always put up a warning. */
+  WARNMS2(cinfo, JWRN_MUST_RESYNC, marker, desired);
+  
+  /* Outer loop handles repeated decision after scanning forward. */
+  for (;;) {
+    if (marker < (int) M_SOF0)
+      action = 2;		/* invalid marker */
+    else if (marker < (int) M_RST0 || marker > (int) M_RST7)
+      action = 3;		/* valid non-restart marker */
+    else {
+      if (marker == ((int) M_RST0 + ((desired+1) & 7)) ||
+	  marker == ((int) M_RST0 + ((desired+2) & 7)))
+	action = 3;		/* one of the next two expected restarts */
+      else if (marker == ((int) M_RST0 + ((desired-1) & 7)) ||
+	       marker == ((int) M_RST0 + ((desired-2) & 7)))
+	action = 2;		/* a prior restart, so advance */
+      else
+	action = 1;		/* desired restart or too far away */
+    }
+    TRACEMS2(cinfo, 4, JTRC_RECOVERY_ACTION, marker, action);
+    switch (action) {
+    case 1:
+      /* Discard marker and let entropy decoder resume processing. */
+      cinfo->unread_marker = 0;
+      return TRUE;
+    case 2:
+      /* Scan to the next marker, and repeat the decision loop. */
+      if (! next_marker(cinfo))
+	return FALSE;
+      marker = cinfo->unread_marker;
+      break;
+    case 3:
+      /* Return without advancing past this marker. */
+      /* Entropy decoder will be forced to process an empty segment. */
+      return TRUE;
+    }
+  } /* end loop */
+}
+
+
+/*
+ * Reset marker processing state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  cinfo->comp_info = NULL;		/* until allocated by get_sof */
+  cinfo->input_scan_number = 0;		/* no SOS seen yet */
+  cinfo->unread_marker = 0;		/* no pending marker */
+  marker->pub.saw_SOI = FALSE;		/* set internal state too */
+  marker->pub.saw_SOF = FALSE;
+  marker->pub.discarded_bytes = 0;
+  marker->cur_marker = NULL;
+}
+
+
+/*
+ * Initialize the marker reader module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker;
+  int i;
+
+  /* Create subobject in permanent pool */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_marker_reader));
+  cinfo->marker = (struct jpeg_marker_reader *) marker;
+  /* Initialize public method pointers */
+  marker->pub.reset_marker_reader = reset_marker_reader;
+  marker->pub.read_markers = read_markers;
+  marker->pub.read_restart_marker = read_restart_marker;
+  /* Initialize COM/APPn processing.
+   * By default, we examine and then discard APP0 and APP14,
+   * but simply discard COM and all other APPn.
+   */
+  marker->process_COM = skip_variable;
+  marker->length_limit_COM = 0;
+  for (i = 0; i < 16; i++) {
+    marker->process_APPn[i] = skip_variable;
+    marker->length_limit_APPn[i] = 0;
+  }
+  marker->process_APPn[0] = get_interesting_appn;
+  marker->process_APPn[14] = get_interesting_appn;
+  /* Reset marker processing state */
+  reset_marker_reader(cinfo);
+}
+
+
+/*
+ * Control saving of COM and APPn markers into marker_list.
+ */
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+GLOBAL(void)
+jpeg_save_markers (j_decompress_ptr cinfo, int marker_code,
+		   unsigned int length_limit)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  long maxlength;
+  jpeg_marker_parser_method processor;
+
+  /* Length limit mustn't be larger than what we can allocate
+   * (should only be a concern in a 16-bit environment).
+   */
+  maxlength = cinfo->mem->max_alloc_chunk - SIZEOF(struct jpeg_marker_struct);
+  if (((long) length_limit) > maxlength)
+    length_limit = (unsigned int) maxlength;
+
+  /* Choose processor routine to use.
+   * APP0/APP14 have special requirements.
+   */
+  if (length_limit) {
+    processor = save_marker;
+    /* If saving APP0/APP14, save at least enough for our internal use. */
+    if (marker_code == (int) M_APP0 && length_limit < APP0_DATA_LEN)
+      length_limit = APP0_DATA_LEN;
+    else if (marker_code == (int) M_APP14 && length_limit < APP14_DATA_LEN)
+      length_limit = APP14_DATA_LEN;
+  } else {
+    processor = skip_variable;
+    /* If discarding APP0/APP14, use our regular on-the-fly processor. */
+    if (marker_code == (int) M_APP0 || marker_code == (int) M_APP14)
+      processor = get_interesting_appn;
+  }
+
+  if (marker_code == (int) M_COM) {
+    marker->process_COM = processor;
+    marker->length_limit_COM = length_limit;
+  } else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15) {
+    marker->process_APPn[marker_code - (int) M_APP0] = processor;
+    marker->length_limit_APPn[marker_code - (int) M_APP0] = length_limit;
+  } else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+/*
+ * Install a special processing method for COM or APPn markers.
+ */
+
+GLOBAL(void)
+jpeg_set_marker_processor (j_decompress_ptr cinfo, int marker_code,
+			   jpeg_marker_parser_method routine)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  if (marker_code == (int) M_COM)
+    marker->process_COM = routine;
+  else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15)
+    marker->process_APPn[marker_code - (int) M_APP0] = routine;
+  else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdmaster.cpp b/ml/dlib/dlib/external/libjpeg/jdmaster.cpp
new file mode 100644
index 000000000..8aea1c8ea
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdmaster.cpp
@@ -0,0 +1,557 @@
+/*
+ * jdmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG decompressor.
+ * These routines are concerned with selecting the modules to be executed
+ * and with determining the number of passes and the work to be done in each
+ * pass.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_decomp_master pub; /* public fields */
+
+  int pass_number;		/* # of passes completed */
+
+  int using_merged_upsample; /* TRUE if using merged upsample/cconvert */
+
+  /* Saved references to initialized quantizer modules,
+   * in case we need to switch modes.
+   */
+  struct jpeg_color_quantizer * quantizer_1pass;
+  struct jpeg_color_quantizer * quantizer_2pass;
+} my_decomp_master;
+
+typedef my_decomp_master * my_master_ptr;
+
+
+/*
+ * Determine whether merged upsample/color conversion should be used.
+ * CRUCIAL: this must match the actual capabilities of jdmerge.c!
+ */
+
+LOCAL(int)
+use_merged_upsample (j_decompress_ptr cinfo)
+{
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  /* Merging is the equivalent of plain box-filter upsampling */
+  if (cinfo->do_fancy_upsampling || cinfo->CCIR601_sampling)
+    return FALSE;
+  /* jdmerge.c only supports YCC=>RGB color conversion */
+  if (cinfo->jpeg_color_space != JCS_YCbCr || cinfo->num_components != 3 ||
+      cinfo->out_color_space != JCS_RGB ||
+      cinfo->out_color_components != RGB_PIXELSIZE)
+    return FALSE;
+  /* and it only handles 2h1v or 2h2v sampling ratios */
+  if (cinfo->comp_info[0].h_samp_factor != 2 ||
+      cinfo->comp_info[1].h_samp_factor != 1 ||
+      cinfo->comp_info[2].h_samp_factor != 1 ||
+      cinfo->comp_info[0].v_samp_factor >  2 ||
+      cinfo->comp_info[1].v_samp_factor != 1 ||
+      cinfo->comp_info[2].v_samp_factor != 1)
+    return FALSE;
+  /* furthermore, it doesn't work if we've scaled the IDCTs differently */
+  if (cinfo->comp_info[0].DCT_scaled_size != cinfo->min_DCT_scaled_size ||
+      cinfo->comp_info[1].DCT_scaled_size != cinfo->min_DCT_scaled_size ||
+      cinfo->comp_info[2].DCT_scaled_size != cinfo->min_DCT_scaled_size)
+    return FALSE;
+  /* ??? also need to test for upsample-time rescaling, when & if supported */
+  return TRUE;			/* by golly, it'll work... */
+#else
+  return FALSE;
+#endif
+}
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ * Also note that it may be called before the master module is initialized!
+ */
+
+GLOBAL(void)
+jpeg_calc_output_dimensions (j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+#endif
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_READY)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+  /* Compute actual output image dimensions and DCT scaling choices. */
+  if (cinfo->scale_num * 8 <= cinfo->scale_denom) {
+    /* Provide 1/8 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 8L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 8L);
+    cinfo->min_DCT_scaled_size = 1;
+  } else if (cinfo->scale_num * 4 <= cinfo->scale_denom) {
+    /* Provide 1/4 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 4L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 4L);
+    cinfo->min_DCT_scaled_size = 2;
+  } else if (cinfo->scale_num * 2 <= cinfo->scale_denom) {
+    /* Provide 1/2 scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, 2L);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, 2L);
+    cinfo->min_DCT_scaled_size = 4;
+  } else {
+    /* Provide 1/1 scaling */
+    cinfo->output_width = cinfo->image_width;
+    cinfo->output_height = cinfo->image_height;
+    cinfo->min_DCT_scaled_size = DCTSIZE;
+  }
+  /* In selecting the actual DCT scaling for each component, we try to
+   * scale up the chroma components via IDCT scaling rather than upsampling.
+   * This saves time if the upsampler gets to use 1:1 scaling.
+   * Note this code assumes that the supported DCT scalings are powers of 2.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    int ssize = cinfo->min_DCT_scaled_size;
+    while (ssize < DCTSIZE &&
+	   (compptr->h_samp_factor * ssize * 2 <=
+	    cinfo->max_h_samp_factor * cinfo->min_DCT_scaled_size) &&
+	   (compptr->v_samp_factor * ssize * 2 <=
+	    cinfo->max_v_samp_factor * cinfo->min_DCT_scaled_size)) {
+      ssize = ssize * 2;
+    }
+    compptr->DCT_scaled_size = ssize;
+  }
+
+  /* Recompute downsampled dimensions of components;
+   * application needs to know these if using raw downsampled data.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Size in samples, after IDCT scaling */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width *
+		    (long) (compptr->h_samp_factor * compptr->DCT_scaled_size),
+		    (long) (cinfo->max_h_samp_factor * DCTSIZE));
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height *
+		    (long) (compptr->v_samp_factor * compptr->DCT_scaled_size),
+		    (long) (cinfo->max_v_samp_factor * DCTSIZE));
+  }
+
+#else /* !IDCT_SCALING_SUPPORTED */
+
+  /* Hardwire it to "no scaling" */
+  cinfo->output_width = cinfo->image_width;
+  cinfo->output_height = cinfo->image_height;
+  /* jdinput.c has already initialized DCT_scaled_size to DCTSIZE,
+   * and has computed unscaled downsampled_width and downsampled_height.
+   */
+
+#endif /* IDCT_SCALING_SUPPORTED */
+
+  /* Report number of components in selected colorspace. */
+  /* Probably this should be in the color conversion module... */
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    break;
+  case JCS_RGB:
+#if RGB_PIXELSIZE != 3
+    cinfo->out_color_components = RGB_PIXELSIZE;
+    break;
+#endif /* else share code with YCbCr */
+  case JCS_YCbCr:
+    cinfo->out_color_components = 3;
+    break;
+  case JCS_CMYK:
+  case JCS_YCCK:
+    cinfo->out_color_components = 4;
+    break;
+  default:			/* else must be same colorspace as in file */
+    cinfo->out_color_components = cinfo->num_components;
+    break;
+  }
+  cinfo->output_components = (cinfo->quantize_colors ? 1 :
+			      cinfo->out_color_components);
+
+  /* See if upsampler will want to emit more than one row at a time */
+  if (use_merged_upsample(cinfo))
+    cinfo->rec_outbuf_height = cinfo->max_v_samp_factor;
+  else
+    cinfo->rec_outbuf_height = 1;
+}
+
+
+/*
+ * Several decompression processes need to range-limit values to the range
+ * 0..MAXJSAMPLE; the input value may fall somewhat outside this range
+ * due to noise introduced by quantization, roundoff error, etc.  These
+ * processes are inner loops and need to be as fast as possible.  On most
+ * machines, particularly CPUs with pipelines or instruction prefetch,
+ * a (subscript-check-less) C table lookup
+ *		x = sample_range_limit[x];
+ * is faster than explicit tests
+ *		if (x < 0)  x = 0;
+ *		else if (x > MAXJSAMPLE)  x = MAXJSAMPLE;
+ * These processes all use a common table prepared by the routine below.
+ *
+ * For most steps we can mathematically guarantee that the initial value
+ * of x is within MAXJSAMPLE+1 of the legal range, so a table running from
+ * -(MAXJSAMPLE+1) to 2*MAXJSAMPLE+1 is sufficient.  But for the initial
+ * limiting step (just after the IDCT), a wildly out-of-range value is 
+ * possible if the input data is corrupt.  To avoid any chance of indexing
+ * off the end of memory and getting a bad-pointer trap, we perform the
+ * post-IDCT limiting thus:
+ *		x = range_limit[x & MASK];
+ * where MASK is 2 bits wider than legal sample data, ie 10 bits for 8-bit
+ * samples.  Under normal circumstances this is more than enough range and
+ * a correct output will be generated; with bogus input data the mask will
+ * cause wraparound, and we will safely generate a bogus-but-in-range output.
+ * For the post-IDCT step, we want to convert the data from signed to unsigned
+ * representation by adding CENTERJSAMPLE at the same time that we limit it.
+ * So the post-IDCT limiting table ends up looking like this:
+ *   CENTERJSAMPLE,CENTERJSAMPLE+1,...,MAXJSAMPLE,
+ *   MAXJSAMPLE (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0          (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0,1,...,CENTERJSAMPLE-1
+ * Negative inputs select values from the upper half of the table after
+ * masking.
+ *
+ * We can save some space by overlapping the start of the post-IDCT table
+ * with the simpler range limiting table.  The post-IDCT table begins at
+ * sample_range_limit + CENTERJSAMPLE.
+ *
+ * Note that the table is allocated in near data space on PCs; it's small
+ * enough and used often enough to justify this.
+ */
+
+LOCAL(void)
+prepare_range_limit_table (j_decompress_ptr cinfo)
+/* Allocate and fill in the sample_range_limit table */
+{
+  JSAMPLE * table;
+  int i;
+
+  table = (JSAMPLE *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(5 * (MAXJSAMPLE+1) + CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  table += (MAXJSAMPLE+1);	/* allow negative subscripts of simple table */
+  cinfo->sample_range_limit = table;
+  /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+  MEMZERO(table - (MAXJSAMPLE+1), (MAXJSAMPLE+1) * SIZEOF(JSAMPLE));
+  /* Main part of "simple" table: limit[x] = x */
+  for (i = 0; i <= MAXJSAMPLE; i++)
+    table[i] = (JSAMPLE) i;
+  table += CENTERJSAMPLE;	/* Point to where post-IDCT table starts */
+  /* End of simple table, rest of first half of post-IDCT table */
+  for (i = CENTERJSAMPLE; i < 2*(MAXJSAMPLE+1); i++)
+    table[i] = MAXJSAMPLE;
+  /* Second half of post-IDCT table */
+  MEMZERO(table + (2 * (MAXJSAMPLE+1)),
+	  (2 * (MAXJSAMPLE+1) - CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  MEMCOPY(table + (4 * (MAXJSAMPLE+1) - CENTERJSAMPLE),
+	  cinfo->sample_range_limit, CENTERJSAMPLE * SIZEOF(JSAMPLE));
+}
+
+
+/*
+ * Master selection of decompression modules.
+ * This is done once at jpeg_start_decompress time.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ * We also initialize the decompressor input side to begin consuming data.
+ *
+ * Since jpeg_read_header has finished, we know what is in the SOF
+ * and (first) SOS markers.  We also have all the application parameter
+ * settings.
+ */
+
+LOCAL(void)
+master_selection (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+  int use_c_buffer;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  /* Initialize dimensions and other stuff */
+  jpeg_calc_output_dimensions(cinfo);
+  prepare_range_limit_table(cinfo);
+
+  /* Width of an output scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->output_width * (long) cinfo->out_color_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* Initialize my private state */
+  master->pass_number = 0;
+  master->using_merged_upsample = use_merged_upsample(cinfo);
+
+  /* Color quantizer selection */
+  master->quantizer_1pass = NULL;
+  master->quantizer_2pass = NULL;
+  /* No mode changes if not using buffered-image mode. */
+  if (! cinfo->quantize_colors || ! cinfo->buffered_image) {
+    cinfo->enable_1pass_quant = FALSE;
+    cinfo->enable_external_quant = FALSE;
+    cinfo->enable_2pass_quant = FALSE;
+  }
+  if (cinfo->quantize_colors) {
+    if (cinfo->raw_data_out)
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    /* 2-pass quantizer only works in 3-component color space. */
+    if (cinfo->out_color_components != 3) {
+      cinfo->enable_1pass_quant = TRUE;
+      cinfo->enable_external_quant = FALSE;
+      cinfo->enable_2pass_quant = FALSE;
+      cinfo->colormap = NULL;
+    } else if (cinfo->colormap != NULL) {
+      cinfo->enable_external_quant = TRUE;
+    } else if (cinfo->two_pass_quantize) {
+      cinfo->enable_2pass_quant = TRUE;
+    } else {
+      cinfo->enable_1pass_quant = TRUE;
+    }
+
+    if (cinfo->enable_1pass_quant) {
+#ifdef QUANT_1PASS_SUPPORTED
+      jinit_1pass_quantizer(cinfo);
+      master->quantizer_1pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+
+    /* We use the 2-pass code to map to external colormaps. */
+    if (cinfo->enable_2pass_quant || cinfo->enable_external_quant) {
+#ifdef QUANT_2PASS_SUPPORTED
+      jinit_2pass_quantizer(cinfo);
+      master->quantizer_2pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+    /* If both quantizers are initialized, the 2-pass one is left active;
+     * this is necessary for starting with quantization to an external map.
+     */
+  }
+
+  /* Post-processing: in particular, color conversion first */
+  if (! cinfo->raw_data_out) {
+    if (master->using_merged_upsample) {
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+      jinit_merged_upsampler(cinfo); /* does color conversion too */
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else {
+      jinit_color_deconverter(cinfo);
+      jinit_upsampler(cinfo);
+    }
+    jinit_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+  }
+  /* Inverse DCT */
+  jinit_inverse_dct(cinfo);
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code) {
+    ERREXIT(cinfo, JERR_ARITH_NOTIMPL);
+  } else {
+    if (cinfo->progressive_mode) {
+#ifdef D_PROGRESSIVE_SUPPORTED
+      jinit_phuff_decoder(cinfo);
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else
+      jinit_huff_decoder(cinfo);
+  }
+
+  /* Initialize principal buffer controllers. */
+  use_c_buffer = cinfo->inputctl->has_multiple_scans || cinfo->buffered_image;
+  jinit_d_coef_controller(cinfo, use_c_buffer);
+
+  if (! cinfo->raw_data_out)
+    jinit_d_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* If jpeg_start_decompress will read the whole file, initialize
+   * progress monitoring appropriately.  The input step is counted
+   * as one pass.
+   */
+  if (cinfo->progress != NULL && ! cinfo->buffered_image &&
+      cinfo->inputctl->has_multiple_scans) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = (cinfo->enable_2pass_quant ? 3 : 2);
+    /* Count the input pass as done */
+    master->pass_number++;
+  }
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each output pass.  We determine which
+ * modules will be active during this pass and give them appropriate
+ * start_pass calls.  We also set is_dummy_pass to indicate whether this
+ * is a "real" output pass or a dummy pass for color quantization.
+ * (In the latter case, jdapistd.c will crank the pass to completion.)
+ */
+
+METHODDEF(void)
+prepare_for_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (master->pub.is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Final pass of 2-pass quantization */
+    master->pub.is_dummy_pass = FALSE;
+    (*cinfo->cquantize->start_pass) (cinfo, FALSE);
+    (*cinfo->post->start_pass) (cinfo, JBUF_CRANK_DEST);
+    (*cinfo->main->start_pass) (cinfo, JBUF_CRANK_DEST);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  } else {
+    if (cinfo->quantize_colors && cinfo->colormap == NULL) {
+      /* Select new quantization method */
+      if (cinfo->two_pass_quantize && cinfo->enable_2pass_quant) {
+	cinfo->cquantize = master->quantizer_2pass;
+	master->pub.is_dummy_pass = TRUE;
+      } else if (cinfo->enable_1pass_quant) {
+	cinfo->cquantize = master->quantizer_1pass;
+      } else {
+	ERREXIT(cinfo, JERR_MODE_CHANGE);
+      }
+    }
+    (*cinfo->idct->start_pass) (cinfo);
+    (*cinfo->coef->start_output_pass) (cinfo);
+    if (! cinfo->raw_data_out) {
+      if (! master->using_merged_upsample)
+	(*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->upsample->start_pass) (cinfo);
+      if (cinfo->quantize_colors)
+	(*cinfo->cquantize->start_pass) (cinfo, master->pub.is_dummy_pass);
+      (*cinfo->post->start_pass) (cinfo,
+	    (master->pub.is_dummy_pass ? JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+      (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+  }
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->pass_number +
+				    (master->pub.is_dummy_pass ? 2 : 1);
+    /* In buffered-image mode, we assume one more output pass if EOI not
+     * yet reached, but no more passes if EOI has been reached.
+     */
+    if (cinfo->buffered_image && ! cinfo->inputctl->eoi_reached) {
+      cinfo->progress->total_passes += (cinfo->enable_2pass_quant ? 2 : 1);
+    }
+  }
+}
+
+
+/*
+ * Finish up at end of an output pass.
+ */
+
+METHODDEF(void)
+finish_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (cinfo->quantize_colors)
+    (*cinfo->cquantize->finish_pass) (cinfo);
+  master->pass_number++;
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+GLOBAL(void)
+jpeg_new_colormap (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_BUFIMAGE)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (cinfo->quantize_colors && cinfo->enable_external_quant &&
+      cinfo->colormap != NULL) {
+    /* Select 2-pass quantizer for external colormap use */
+    cinfo->cquantize = master->quantizer_2pass;
+    /* Notify quantizer of colormap change */
+    (*cinfo->cquantize->new_color_map) (cinfo);
+    master->pub.is_dummy_pass = FALSE; /* just in case */
+  } else
+    ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+/*
+ * Initialize master decompression control and select active modules.
+ * This is performed at the start of jpeg_start_decompress.
+ */
+
+GLOBAL(void)
+jinit_master_decompress (j_decompress_ptr cinfo)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_decomp_master));
+  cinfo->master = (struct jpeg_decomp_master *) master;
+  master->pub.prepare_for_output_pass = prepare_for_output_pass;
+  master->pub.finish_output_pass = finish_output_pass;
+
+  master->pub.is_dummy_pass = FALSE;
+
+  master_selection(cinfo);
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdmerge.cpp b/ml/dlib/dlib/external/libjpeg/jdmerge.cpp
new file mode 100644
index 000000000..38a692ff8
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdmerge.cpp
@@ -0,0 +1,400 @@
+/*
+ * jdmerge.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ *
+ * This file combines functions from jdsample.c and jdcolor.c;
+ * read those files first to understand what's going on.
+ *
+ * When the chroma components are to be upsampled by simple replication
+ * (ie, box filtering), we can save some work in color conversion by
+ * calculating all the output pixels corresponding to a pair of chroma
+ * samples at one time.  In the conversion equations
+ *	R = Y           + K1 * Cr
+ *	G = Y + K2 * Cb + K3 * Cr
+ *	B = Y + K4 * Cb
+ * only the Y term varies among the group of pixels corresponding to a pair
+ * of chroma samples, so the rest of the terms can be calculated just once.
+ * At typical sampling ratios, this eliminates half or three-quarters of the
+ * multiplications needed for color conversion.
+ *
+ * This file currently provides implementations for the following cases:
+ *	YCbCr => RGB color conversion only.
+ *	Sampling ratios of 2h1v or 2h2v.
+ *	No scaling needed at upsample time.
+ *	Corner-aligned (non-CCIR601) sampling alignment.
+ * Other special cases could be added, but in most applications these are
+ * the only common cases.  (For uncommon cases we fall back on the more
+ * general code in jdsample.c and jdcolor.c.)
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Pointer to routine to do actual upsampling/conversion of one row group */
+  JMETHOD(void, upmethod, (j_decompress_ptr cinfo,
+			   JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+			   JSAMPARRAY output_buf));
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  long * Cr_g_tab;		/* => table for Cr to G conversion */
+  long * Cb_g_tab;		/* => table for Cb to G conversion */
+
+  /* For 2:1 vertical sampling, we produce two output rows at a time.
+   * We need a "spare" row buffer to hold the second output row if the
+   * application provides just a one-row buffer; we also use the spare
+   * to discard the dummy last row if the image height is odd.
+   */
+  JSAMPROW spare_row;
+  int spare_full;		/* T if spare buffer is occupied */
+
+  JDIMENSION out_row_width;	/* samples per output row */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((long) 1 << (SCALEBITS-1))
+#define FIX(x)		((long) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ * This is taken directly from jdcolor.c; see that file for more info.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int i;
+  long x;
+  SHIFT_TEMPS
+
+  upsample->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cr_g_tab = (long *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(long));
+  upsample->Cb_g_tab = (long *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(long));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    upsample->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    upsample->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    upsample->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    upsample->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_merged_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the spare buffer empty */
+  upsample->spare_full = FALSE;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * The control routine just handles the row buffering considerations.
+ */
+
+METHODDEF(void)
+merged_2v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION ,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+/* 2:1 vertical sampling case: may need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPROW work_ptrs[2];
+  JDIMENSION num_rows;		/* number of rows returned to caller */
+
+  if (upsample->spare_full) {
+    /* If we have a spare row saved from a previous cycle, just return it. */
+    jcopy_sample_rows(& upsample->spare_row, 0, output_buf + *out_row_ctr, 0,
+		      1, upsample->out_row_width);
+    num_rows = 1;
+    upsample->spare_full = FALSE;
+  } else {
+    /* Figure number of rows to return to caller. */
+    num_rows = 2;
+    /* Not more than the distance to the end of the image. */
+    if (num_rows > upsample->rows_to_go)
+      num_rows = upsample->rows_to_go;
+    /* And not more than what the client can accept: */
+    out_rows_avail -= *out_row_ctr;
+    if (num_rows > out_rows_avail)
+      num_rows = out_rows_avail;
+    /* Create output pointer array for upsampler. */
+    work_ptrs[0] = output_buf[*out_row_ctr];
+    if (num_rows > 1) {
+      work_ptrs[1] = output_buf[*out_row_ctr + 1];
+    } else {
+      work_ptrs[1] = upsample->spare_row;
+      upsample->spare_full = TRUE;
+    }
+    /* Now do the upsampling. */
+    (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr, work_ptrs);
+  }
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (! upsample->spare_full)
+    (*in_row_group_ctr)++;
+}
+
+
+METHODDEF(void)
+merged_1v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION ,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION )
+/* 1:1 vertical sampling case: much easier, never need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Just do the upsampling. */
+  (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr,
+			 output_buf + *out_row_ctr);
+  /* Adjust counts */
+  (*out_row_ctr)++;
+  (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by the control routines to do
+ * the actual upsampling/conversion.  One row group is processed per call.
+ *
+ * Note: since we may be writing directly into application-supplied buffers,
+ * we have to be honest about the output width; we can't assume the buffer
+ * has been rounded up to an even width.
+ */
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+METHODDEF(void)
+h2v1_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int y, cred, cgreen, cblue;
+  int cb, cr;
+  JSAMPROW outptr;
+  JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  long * Crgtab = upsample->Cr_g_tab;
+  long * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+    outptr += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+    outptr += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr0);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ */
+
+METHODDEF(void)
+h2v2_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int y, cred, cgreen, cblue;
+  int cb, cr;
+  JSAMPROW outptr0, outptr1;
+  JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  long * Crgtab = upsample->Cr_g_tab;
+  long * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr*2];
+  inptr01 = input_buf[0][in_row_group_ctr*2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr00);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    y  = GETJSAMPLE(*inptr01);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Module initialization routine for merged upsampling/color conversion.
+ *
+ * NB: this is called under the conditions determined by use_merged_upsample()
+ * in jdmaster.c.  That routine MUST correspond to the actual capabilities
+ * of this module; no safety checks are made here.
+ */
+
+GLOBAL(void)
+jinit_merged_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_merged_upsample;
+  upsample->pub.need_context_rows = FALSE;
+
+  upsample->out_row_width = cinfo->output_width * cinfo->out_color_components;
+
+  if (cinfo->max_v_samp_factor == 2) {
+    upsample->pub.upsample = merged_2v_upsample;
+    upsample->upmethod = h2v2_merged_upsample;
+    /* Allocate a spare row buffer */
+    upsample->spare_row = (JSAMPROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(size_t) (upsample->out_row_width * SIZEOF(JSAMPLE)));
+  } else {
+    upsample->pub.upsample = merged_1v_upsample;
+    upsample->upmethod = h2v1_merged_upsample;
+    /* No spare row needed */
+    upsample->spare_row = NULL;
+  }
+
+  build_ycc_rgb_table(cinfo);
+}
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jdphuff.cpp b/ml/dlib/dlib/external/libjpeg/jdphuff.cpp
new file mode 100644
index 000000000..d9c02fe0b
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdphuff.cpp
@@ -0,0 +1,671 @@
+/*
+ * jdphuff.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy decoding routines for progressive JPEG.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdhuff.h"		/* Declarations shared with jdhuff.c */
+
+#ifdef __GNUC__
+#pragma GCC diagnostic ignored "-Wshift-negative-value"
+#endif
+
+#ifdef D_PROGRESSIVE_SUPPORTED
+
+/*
+ * Expanded entropy decoder object for progressive Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  unsigned int EOBRUN;			/* remaining EOBs in EOBRUN */
+  int last_dc_val[MAX_COMPS_IN_SCAN];	/* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).EOBRUN = (src).EOBRUN, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;	/* Bit buffer at start of MCU */
+  savable_state saved;		/* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * derived_tbls[NUM_HUFF_TBLS];
+
+  d_derived_tbl * ac_derived_tbl; /* active table during an AC scan */
+} phuff_entropy_decoder;
+
+typedef phuff_entropy_decoder * phuff_entropy_ptr;
+
+/* Forward declarations */
+METHODDEF(int) decode_mcu_DC_first JPP((j_decompress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(int) decode_mcu_AC_first JPP((j_decompress_ptr cinfo,
+					    JBLOCKROW *MCU_data));
+METHODDEF(int) decode_mcu_DC_refine JPP((j_decompress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+METHODDEF(int) decode_mcu_AC_refine JPP((j_decompress_ptr cinfo,
+					     JBLOCKROW *MCU_data));
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_phuff_decoder (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int is_DC_band, bad;
+  int ci, coefi, tbl;
+  int *coef_bit_ptr;
+  jpeg_component_info * compptr;
+
+  is_DC_band = (cinfo->Ss == 0);
+
+  /* Validate scan parameters */
+  bad = FALSE;
+  if (is_DC_band) {
+    if (cinfo->Se != 0)
+      bad = TRUE;
+  } else {
+    /* need not check Ss/Se < 0 since they came from unsigned bytes */
+    if (cinfo->Ss > cinfo->Se || cinfo->Se >= DCTSIZE2)
+      bad = TRUE;
+    /* AC scans may have only one component */
+    if (cinfo->comps_in_scan != 1)
+      bad = TRUE;
+  }
+  if (cinfo->Ah != 0) {
+    /* Successive approximation refinement scan: must have Al = Ah-1. */
+    if (cinfo->Al != cinfo->Ah-1)
+      bad = TRUE;
+  }
+  if (cinfo->Al > 13)		/* need not check for < 0 */
+    bad = TRUE;
+  /* Arguably the maximum Al value should be less than 13 for 8-bit precision,
+   * but the spec doesn't say so, and we try to be liberal about what we
+   * accept.  Note: large Al values could result in out-of-range DC
+   * coefficients during early scans, leading to bizarre displays due to
+   * overflows in the IDCT math.  But we won't crash.
+   */
+  if (bad)
+    ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	     cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+  /* Update progression status, and verify that scan order is legal.
+   * Note that inter-scan inconsistencies are treated as warnings
+   * not fatal errors ... not clear if this is right way to behave.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    int cindex = cinfo->cur_comp_info[ci]->component_index;
+    coef_bit_ptr = & cinfo->coef_bits[cindex][0];
+    if (!is_DC_band && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+      WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+    for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+      int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+      if (cinfo->Ah != expected)
+	WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+      coef_bit_ptr[coefi] = cinfo->Al;
+    }
+  }
+
+  /* Select MCU decoding routine */
+  if (cinfo->Ah == 0) {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_first;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_first;
+  } else {
+    if (is_DC_band)
+      entropy->pub.decode_mcu = decode_mcu_DC_refine;
+    else
+      entropy->pub.decode_mcu = decode_mcu_AC_refine;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* Make sure requested tables are present, and compute derived tables.
+     * We may build same derived table more than once, but it's not expensive.
+     */
+    if (is_DC_band) {
+      if (cinfo->Ah == 0) {	/* DC refinement needs no table */
+	tbl = compptr->dc_tbl_no;
+	jpeg_make_d_derived_tbl(cinfo, TRUE, tbl,
+				& entropy->derived_tbls[tbl]);
+      }
+    } else {
+      tbl = compptr->ac_tbl_no;
+      jpeg_make_d_derived_tbl(cinfo, FALSE, tbl,
+			      & entropy->derived_tbls[tbl]);
+      /* remember the single active table */
+      entropy->ac_derived_tbl = entropy->derived_tbls[tbl];
+    }
+    /* Initialize DC predictions to 0 */
+    entropy->saved.last_dc_val[ci] = 0;
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->pub.insufficient_data = FALSE;
+
+  /* Initialize private state variables */
+  entropy->saved.EOBRUN = 0;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and add will be faster than a table lookup.
+ */
+
+#ifdef AVOID_TABLES
+
+#define HUFF_EXTEND(x,s)  ((x) < (1<<((s)-1)) ? (x) + (((-1)<<(s)) + 1) : (x))
+
+#else
+
+#define HUFF_EXTEND(x,s)  ((x) < extend_test[s] ? (x) + extend_offset[s] : (x))
+
+static const int extend_test[16] =   /* entry n is 2**(n-1) */
+  { 0, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
+    0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000 };
+
+static const int extend_offset[16] = /* entry n is (-1 << n) + 1 */
+  { 0, ((-1)<<1) + 1, ((-1)<<2) + 1, ((-1)<<3) + 1, ((-1)<<4) + 1,
+    ((-1)<<5) + 1, ((-1)<<6) + 1, ((-1)<<7) + 1, ((-1)<<8) + 1,
+    ((-1)<<9) + 1, ((-1)<<10) + 1, ((-1)<<11) + 1, ((-1)<<12) + 1,
+    ((-1)<<13) + 1, ((-1)<<14) + 1, ((-1)<<15) + 1 };
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(int)
+process_restart (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+  /* Re-init EOB run count, too */
+  entropy->saved.EOBRUN = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->pub.insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Huffman MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * Huffman-compressed coefficients. 
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ *
+ * We return FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * spectral selection, since we'll just re-assign them on the next call.
+ * Successive approximation AC refinement has to be more careful, however.)
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(int)
+decode_mcu_DC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Al = cinfo->Al;
+  int s, r;
+  int blkn, ci;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  savable_state state;
+  d_derived_tbl * tbl;
+  jpeg_component_info * compptr;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      block = MCU_data[blkn];
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      tbl = entropy->derived_tbls[compptr->dc_tbl_no];
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      HUFF_DECODE(s, br_state, tbl, return FALSE, label1);
+      if (s) {
+	CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	r = GET_BITS(s);
+	s = HUFF_EXTEND(r, s);
+      }
+
+      /* Convert DC difference to actual value, update last_dc_val */
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      /* Scale and output the coefficient (assumes jpeg_natural_order[0]=0) */
+      (*block)[0] = (JCOEF) (s << Al);
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(int)
+decode_mcu_AC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Se = cinfo->Se;
+  int Al = cinfo->Al;
+  int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state.
+     * We can avoid loading/saving bitread state if in an EOB run.
+     */
+    EOBRUN = entropy->saved.EOBRUN;	/* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+
+    if (EOBRUN > 0)		/* if it's a band of zeroes... */
+      EOBRUN--;			/* ...process it now (we do nothing) */
+    else {
+      BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+      block = MCU_data[0];
+      tbl = entropy->ac_derived_tbl;
+
+      for (k = cinfo->Ss; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, return FALSE, label2);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  k += r;
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  r = GET_BITS(s);
+	  s = HUFF_EXTEND(r, s);
+	  /* Scale and output coefficient in natural (dezigzagged) order */
+	  (*block)[jpeg_natural_order[k]] = (JCOEF) (s << Al);
+	} else {
+	  if (r == 15) {	/* ZRL */
+	    k += 15;		/* skip 15 zeroes in band */
+	  } else {		/* EOBr, run length is 2^r + appended bits */
+	    EOBRUN = 1 << r;
+	    if (r) {		/* EOBr, r > 0 */
+	      CHECK_BIT_BUFFER(br_state, r, return FALSE);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    EOBRUN--;		/* this band is processed at this moment */
+	    break;		/* force end-of-band */
+	  }
+	}
+      }
+
+      BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    }
+
+    /* Completed MCU, so update state */
+    entropy->saved.EOBRUN = EOBRUN;	/* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(int)
+decode_mcu_DC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+  int blkn;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* Not worth the cycles to check insufficient_data here,
+   * since we will not change the data anyway if we read zeroes.
+   */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    CHECK_BIT_BUFFER(br_state, 1, return FALSE);
+    if (GET_BITS(1))
+      (*block)[0] |= p1;
+    /* Note: since we use |=, repeating the assignment later is safe */
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(int)
+decode_mcu_AC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  phuff_entropy_ptr entropy = (phuff_entropy_ptr) cinfo->entropy;
+  int Se = cinfo->Se;
+  int p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+  int m1 = (-1) << cinfo->Al;	/* -1 in the bit position being coded */
+  int s, k, r;
+  unsigned int EOBRUN;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+  int num_newnz;
+  int newnz_pos[DCTSIZE2];
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, don't modify the MCU.
+   */
+  if (! entropy->pub.insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    EOBRUN = entropy->saved.EOBRUN; /* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+    block = MCU_data[0];
+    tbl = entropy->ac_derived_tbl;
+
+    /* If we are forced to suspend, we must undo the assignments to any newly
+     * nonzero coefficients in the block, because otherwise we'd get confused
+     * next time about which coefficients were already nonzero.
+     * But we need not undo addition of bits to already-nonzero coefficients;
+     * instead, we can test the current bit to see if we already did it.
+     */
+    num_newnz = 0;
+
+    /* initialize coefficient loop counter to start of band */
+    k = cinfo->Ss;
+
+    if (EOBRUN == 0) {
+      for (; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, goto undoit, label3);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  if (s != 1)		/* size of new coef should always be 1 */
+	    WARNMS(cinfo, JWRN_HUFF_BAD_CODE);
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1))
+	    s = p1;		/* newly nonzero coef is positive */
+	  else
+	    s = m1;		/* newly nonzero coef is negative */
+	} else {
+	  if (r != 15) {
+	    EOBRUN = 1 << r;	/* EOBr, run length is 2^r + appended bits */
+	    if (r) {
+	      CHECK_BIT_BUFFER(br_state, r, goto undoit);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    break;		/* rest of block is handled by EOB logic */
+	  }
+	  /* note s = 0 for processing ZRL */
+	}
+	/* Advance over already-nonzero coefs and r still-zero coefs,
+	 * appending correction bits to the nonzeroes.  A correction bit is 1
+	 * if the absolute value of the coefficient must be increased.
+	 */
+	do {
+	  thiscoef = *block + jpeg_natural_order[k];
+	  if (*thiscoef != 0) {
+	    CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	    if (GET_BITS(1)) {
+	      if ((*thiscoef & p1) == 0) { /* do nothing if already set it */
+		if (*thiscoef >= 0)
+		  *thiscoef += p1;
+		else
+		  *thiscoef += m1;
+	      }
+	    }
+	  } else {
+	    if (--r < 0)
+	      break;		/* reached target zero coefficient */
+	  }
+	  k++;
+	} while (k <= Se);
+	if (s) {
+	  int pos = jpeg_natural_order[k];
+	  /* Output newly nonzero coefficient */
+	  (*block)[pos] = (JCOEF) s;
+	  /* Remember its position in case we have to suspend */
+	  newnz_pos[num_newnz++] = pos;
+	}
+      }
+    }
+
+    if (EOBRUN > 0) {
+      /* Scan any remaining coefficient positions after the end-of-band
+       * (the last newly nonzero coefficient, if any).  Append a correction
+       * bit to each already-nonzero coefficient.  A correction bit is 1
+       * if the absolute value of the coefficient must be increased.
+       */
+      for (; k <= Se; k++) {
+	thiscoef = *block + jpeg_natural_order[k];
+	if (*thiscoef != 0) {
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1)) {
+	    if ((*thiscoef & p1) == 0) { /* do nothing if already changed it */
+	      if (*thiscoef >= 0)
+		*thiscoef += p1;
+	      else
+		*thiscoef += m1;
+	    }
+	  }
+	}
+      }
+      /* Count one block completed in EOB run */
+      EOBRUN--;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    entropy->saved.EOBRUN = EOBRUN; /* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+
+undoit:
+  /* Re-zero any output coefficients that we made newly nonzero */
+  while (num_newnz > 0)
+    (*block)[newnz_pos[--num_newnz]] = 0;
+
+  return FALSE;
+}
+
+
+/*
+ * Module initialization routine for progressive Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_phuff_decoder (j_decompress_ptr cinfo)
+{
+  phuff_entropy_ptr entropy;
+  int *coef_bit_ptr;
+  int ci, i;
+
+  entropy = (phuff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(phuff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass_phuff_decoder;
+
+  /* Mark derived tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->derived_tbls[i] = NULL;
+  }
+
+  /* Create progression status table */
+  cinfo->coef_bits = (int (*)[DCTSIZE2])
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				cinfo->num_components*DCTSIZE2*SIZEOF(int));
+  coef_bit_ptr = & cinfo->coef_bits[0][0];
+  for (ci = 0; ci < cinfo->num_components; ci++) 
+    for (i = 0; i < DCTSIZE2; i++)
+      *coef_bit_ptr++ = -1;
+}
+
+#endif /* D_PROGRESSIVE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jdpostct.cpp b/ml/dlib/dlib/external/libjpeg/jdpostct.cpp
new file mode 100644
index 000000000..63e10ec17
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdpostct.cpp
@@ -0,0 +1,290 @@
+/*
+ * jdpostct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the decompression postprocessing controller.
+ * This controller manages the upsampling, color conversion, and color
+ * quantization/reduction steps; specifically, it controls the buffering
+ * between upsample/color conversion and color quantization/reduction.
+ *
+ * If no color quantization/reduction is required, then this module has no
+ * work to do, and it just hands off to the upsample/color conversion code.
+ * An integrated upsample/convert/quantize process would replace this module
+ * entirely.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_post_controller pub; /* public fields */
+
+  /* Color quantization source buffer: this holds output data from
+   * the upsample/color conversion step to be passed to the quantizer.
+   * For two-pass color quantization, we need a full-image buffer;
+   * for one-pass operation, a strip buffer is sufficient.
+   */
+  jvirt_sarray_ptr whole_image;	/* virtual array, or NULL if one-pass */
+  JSAMPARRAY buffer;		/* strip buffer, or current strip of virtual */
+  JDIMENSION strip_height;	/* buffer size in rows */
+  /* for two-pass mode only: */
+  JDIMENSION starting_row;	/* row # of first row in current strip */
+  JDIMENSION next_row;		/* index of next row to fill/empty in strip */
+} my_post_controller;
+
+typedef my_post_controller * my_post_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) post_process_1pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) post_process_prepass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+METHODDEF(void) post_process_2pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_dpost (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->quantize_colors) {
+      /* Single-pass processing with color quantization. */
+      post->pub.post_process_data = post_process_1pass;
+      /* We could be doing buffered-image output before starting a 2-pass
+       * color quantization; in that case, jinit_d_post_controller did not
+       * allocate a strip buffer.  Use the virtual-array buffer as workspace.
+       */
+      if (post->buffer == NULL) {
+	post->buffer = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, post->whole_image,
+	   (JDIMENSION) 0, post->strip_height, TRUE);
+      }
+    } else {
+      /* For single-pass processing without color quantization,
+       * I have no work to do; just call the upsampler directly.
+       */
+      post->pub.post_process_data = cinfo->upsample->upsample;
+    }
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    /* First pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_prepass;
+    break;
+  case JBUF_CRANK_DEST:
+    /* Second pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_2pass;
+    break;
+#endif /* QUANT_2PASS_SUPPORTED */
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+  post->starting_row = post->next_row = 0;
+}
+
+
+/*
+ * Process some data in the one-pass (strip buffer) case.
+ * This is used for color precision reduction as well as one-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_1pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Fill the buffer, but not more than what we can dump out in one go. */
+  /* Note we rely on the upsampler to detect bottom of image. */
+  max_rows = out_rows_avail - *out_row_ctr;
+  if (max_rows > post->strip_height)
+    max_rows = post->strip_height;
+  num_rows = 0;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &num_rows, max_rows);
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer, output_buf + *out_row_ctr, (int) num_rows);
+  *out_row_ctr += num_rows;
+}
+
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+/*
+ * Process some data in the first pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_prepass (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		      JDIMENSION in_row_groups_avail,
+		      JSAMPARRAY , JDIMENSION *out_row_ctr,
+		      JDIMENSION )
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION old_next_row, num_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, TRUE);
+  }
+
+  /* Upsample some data (up to a strip height's worth). */
+  old_next_row = post->next_row;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &post->next_row, post->strip_height);
+
+  /* Allow quantizer to scan new data.  No data is emitted, */
+  /* but we advance out_row_ctr so outer loop can tell when we're done. */
+  if (post->next_row > old_next_row) {
+    num_rows = post->next_row - old_next_row;
+    (*cinfo->cquantize->color_quantize) (cinfo, post->buffer + old_next_row,
+					 (JSAMPARRAY) NULL, (int) num_rows);
+    *out_row_ctr += num_rows;
+  }
+
+  /* Advance if we filled the strip. */
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+
+/*
+ * Process some data in the second pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_2pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE , JDIMENSION *,
+		    JDIMENSION ,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, FALSE);
+  }
+
+  /* Determine number of rows to emit. */
+  num_rows = post->strip_height - post->next_row; /* available in strip */
+  max_rows = out_rows_avail - *out_row_ctr; /* available in output area */
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+  /* We have to check bottom of image here, can't depend on upsampler. */
+  max_rows = cinfo->output_height - post->starting_row;
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer + post->next_row, output_buf + *out_row_ctr,
+		(int) num_rows);
+  *out_row_ctr += num_rows;
+
+  /* Advance if we filled the strip. */
+  post->next_row += num_rows;
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize postprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_d_post_controller (j_decompress_ptr cinfo, int need_full_buffer)
+{
+  my_post_ptr post;
+
+  post = (my_post_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_post_controller));
+  cinfo->post = (struct jpeg_d_post_controller *) post;
+  post->pub.start_pass = start_pass_dpost;
+  post->whole_image = NULL;	/* flag for no virtual arrays */
+  post->buffer = NULL;		/* flag for no strip buffer */
+
+  /* Create the quantization buffer, if needed */
+  if (cinfo->quantize_colors) {
+    /* The buffer strip height is max_v_samp_factor, which is typically
+     * an efficient number of rows for upsampling to return.
+     * (In the presence of output rescaling, we might want to be smarter?)
+     */
+    post->strip_height = (JDIMENSION) cinfo->max_v_samp_factor;
+    if (need_full_buffer) {
+      /* Two-pass color quantization: need full-image storage. */
+      /* We round up the number of rows to a multiple of the strip height. */
+#ifdef QUANT_2PASS_SUPPORTED
+      post->whole_image = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 (JDIMENSION) jround_up((long) cinfo->output_height,
+				(long) post->strip_height),
+	 post->strip_height);
+#else
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif /* QUANT_2PASS_SUPPORTED */
+    } else {
+      /* One-pass color quantization: just make a strip buffer. */
+      post->buffer = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 post->strip_height);
+    }
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jdsample.cpp b/ml/dlib/dlib/external/libjpeg/jdsample.cpp
new file mode 100644
index 000000000..647c8bf45
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jdsample.cpp
@@ -0,0 +1,478 @@
+/*
+ * jdsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains upsampling routines.
+ *
+ * Upsampling input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  Upsampling will normally produce
+ * max_v_samp_factor pixel rows from each row group (but this could vary
+ * if the upsampler is applying a scale factor of its own).
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Pointer to routine to upsample a single component */
+typedef JMETHOD(void, upsample1_ptr,
+		(j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Color conversion buffer.  When using separate upsampling and color
+   * conversion steps, this buffer holds one upsampled row group until it
+   * has been color converted and output.
+   * Note: we do not allocate any storage for component(s) which are full-size,
+   * ie do not need rescaling.  The corresponding entry of color_buf[] is
+   * simply set to point to the input data array, thereby avoiding copying.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  /* Per-component upsampling method pointers */
+  upsample1_ptr methods[MAX_COMPONENTS];
+
+  int next_row_out;		/* counts rows emitted from color_buf */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+
+  /* Height of an input row group for each component. */
+  int rowgroup_height[MAX_COMPONENTS];
+
+  /* These arrays save pixel expansion factors so that int_expand need not
+   * recompute them each time.  They are unused for other upsampling methods.
+   */
+  unsigned char h_expand[MAX_COMPONENTS];
+  unsigned char v_expand[MAX_COMPONENTS];
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the conversion buffer empty */
+  upsample->next_row_out = cinfo->max_v_samp_factor;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * In this version we upsample each component independently.
+ * We upsample one row group into the conversion buffer, then apply
+ * color conversion a row at a time.
+ */
+
+METHODDEF(void)
+sep_upsample (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	      JDIMENSION ,
+	      JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	      JDIMENSION out_rows_avail)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JDIMENSION num_rows;
+
+  /* Fill the conversion buffer, if it's empty */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor) {
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      /* Invoke per-component upsample method.  Notice we pass a POINTER
+       * to color_buf[ci], so that fullsize_upsample can change it.
+       */
+      (*upsample->methods[ci]) (cinfo, compptr,
+	input_buf[ci] + (*in_row_group_ctr * upsample->rowgroup_height[ci]),
+	upsample->color_buf + ci);
+    }
+    upsample->next_row_out = 0;
+  }
+
+  /* Color-convert and emit rows */
+
+  /* How many we have in the buffer: */
+  num_rows = (JDIMENSION) (cinfo->max_v_samp_factor - upsample->next_row_out);
+  /* Not more than the distance to the end of the image.  Need this test
+   * in case the image height is not a multiple of max_v_samp_factor:
+   */
+  if (num_rows > upsample->rows_to_go) 
+    num_rows = upsample->rows_to_go;
+  /* And not more than what the client can accept: */
+  out_rows_avail -= *out_row_ctr;
+  if (num_rows > out_rows_avail)
+    num_rows = out_rows_avail;
+
+  (*cinfo->cconvert->color_convert) (cinfo, upsample->color_buf,
+				     (JDIMENSION) upsample->next_row_out,
+				     output_buf + *out_row_ctr,
+				     (int) num_rows);
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  upsample->next_row_out += num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor)
+    (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by sep_upsample to upsample pixel values
+ * of a single component.  One row group is processed per call.
+ */
+
+
+/*
+ * For full-size components, we just make color_buf[ci] point at the
+ * input buffer, and thus avoid copying any data.  Note that this is
+ * safe only because sep_upsample doesn't declare the input row group
+ * "consumed" until we are done color converting and emitting it.
+ */
+
+METHODDEF(void)
+fullsize_upsample (j_decompress_ptr , jpeg_component_info * ,
+		   JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = input_data;
+}
+
+
+/*
+ * This is a no-op version used for "uninteresting" components.
+ * These components will not be referenced by color conversion.
+ */
+
+METHODDEF(void)
+noop_upsample (j_decompress_ptr , jpeg_component_info * ,
+	       JSAMPARRAY , JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = NULL;	/* safety check */
+}
+
+
+/*
+ * This version handles any integral sampling ratios.
+ * This is not used for typical JPEG files, so it need not be fast.
+ * Nor, for that matter, is it particularly accurate: the algorithm is
+ * simple replication of the input pixel onto the corresponding output
+ * pixels.  The hi-falutin sampling literature refers to this as a
+ * "box filter".  A box filter tends to introduce visible artifacts,
+ * so if you are actually going to use 3:1 or 4:1 sampling ratios
+ * you would be well advised to improve this code.
+ */
+
+METHODDEF(void)
+int_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	      JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  JSAMPLE invalue;
+  int h;
+  JSAMPROW outend;
+  int h_expand, v_expand;
+  int inrow, outrow;
+
+  h_expand = upsample->h_expand[compptr->component_index];
+  v_expand = upsample->v_expand[compptr->component_index];
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    /* Generate one output row with proper horizontal expansion */
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      for (h = h_expand; h > 0; h--) {
+	*outptr++ = invalue;
+      }
+    }
+    /* Generate any additional output rows by duplicating the first one */
+    if (v_expand > 1) {
+      jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+			v_expand-1, cinfo->output_width);
+    }
+    inrow++;
+    outrow += v_expand;
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v1_upsample (j_decompress_ptr cinfo, jpeg_component_info * ,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  JSAMPLE invalue;
+  JSAMPROW outend;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v2_upsample (j_decompress_ptr cinfo, jpeg_component_info * ,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  JSAMPLE invalue;
+  JSAMPROW outend;
+  int inrow, outrow;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+    jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+		      1, cinfo->output_width);
+    inrow++;
+    outrow += 2;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
+ *
+ * The upsampling algorithm is linear interpolation between pixel centers,
+ * also known as a "triangle filter".  This is a good compromise between
+ * speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
+ * of the way between input pixel centers.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_fancy_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr, outptr;
+  int invalue;
+  JDIMENSION colctr;
+  int inrow;
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    inptr = input_data[inrow];
+    outptr = output_data[inrow];
+    /* Special case for first column */
+    invalue = GETJSAMPLE(*inptr++);
+    *outptr++ = (JSAMPLE) invalue;
+    *outptr++ = (JSAMPLE) ((invalue * 3 + GETJSAMPLE(*inptr) + 2) >> 2);
+
+    for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+      /* General case: 3/4 * nearer pixel + 1/4 * further pixel */
+      invalue = GETJSAMPLE(*inptr++) * 3;
+      *outptr++ = (JSAMPLE) ((invalue + GETJSAMPLE(inptr[-2]) + 1) >> 2);
+      *outptr++ = (JSAMPLE) ((invalue + GETJSAMPLE(*inptr) + 2) >> 2);
+    }
+
+    /* Special case for last column */
+    invalue = GETJSAMPLE(*inptr);
+    *outptr++ = (JSAMPLE) ((invalue * 3 + GETJSAMPLE(inptr[-1]) + 1) >> 2);
+    *outptr++ = (JSAMPLE) invalue;
+  }
+}
+
+
+/*
+ * Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * Again a triangle filter; see comments for h2v1 case, above.
+ *
+ * It is OK for us to reference the adjacent input rows because we demanded
+ * context from the main buffer controller (see initialization code).
+ */
+
+METHODDEF(void)
+h2v2_fancy_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  JSAMPROW inptr0, inptr1, outptr;
+#if BITS_IN_JSAMPLE == 8
+  int thiscolsum, lastcolsum, nextcolsum;
+#else
+  long thiscolsum, lastcolsum, nextcolsum;
+#endif
+  JDIMENSION colctr;
+  int inrow, outrow, v;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    for (v = 0; v < 2; v++) {
+      /* inptr0 points to nearest input row, inptr1 points to next nearest */
+      inptr0 = input_data[inrow];
+      if (v == 0)		/* next nearest is row above */
+	inptr1 = input_data[inrow-1];
+      else			/* next nearest is row below */
+	inptr1 = input_data[inrow+1];
+      outptr = output_data[outrow++];
+
+      /* Special case for first column */
+      thiscolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+      nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 4 + 8) >> 4);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 3 + nextcolsum + 7) >> 4);
+      lastcolsum = thiscolsum; thiscolsum = nextcolsum;
+
+      for (colctr = compptr->downsampled_width - 2; colctr > 0; colctr--) {
+	/* General case: 3/4 * nearer pixel + 1/4 * further pixel in each */
+	/* dimension, thus 9/16, 3/16, 3/16, 1/16 overall */
+	nextcolsum = GETJSAMPLE(*inptr0++) * 3 + GETJSAMPLE(*inptr1++);
+	*outptr++ = (JSAMPLE) ((thiscolsum * 3 + lastcolsum + 8) >> 4);
+	*outptr++ = (JSAMPLE) ((thiscolsum * 3 + nextcolsum + 7) >> 4);
+	lastcolsum = thiscolsum; thiscolsum = nextcolsum;
+      }
+
+      /* Special case for last column */
+      *outptr++ = (JSAMPLE) ((thiscolsum * 3 + lastcolsum + 8) >> 4);
+      *outptr++ = (JSAMPLE) ((thiscolsum * 4 + 7) >> 4);
+    }
+    inrow++;
+  }
+}
+
+
+/*
+ * Module initialization routine for upsampling.
+ */
+
+GLOBAL(void)
+jinit_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  int need_buffer, do_fancy;
+  int h_in_group, v_in_group, h_out_group, v_out_group;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_upsample;
+  upsample->pub.upsample = sep_upsample;
+  upsample->pub.need_context_rows = FALSE; /* until we find out differently */
+
+  if (cinfo->CCIR601_sampling)	/* this isn't supported */
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* jdmainct.c doesn't support context rows when min_DCT_scaled_size = 1,
+   * so don't ask for it.
+   */
+  do_fancy = cinfo->do_fancy_upsampling && cinfo->min_DCT_scaled_size > 1;
+
+  /* Verify we can handle the sampling factors, select per-component methods,
+   * and create storage as needed.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Compute size of an "input group" after IDCT scaling.  This many samples
+     * are to be converted to max_h_samp_factor * max_v_samp_factor pixels.
+     */
+    h_in_group = (compptr->h_samp_factor * compptr->DCT_scaled_size) /
+		 cinfo->min_DCT_scaled_size;
+    v_in_group = (compptr->v_samp_factor * compptr->DCT_scaled_size) /
+		 cinfo->min_DCT_scaled_size;
+    h_out_group = cinfo->max_h_samp_factor;
+    v_out_group = cinfo->max_v_samp_factor;
+    upsample->rowgroup_height[ci] = v_in_group; /* save for use later */
+    need_buffer = TRUE;
+    if (! compptr->component_needed) {
+      /* Don't bother to upsample an uninteresting component. */
+      upsample->methods[ci] = noop_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group == h_out_group && v_in_group == v_out_group) {
+      /* Fullsize components can be processed without any work. */
+      upsample->methods[ci] = fullsize_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group == v_out_group) {
+      /* Special cases for 2h1v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2)
+	upsample->methods[ci] = h2v1_fancy_upsample;
+      else
+	upsample->methods[ci] = h2v1_upsample;
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group * 2 == v_out_group) {
+      /* Special cases for 2h2v upsampling */
+      if (do_fancy && compptr->downsampled_width > 2) {
+	upsample->methods[ci] = h2v2_fancy_upsample;
+	upsample->pub.need_context_rows = TRUE;
+      } else
+	upsample->methods[ci] = h2v2_upsample;
+    } else if ((h_out_group % h_in_group) == 0 &&
+	       (v_out_group % v_in_group) == 0) {
+      /* Generic integral-factors upsampling method */
+      upsample->methods[ci] = int_upsample;
+      upsample->h_expand[ci] = (unsigned char) (h_out_group / h_in_group);
+      upsample->v_expand[ci] = (unsigned char) (v_out_group / v_in_group);
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+    if (need_buffer) {
+      upsample->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) jround_up((long) cinfo->output_width,
+				(long) cinfo->max_h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jerror.cpp b/ml/dlib/dlib/external/libjpeg/jerror.cpp
new file mode 100644
index 000000000..117bc4829
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jerror.cpp
@@ -0,0 +1,252 @@
+/*
+ * jerror.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains simple error-reporting and trace-message routines.
+ * These are suitable for Unix-like systems and others where writing to
+ * stderr is the right thing to do.  Many applications will want to replace
+ * some or all of these routines.
+ *
+ * If you define USE_WINDOWS_MESSAGEBOX in jconfig.h or in the makefile,
+ * you get a Windows-specific hack to display error messages in a dialog box.
+ * It ain't much, but it beats dropping error messages into the bit bucket,
+ * which is what happens to output to stderr under most Windows C compilers.
+ *
+ * These routines are used by both the compression and decompression code.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jversion.h"
+#include "jerror.h"
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+#include <windows.h>
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+
+
+/*
+ * Create the message string table.
+ * We do this from the master message list in jerror.h by re-reading
+ * jerror.h with a suitable definition for macro JMESSAGE.
+ * The message table is made an external symbol just in case any applications
+ * want to refer to it directly.
+ */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_std_message_table	jMsgTable
+#endif
+
+#define JMESSAGE(code,string)	string ,
+
+const char * const jpeg_std_message_table[] = {
+#include "jerror.h"
+  NULL
+};
+
+
+/*
+ * Error exit handler: must not return to caller.
+ *
+ * Applications may override this if they want to get control back after
+ * an error.  Typically one would longjmp somewhere instead of exiting.
+ * The setjmp buffer can be made a private field within an expanded error
+ * handler object.  Note that the info needed to generate an error message
+ * is stored in the error object, so you can generate the message now or
+ * later, at your convenience.
+ * You should make sure that the JPEG object is cleaned up (with jpeg_abort
+ * or jpeg_destroy) at some point.
+ */
+
+METHODDEF(void)
+error_exit (j_common_ptr cinfo)
+{
+  /* Always display the message */
+  (*cinfo->err->output_message) (cinfo);
+
+  /* Let the memory manager delete any temp files before we die */
+  jpeg_destroy(cinfo);
+
+  exit(EXIT_FAILURE);
+}
+
+
+/*
+ * Actual output of an error or trace message.
+ * Applications may override this method to send JPEG messages somewhere
+ * other than stderr.
+ *
+ * On Windows, printing to stderr is generally completely useless,
+ * so we provide optional code to produce an error-dialog popup.
+ * Most Windows applications will still prefer to override this routine,
+ * but if they don't, it'll do something at least marginally useful.
+ *
+ * NOTE: to use the library in an environment that doesn't support the
+ * C stdio library, you may have to delete the call to fprintf() entirely,
+ * not just not use this routine.
+ */
+
+METHODDEF(void)
+output_message (j_common_ptr cinfo)
+{
+  char buffer[JMSG_LENGTH_MAX];
+
+  /* Create the message */
+  (*cinfo->err->format_message) (cinfo, buffer);
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+  /* Display it in a message dialog box */
+  MessageBox(GetActiveWindow(), buffer, "JPEG Library Error",
+	     MB_OK | MB_ICONERROR);
+#else
+  /* Send it to stderr, adding a newline */
+  fprintf(stderr, "%s\n", buffer);
+#endif
+}
+
+
+/*
+ * Decide whether to emit a trace or warning message.
+ * msg_level is one of:
+ *   -1: recoverable corrupt-data warning, may want to abort.
+ *    0: important advisory messages (always display to user).
+ *    1: first level of tracing detail.
+ *    2,3,...: successively more detailed tracing messages.
+ * An application might override this method if it wanted to abort on warnings
+ * or change the policy about which messages to display.
+ */
+
+METHODDEF(void)
+emit_message (j_common_ptr cinfo, int msg_level)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+
+  if (msg_level < 0) {
+    /* It's a warning message.  Since corrupt files may generate many warnings,
+     * the policy implemented here is to show only the first warning,
+     * unless trace_level >= 3.
+     */
+    if (err->num_warnings == 0 || err->trace_level >= 3)
+      (*err->output_message) (cinfo);
+    /* Always count warnings in num_warnings. */
+    err->num_warnings++;
+  } else {
+    /* It's a trace message.  Show it if trace_level >= msg_level. */
+    if (err->trace_level >= msg_level)
+      (*err->output_message) (cinfo);
+  }
+}
+
+
+/*
+ * Format a message string for the most recent JPEG error or message.
+ * The message is stored into buffer, which should be at least JMSG_LENGTH_MAX
+ * characters.  Note that no '\n' character is added to the string.
+ * Few applications should need to override this method.
+ */
+
+METHODDEF(void)
+format_message (j_common_ptr cinfo, char * buffer)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+  int msg_code = err->msg_code;
+  const char * msgtext = NULL;
+  const char * msgptr;
+  char ch;
+  int isstring;
+
+  /* Look up message string in proper table */
+  if (msg_code > 0 && msg_code <= err->last_jpeg_message) {
+    msgtext = err->jpeg_message_table[msg_code];
+  } else if (err->addon_message_table != NULL &&
+	     msg_code >= err->first_addon_message &&
+	     msg_code <= err->last_addon_message) {
+    msgtext = err->addon_message_table[msg_code - err->first_addon_message];
+  }
+
+  /* Defend against bogus message number */
+  if (msgtext == NULL) {
+    err->msg_parm.i[0] = msg_code;
+    msgtext = err->jpeg_message_table[0];
+  }
+
+  /* Check for string parameter, as indicated by %s in the message text */
+  isstring = FALSE;
+  msgptr = msgtext;
+  while ((ch = *msgptr++) != '\0') {
+    if (ch == '%') {
+      if (*msgptr == 's') isstring = TRUE;
+      break;
+    }
+  }
+
+  /* Format the message into the passed buffer */
+  if (isstring)
+    sprintf(buffer, msgtext, err->msg_parm.s);
+  else
+    sprintf(buffer, msgtext,
+	    err->msg_parm.i[0], err->msg_parm.i[1],
+	    err->msg_parm.i[2], err->msg_parm.i[3],
+	    err->msg_parm.i[4], err->msg_parm.i[5],
+	    err->msg_parm.i[6], err->msg_parm.i[7]);
+}
+
+
+/*
+ * Reset error state variables at start of a new image.
+ * This is called during compression startup to reset trace/error
+ * processing to default state, without losing any application-specific
+ * method pointers.  An application might possibly want to override
+ * this method if it has additional error processing state.
+ */
+
+METHODDEF(void)
+reset_error_mgr (j_common_ptr cinfo)
+{
+  cinfo->err->num_warnings = 0;
+  /* trace_level is not reset since it is an application-supplied parameter */
+  cinfo->err->msg_code = 0;	/* may be useful as a flag for "no error" */
+}
+
+
+/*
+ * Fill in the standard error-handling methods in a jpeg_error_mgr object.
+ * Typical call is:
+ *	struct jpeg_compress_struct cinfo;
+ *	struct jpeg_error_mgr err;
+ *
+ *	cinfo.err = jpeg_std_error(&err);
+ * after which the application may override some of the methods.
+ */
+
+GLOBAL(struct jpeg_error_mgr *)
+jpeg_std_error (struct jpeg_error_mgr * err)
+{
+  err->error_exit = error_exit;
+  err->emit_message = emit_message;
+  err->output_message = output_message;
+  err->format_message = format_message;
+  err->reset_error_mgr = reset_error_mgr;
+
+  err->trace_level = 0;		/* default = no tracing */
+  err->num_warnings = 0;	/* no warnings emitted yet */
+  err->msg_code = 0;		/* may be useful as a flag for "no error" */
+
+  /* Initialize message table pointers */
+  err->jpeg_message_table = jpeg_std_message_table;
+  err->last_jpeg_message = (int) JMSG_LASTMSGCODE - 1;
+
+  err->addon_message_table = NULL;
+  err->first_addon_message = 0;	/* for safety */
+  err->last_addon_message = 0;
+
+  return err;
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jerror.h b/ml/dlib/dlib/external/libjpeg/jerror.h
new file mode 100644
index 000000000..fc2fffeac
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jerror.h
@@ -0,0 +1,291 @@
+/*
+ * jerror.h
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the error and message codes for the JPEG library.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ * A set of error-reporting macros are defined too.  Some applications using
+ * the JPEG library may wish to include this file to get the error codes
+ * and/or the macros.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef JERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code,string)
+#endif /* JERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code,string)	code ,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_NOMESSAGE, "Bogus message code %d") /* Must be first entry! */
+
+/* For maintenance convenience, list is alphabetical by message code name */
+JMESSAGE(JERR_ARITH_NOTIMPL,
+	 "Sorry, there are legal restrictions on arithmetic coding")
+JMESSAGE(JERR_BAD_ALIGN_TYPE, "ALIGN_TYPE is wrong, please fix")
+JMESSAGE(JERR_BAD_ALLOC_CHUNK, "MAX_ALLOC_CHUNK is wrong, please fix")
+JMESSAGE(JERR_BAD_BUFFER_MODE, "Bogus buffer control mode")
+JMESSAGE(JERR_BAD_COMPONENT_ID, "Invalid component ID %d in SOS")
+JMESSAGE(JERR_BAD_DCT_COEF, "DCT coefficient out of range")
+JMESSAGE(JERR_BAD_DCTSIZE, "IDCT output block size %d not supported")
+JMESSAGE(JERR_BAD_HUFF_TABLE, "Bogus Huffman table definition")
+JMESSAGE(JERR_BAD_IN_COLORSPACE, "Bogus input colorspace")
+JMESSAGE(JERR_BAD_J_COLORSPACE, "Bogus JPEG colorspace")
+JMESSAGE(JERR_BAD_LENGTH, "Bogus marker length")
+JMESSAGE(JERR_BAD_LIB_VERSION,
+	 "Wrong JPEG library version: library is %d, caller expects %d")
+JMESSAGE(JERR_BAD_MCU_SIZE, "Sampling factors too large for interleaved scan")
+JMESSAGE(JERR_BAD_POOL_ID, "Invalid memory pool code %d")
+JMESSAGE(JERR_BAD_PRECISION, "Unsupported JPEG data precision %d")
+JMESSAGE(JERR_BAD_PROGRESSION,
+	 "Invalid progressive parameters Ss=%d Se=%d Ah=%d Al=%d")
+JMESSAGE(JERR_BAD_PROG_SCRIPT,
+	 "Invalid progressive parameters at scan script entry %d")
+JMESSAGE(JERR_BAD_SAMPLING, "Bogus sampling factors")
+JMESSAGE(JERR_BAD_SCAN_SCRIPT, "Invalid scan script at entry %d")
+JMESSAGE(JERR_BAD_STATE, "Improper call to JPEG library in state %d")
+JMESSAGE(JERR_BAD_STRUCT_SIZE,
+	 "JPEG parameter struct mismatch: library thinks size is %u, caller expects %u")
+JMESSAGE(JERR_BAD_VIRTUAL_ACCESS, "Bogus virtual array access")
+JMESSAGE(JERR_BUFFER_SIZE, "Buffer passed to JPEG library is too small")
+JMESSAGE(JERR_CANT_SUSPEND, "Suspension not allowed here")
+JMESSAGE(JERR_CCIR601_NOTIMPL, "CCIR601 sampling not implemented yet")
+JMESSAGE(JERR_COMPONENT_COUNT, "Too many color components: %d, max %d")
+JMESSAGE(JERR_CONVERSION_NOTIMPL, "Unsupported color conversion request")
+JMESSAGE(JERR_DAC_INDEX, "Bogus DAC index %d")
+JMESSAGE(JERR_DAC_VALUE, "Bogus DAC value 0x%x")
+JMESSAGE(JERR_DHT_INDEX, "Bogus DHT index %d")
+JMESSAGE(JERR_DQT_INDEX, "Bogus DQT index %d")
+JMESSAGE(JERR_EMPTY_IMAGE, "Empty JPEG image (DNL not supported)")
+JMESSAGE(JERR_EMS_READ, "Read from EMS failed")
+JMESSAGE(JERR_EMS_WRITE, "Write to EMS failed")
+JMESSAGE(JERR_EOI_EXPECTED, "Didn't expect more than one scan")
+JMESSAGE(JERR_FILE_READ, "Input file read error")
+JMESSAGE(JERR_FILE_WRITE, "Output file write error --- out of disk space?")
+JMESSAGE(JERR_FRACT_SAMPLE_NOTIMPL, "Fractional sampling not implemented yet")
+JMESSAGE(JERR_HUFF_CLEN_OVERFLOW, "Huffman code size table overflow")
+JMESSAGE(JERR_HUFF_MISSING_CODE, "Missing Huffman code table entry")
+JMESSAGE(JERR_IMAGE_TOO_BIG, "Maximum supported image dimension is %u pixels")
+JMESSAGE(JERR_INPUT_EMPTY, "Empty input file")
+JMESSAGE(JERR_INPUT_EOF, "Premature end of input file")
+JMESSAGE(JERR_MISMATCHED_QUANT_TABLE,
+	 "Cannot transcode due to multiple use of quantization table %d")
+JMESSAGE(JERR_MISSING_DATA, "Scan script does not transmit all data")
+JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
+JMESSAGE(JERR_NOTIMPL, "Not implemented yet")
+JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
+JMESSAGE(JERR_NO_BACKING_STORE, "Backing store not supported")
+JMESSAGE(JERR_NO_HUFF_TABLE, "Huffman table 0x%02x was not defined")
+JMESSAGE(JERR_NO_IMAGE, "JPEG datastream contains no image")
+JMESSAGE(JERR_NO_QUANT_TABLE, "Quantization table 0x%02x was not defined")
+JMESSAGE(JERR_NO_SOI, "Not a JPEG file: starts with 0x%02x 0x%02x")
+JMESSAGE(JERR_OUT_OF_MEMORY, "Insufficient memory (case %d)")
+JMESSAGE(JERR_QUANT_COMPONENTS,
+	 "Cannot quantize more than %d color components")
+JMESSAGE(JERR_QUANT_FEW_COLORS, "Cannot quantize to fewer than %d colors")
+JMESSAGE(JERR_QUANT_MANY_COLORS, "Cannot quantize to more than %d colors")
+JMESSAGE(JERR_SOF_DUPLICATE, "Invalid JPEG file structure: two SOF markers")
+JMESSAGE(JERR_SOF_NO_SOS, "Invalid JPEG file structure: missing SOS marker")
+JMESSAGE(JERR_SOF_UNSUPPORTED, "Unsupported JPEG process: SOF type 0x%02x")
+JMESSAGE(JERR_SOI_DUPLICATE, "Invalid JPEG file structure: two SOI markers")
+JMESSAGE(JERR_SOS_NO_SOF, "Invalid JPEG file structure: SOS before SOF")
+JMESSAGE(JERR_TFILE_CREATE, "Failed to create temporary file %s")
+JMESSAGE(JERR_TFILE_READ, "Read failed on temporary file")
+JMESSAGE(JERR_TFILE_SEEK, "Seek failed on temporary file")
+JMESSAGE(JERR_TFILE_WRITE,
+	 "Write failed on temporary file --- out of disk space?")
+JMESSAGE(JERR_TOO_LITTLE_DATA, "Application transferred too few scanlines")
+JMESSAGE(JERR_UNKNOWN_MARKER, "Unsupported marker type 0x%02x")
+JMESSAGE(JERR_VIRTUAL_BUG, "Virtual array controller messed up")
+JMESSAGE(JERR_WIDTH_OVERFLOW, "Image too wide for this implementation")
+JMESSAGE(JERR_XMS_READ, "Read from XMS failed")
+JMESSAGE(JERR_XMS_WRITE, "Write to XMS failed")
+JMESSAGE(JMSG_COPYRIGHT, JCOPYRIGHT)
+JMESSAGE(JMSG_VERSION, JVERSION)
+JMESSAGE(JTRC_16BIT_TABLES,
+	 "Caution: quantization tables are too coarse for baseline JPEG")
+JMESSAGE(JTRC_ADOBE,
+	 "Adobe APP14 marker: version %d, flags 0x%04x 0x%04x, transform %d")
+JMESSAGE(JTRC_APP0, "Unknown APP0 marker (not JFIF), length %u")
+JMESSAGE(JTRC_APP14, "Unknown APP14 marker (not Adobe), length %u")
+JMESSAGE(JTRC_DAC, "Define Arithmetic Table 0x%02x: 0x%02x")
+JMESSAGE(JTRC_DHT, "Define Huffman Table 0x%02x")
+JMESSAGE(JTRC_DQT, "Define Quantization Table %d  precision %d")
+JMESSAGE(JTRC_DRI, "Define Restart Interval %u")
+JMESSAGE(JTRC_EMS_CLOSE, "Freed EMS handle %u")
+JMESSAGE(JTRC_EMS_OPEN, "Obtained EMS handle %u")
+JMESSAGE(JTRC_EOI, "End Of Image")
+JMESSAGE(JTRC_HUFFBITS, "        %3d %3d %3d %3d %3d %3d %3d %3d")
+JMESSAGE(JTRC_JFIF, "JFIF APP0 marker: version %d.%02d, density %dx%d  %d")
+JMESSAGE(JTRC_JFIF_BADTHUMBNAILSIZE,
+	 "Warning: thumbnail image size does not match data length %u")
+JMESSAGE(JTRC_JFIF_EXTENSION,
+	 "JFIF extension marker: type 0x%02x, length %u")
+JMESSAGE(JTRC_JFIF_THUMBNAIL, "    with %d x %d thumbnail image")
+JMESSAGE(JTRC_MISC_MARKER, "Miscellaneous marker 0x%02x, length %u")
+JMESSAGE(JTRC_PARMLESS_MARKER, "Unexpected marker 0x%02x")
+JMESSAGE(JTRC_QUANTVALS, "        %4u %4u %4u %4u %4u %4u %4u %4u")
+JMESSAGE(JTRC_QUANT_3_NCOLORS, "Quantizing to %d = %d*%d*%d colors")
+JMESSAGE(JTRC_QUANT_NCOLORS, "Quantizing to %d colors")
+JMESSAGE(JTRC_QUANT_SELECTED, "Selected %d colors for quantization")
+JMESSAGE(JTRC_RECOVERY_ACTION, "At marker 0x%02x, recovery action %d")
+JMESSAGE(JTRC_RST, "RST%d")
+JMESSAGE(JTRC_SMOOTH_NOTIMPL,
+	 "Smoothing not supported with nonstandard sampling ratios")
+JMESSAGE(JTRC_SOF, "Start Of Frame 0x%02x: width=%u, height=%u, components=%d")
+JMESSAGE(JTRC_SOF_COMPONENT, "    Component %d: %dhx%dv q=%d")
+JMESSAGE(JTRC_SOI, "Start of Image")
+JMESSAGE(JTRC_SOS, "Start Of Scan: %d components")
+JMESSAGE(JTRC_SOS_COMPONENT, "    Component %d: dc=%d ac=%d")
+JMESSAGE(JTRC_SOS_PARAMS, "  Ss=%d, Se=%d, Ah=%d, Al=%d")
+JMESSAGE(JTRC_TFILE_CLOSE, "Closed temporary file %s")
+JMESSAGE(JTRC_TFILE_OPEN, "Opened temporary file %s")
+JMESSAGE(JTRC_THUMB_JPEG,
+	 "JFIF extension marker: JPEG-compressed thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_PALETTE,
+	 "JFIF extension marker: palette thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_RGB,
+	 "JFIF extension marker: RGB thumbnail image, length %u")
+JMESSAGE(JTRC_UNKNOWN_IDS,
+	 "Unrecognized component IDs %d %d %d, assuming YCbCr")
+JMESSAGE(JTRC_XMS_CLOSE, "Freed XMS handle %u")
+JMESSAGE(JTRC_XMS_OPEN, "Obtained XMS handle %u")
+JMESSAGE(JWRN_ADOBE_XFORM, "Unknown Adobe color transform code %d")
+JMESSAGE(JWRN_BOGUS_PROGRESSION,
+	 "Inconsistent progression sequence for component %d coefficient %d")
+JMESSAGE(JWRN_EXTRANEOUS_DATA,
+	 "Corrupt JPEG data: %u extraneous bytes before marker 0x%02x")
+JMESSAGE(JWRN_HIT_MARKER, "Corrupt JPEG data: premature end of data segment")
+JMESSAGE(JWRN_HUFF_BAD_CODE, "Corrupt JPEG data: bad Huffman code")
+JMESSAGE(JWRN_JFIF_MAJOR, "Warning: unknown JFIF revision number %d.%02d")
+JMESSAGE(JWRN_JPEG_EOF, "Premature end of JPEG file")
+JMESSAGE(JWRN_MUST_RESYNC,
+	 "Corrupt JPEG data: found marker 0x%02x instead of RST%d")
+JMESSAGE(JWRN_NOT_SEQUENTIAL, "Invalid SOS parameters for sequential JPEG")
+JMESSAGE(JWRN_TOO_MUCH_DATA, "Application transferred too many scanlines")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTMSGCODE
+} J_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE
+
+
+#ifndef JERROR_H
+#define JERROR_H
+
+/* Macros to simplify using the error and trace message stuff */
+/* The first parameter is either type of cinfo pointer */
+
+/* Fatal errors (print message and exit) */
+#define ERREXIT(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT3(cinfo,code,p1,p2,p3)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT4(cinfo,code,p1,p2,p3,p4)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXITS(cinfo,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+
+#define MAKESTMT(stuff)		do { stuff } while (0)
+
+/* Nonfatal errors (we can keep going, but the data is probably corrupt) */
+#define WARNMS(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+
+/* Informational/debugging messages */
+#define TRACEMS(cinfo,lvl,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS1(cinfo,lvl,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS2(cinfo,lvl,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS3(cinfo,lvl,code,p1,p2,p3)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS4(cinfo,lvl,code,p1,p2,p3,p4)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS5(cinfo,lvl,code,p1,p2,p3,p4,p5)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS8(cinfo,lvl,code,p1,p2,p3,p4,p5,p6,p7,p8)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); _mp[5] = (p6); _mp[6] = (p7); _mp[7] = (p8); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMSS(cinfo,lvl,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+
+#endif /* JERROR_H */
diff --git a/ml/dlib/dlib/external/libjpeg/jfdctflt.cpp b/ml/dlib/dlib/external/libjpeg/jfdctflt.cpp
new file mode 100644
index 000000000..79d7a0078
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jfdctflt.cpp
@@ -0,0 +1,168 @@
+/*
+ * jfdctflt.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * This implementation should be more accurate than either of the integer
+ * DCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_float (FAST_FLOAT * data)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
+  FAST_FLOAT *dataptr;
+  int ctr;
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+    
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+    
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jfdctfst.cpp b/ml/dlib/dlib/external/libjpeg/jfdctfst.cpp
new file mode 100644
index 000000000..4c96d4f3d
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jfdctfst.cpp
@@ -0,0 +1,224 @@
+/*
+ * jfdctfst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jfdctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * Again to save a few shifts, the intermediate results between pass 1 and
+ * pass 2 are not upscaled, but are represented only to integral precision.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#define CONST_BITS  8
+
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_0_382683433  ((long)   98)		/* FIX(0.382683433) */
+#define FIX_0_541196100  ((long)  139)		/* FIX(0.541196100) */
+#define FIX_0_707106781  ((long)  181)		/* FIX(0.707106781) */
+#define FIX_1_306562965  ((long)  334)		/* FIX(1.306562965) */
+#else
+#define FIX_0_382683433  FIX(0.382683433)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_707106781  FIX(0.707106781)
+#define FIX_1_306562965  FIX(1.306562965)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an long constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_ifast (DCTELEM * data)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z1, z2, z3, z4, z5, z11, z13;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part */
+    
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+    
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jfdctint.cpp b/ml/dlib/dlib/external/libjpeg/jfdctint.cpp
new file mode 100644
index 000000000..b6046a26b
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jfdctint.cpp
@@ -0,0 +1,283 @@
+/*
+ * jfdctint.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true DCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D DCT,
+ * because the y0 and y4 outputs need not be divided by sqrt(N).
+ * In the IJG code, this factor of 8 is removed by the quantization step
+ * (in jcdctmgr.c), NOT in this module.
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (For 12-bit sample data, the intermediate
+ * array is long anyway.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((long)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((long)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((long)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((long)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((long)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((long)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((long)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((long)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((long)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((long)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((long)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((long)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an long variable by an long constant to yield an long result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_islow (DCTELEM * data)
+{
+  long tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  long tmp10, tmp11, tmp12, tmp13;
+  long z1, z2, z3, z4, z5;
+  DCTELEM *dataptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[0] + dataptr[7];
+    tmp7 = dataptr[0] - dataptr[7];
+    tmp1 = dataptr[1] + dataptr[6];
+    tmp6 = dataptr[1] - dataptr[6];
+    tmp2 = dataptr[2] + dataptr[5];
+    tmp5 = dataptr[2] - dataptr[5];
+    tmp3 = dataptr[3] + dataptr[4];
+    tmp4 = dataptr[3] - dataptr[4];
+    
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
+    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+				   CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
+				   CONST_BITS-PASS1_BITS);
+    
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+    
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
+    dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
+    
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+    
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS);
+    
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
+					   CONST_BITS+PASS1_BITS);
+    
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents cos(K*pi/16).
+     * i0..i3 in the paper are tmp4..tmp7 here.
+     */
+    
+    z1 = tmp4 + tmp7;
+    z2 = tmp5 + tmp6;
+    z3 = tmp4 + tmp6;
+    z4 = tmp5 + tmp7;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp4 + z1 + z3,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp5 + z2 + z4,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp6 + z2 + z3,
+					   CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp7 + z1 + z4,
+					   CONST_BITS+PASS1_BITS);
+    
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_ISLOW_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jidctflt.cpp b/ml/dlib/dlib/external/libjpeg/jidctflt.cpp
new file mode 100644
index 000000000..3e9b54579
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jidctflt.cpp
@@ -0,0 +1,242 @@
+/*
+ * jidctflt.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * This implementation should be more accurate than either of the integer
+ * IDCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a float result.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  FLOAT_MULT_TYPE * quantptr;
+  FAST_FLOAT * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
+    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE*0] = tmp0 + tmp7;
+    wsptr[DCTSIZE*7] = tmp0 - tmp7;
+    wsptr[DCTSIZE*1] = tmp1 + tmp6;
+    wsptr[DCTSIZE*6] = tmp1 - tmp6;
+    wsptr[DCTSIZE*2] = tmp2 + tmp5;
+    wsptr[DCTSIZE*5] = tmp2 - tmp5;
+    wsptr[DCTSIZE*4] = tmp3 + tmp4;
+    wsptr[DCTSIZE*3] = tmp3 - tmp4;
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * And testing floats for zero is relatively expensive, so we don't bother.
+     */
+    
+    /* Even part */
+
+    tmp10 = wsptr[0] + wsptr[4];
+    tmp11 = wsptr[0] - wsptr[4];
+
+    tmp13 = wsptr[2] + wsptr[6];
+    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = wsptr[5] + wsptr[3];
+    z10 = wsptr[5] - wsptr[3];
+    z11 = wsptr[1] + wsptr[7];
+    z12 = wsptr[1] - wsptr[7];
+
+    tmp7 = z11 + z13;
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */
+    tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] = range_limit[(int) DESCALE((long) (tmp0 + tmp7), 3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) DESCALE((long) (tmp0 - tmp7), 3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE((long) (tmp1 + tmp6), 3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) DESCALE((long) (tmp1 - tmp6), 3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE((long) (tmp2 + tmp5), 3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) DESCALE((long) (tmp2 - tmp5), 3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) DESCALE((long) (tmp3 + tmp4), 3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE((long) (tmp3 - tmp4), 3)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jidctfst.cpp b/ml/dlib/dlib/external/libjpeg/jidctfst.cpp
new file mode 100644
index 000000000..e08835bc5
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jidctfst.cpp
@@ -0,0 +1,368 @@
+/*
+ * jidctfst.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jidctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * The dequantized coefficients are not integers because the AA&N scaling
+ * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
+ * so that the first and second IDCT rounds have the same input scaling.
+ * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
+ * avoid a descaling shift; this compromises accuracy rather drastically
+ * for small quantization table entries, but it saves a lot of shifts.
+ * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
+ * so we use a much larger scaling factor to preserve accuracy.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  8
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  8
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_1_082392200  ((long)  277)		/* FIX(1.082392200) */
+#define FIX_1_414213562  ((long)  362)		/* FIX(1.414213562) */
+#define FIX_1_847759065  ((long)  473)		/* FIX(1.847759065) */
+#define FIX_2_613125930  ((long)  669)		/* FIX(2.613125930) */
+#else
+#define FIX_1_082392200  FIX(1.082392200)
+#define FIX_1_414213562  FIX(1.414213562)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_613125930  FIX(2.613125930)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an long constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
+ * multiplication will do.  For 12-bit data, the multiplier table is
+ * declared long, so a 32-bit multiply will be used.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define DEQUANTIZE(coef,quantval)  (((IFAST_MULT_TYPE) (coef)) * (quantval))
+#else
+#define DEQUANTIZE(coef,quantval)  \
+	DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
+#endif
+
+
+/* Like DESCALE, but applies to a DCTELEM and produces an int.
+ * We assume that int right shift is unsigned if long right shift is.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	DCTELEM ishift_temp;
+#if BITS_IN_JSAMPLE == 8
+#define DCTELEMBITS  16		/* DCTELEM may be 16 or 32 bits */
+#else
+#define DCTELEMBITS  32		/* DCTELEM must be 32 bits */
+#endif
+#define IRIGHT_SHIFT(x,shft)  \
+    ((ishift_temp = (x)) < 0 ? \
+     (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
+     (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+#ifdef USE_ACCURATE_ROUNDING
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
+#else
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT(x, n))
+#endif
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  IFAST_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS			/* for DESCALE */
+  ISHIFT_TEMPS			/* for IDESCALE */
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
+    wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
+    wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
+    wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
+    wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
+    wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
+    wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
+    wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+    
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+
+    tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
+    tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
+
+    tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
+    tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
+	    - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
+    z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
+    z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
+    z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jidctint.cpp b/ml/dlib/dlib/external/libjpeg/jidctint.cpp
new file mode 100644
index 000000000..630b4fb89
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jidctint.cpp
@@ -0,0 +1,389 @@
+/*
+ * jidctint.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true IDCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D IDCT,
+ * because the y0 and y4 inputs need not be divided by sqrt(N).
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (To scale up 12-bit sample data further, an
+ * intermediate long array would be needed.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((long)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((long)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((long)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((long)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((long)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((long)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((long)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((long)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((long)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((long)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((long)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((long)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an long variable by an long constant to yield an long result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  long tmp0, tmp1, tmp2, tmp3;
+  long tmp10, tmp11, tmp12, tmp13;
+  long z1, z2, z3, z4, z5;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+  /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+
+    tmp0 = (z2 + z3) << CONST_BITS;
+    tmp1 = (z2 - z3) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+    
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+    
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
+    
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+    
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((long) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = (long) wsptr[2];
+    z3 = (long) wsptr[6];
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065);
+    tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    
+    tmp0 = ((long) wsptr[0] + (long) wsptr[4]) << CONST_BITS;
+    tmp1 = ((long) wsptr[0] - (long) wsptr[4]) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp3;
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+    
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+    
+    tmp0 = (long) wsptr[7];
+    tmp1 = (long) wsptr[5];
+    tmp2 = (long) wsptr[3];
+    tmp3 = (long) wsptr[1];
+    
+    z1 = tmp0 + tmp3;
+    z2 = tmp1 + tmp2;
+    z3 = tmp0 + tmp2;
+    z4 = tmp1 + tmp3;
+    z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+    
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    
+    z3 += z5;
+    z4 += z5;
+    
+    tmp0 += z1 + z3;
+    tmp1 += z2 + z4;
+    tmp2 += z2 + z3;
+    tmp3 += z1 + z4;
+    
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp3,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) DESCALE(tmp10 - tmp3,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp11 + tmp2,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) DESCALE(tmp11 - tmp2,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE(tmp12 + tmp1,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) DESCALE(tmp12 - tmp1,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE(tmp13 + tmp0,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) DESCALE(tmp13 - tmp0,
+					  CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_ISLOW_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jidctred.cpp b/ml/dlib/dlib/external/libjpeg/jidctred.cpp
new file mode 100644
index 000000000..fa442ac9a
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jidctred.cpp
@@ -0,0 +1,398 @@
+/*
+ * jidctred.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains inverse-DCT routines that produce reduced-size output:
+ * either 4x4, 2x2, or 1x1 pixels from an 8x8 DCT block.
+ *
+ * The implementation is based on the Loeffler, Ligtenberg and Moschytz (LL&M)
+ * algorithm used in jidctint.c.  We simply replace each 8-to-8 1-D IDCT step
+ * with an 8-to-4 step that produces the four averages of two adjacent outputs
+ * (or an 8-to-2 step producing two averages of four outputs, for 2x2 output).
+ * These steps were derived by computing the corresponding values at the end
+ * of the normal LL&M code, then simplifying as much as possible.
+ *
+ * 1x1 is trivial: just take the DC coefficient divided by 8.
+ *
+ * See jidctint.c for additional comments.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling is the same as in jidctint.c. */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_211164243  ((long)  1730)	/* FIX(0.211164243) */
+#define FIX_0_509795579  ((long)  4176)	/* FIX(0.509795579) */
+#define FIX_0_601344887  ((long)  4926)	/* FIX(0.601344887) */
+#define FIX_0_720959822  ((long)  5906)	/* FIX(0.720959822) */
+#define FIX_0_765366865  ((long)  6270)	/* FIX(0.765366865) */
+#define FIX_0_850430095  ((long)  6967)	/* FIX(0.850430095) */
+#define FIX_0_899976223  ((long)  7373)	/* FIX(0.899976223) */
+#define FIX_1_061594337  ((long)  8697)	/* FIX(1.061594337) */
+#define FIX_1_272758580  ((long)  10426)	/* FIX(1.272758580) */
+#define FIX_1_451774981  ((long)  11893)	/* FIX(1.451774981) */
+#define FIX_1_847759065  ((long)  15137)	/* FIX(1.847759065) */
+#define FIX_2_172734803  ((long)  17799)	/* FIX(2.172734803) */
+#define FIX_2_562915447  ((long)  20995)	/* FIX(2.562915447) */
+#define FIX_3_624509785  ((long)  29692)	/* FIX(3.624509785) */
+#else
+#define FIX_0_211164243  FIX(0.211164243)
+#define FIX_0_509795579  FIX(0.509795579)
+#define FIX_0_601344887  FIX(0.601344887)
+#define FIX_0_720959822  FIX(0.720959822)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_850430095  FIX(0.850430095)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_061594337  FIX(1.061594337)
+#define FIX_1_272758580  FIX(1.272758580)
+#define FIX_1_451774981  FIX(1.451774981)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_172734803  FIX(2.172734803)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_624509785  FIX(3.624509785)
+#endif
+
+
+/* Multiply an long variable by an long constant to yield an long result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 4x4 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  long tmp0, tmp2, tmp10, tmp12;
+  long z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE*4];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
+    /* Don't bother to process column 4, because second pass won't use it */
+    if (ctr == DCTSIZE-4)
+      continue;
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*5] == 0 &&
+	inptr[DCTSIZE*6] == 0 && inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero; we need not examine term 4 for 4x4 output */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      
+      continue;
+    }
+    
+    /* Even part */
+    
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= (CONST_BITS+1);
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp2 = MULTIPLY(z2, FIX_1_847759065) + MULTIPLY(z3, - FIX_0_765366865);
+    
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+    
+    /* Odd part */
+    
+    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
+	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
+	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
+	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
+    
+    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
+	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
+	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
+	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
+
+    /* Final output stage */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp2, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*3] = (int) DESCALE(tmp10 - tmp2, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp12 + tmp0, CONST_BITS-PASS1_BITS+1);
+    wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 - tmp0, CONST_BITS-PASS1_BITS+1);
+  }
+  
+  /* Pass 2: process 4 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 4; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* It's not clear whether a zero row test is worthwhile here ... */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((long) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+    
+    tmp0 = ((long) wsptr[0]) << (CONST_BITS+1);
+    
+    tmp2 = MULTIPLY((long) wsptr[2], FIX_1_847759065)
+	 + MULTIPLY((long) wsptr[6], - FIX_0_765366865);
+    
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+    
+    /* Odd part */
+    
+    z1 = (long) wsptr[7];
+    z2 = (long) wsptr[5];
+    z3 = (long) wsptr[3];
+    z4 = (long) wsptr[1];
+    
+    tmp0 = MULTIPLY(z1, - FIX_0_211164243) /* sqrt(2) * (c3-c1) */
+	 + MULTIPLY(z2, FIX_1_451774981) /* sqrt(2) * (c3+c7) */
+	 + MULTIPLY(z3, - FIX_2_172734803) /* sqrt(2) * (-c1-c5) */
+	 + MULTIPLY(z4, FIX_1_061594337); /* sqrt(2) * (c5+c7) */
+    
+    tmp2 = MULTIPLY(z1, - FIX_0_509795579) /* sqrt(2) * (c7-c5) */
+	 + MULTIPLY(z2, - FIX_0_601344887) /* sqrt(2) * (c5-c1) */
+	 + MULTIPLY(z3, FIX_0_899976223) /* sqrt(2) * (c3-c7) */
+	 + MULTIPLY(z4, FIX_2_562915447); /* sqrt(2) * (c1+c3) */
+
+    /* Final output stage */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp2,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) DESCALE(tmp10 - tmp2,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp12 + tmp0,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) DESCALE(tmp12 - tmp0,
+					  CONST_BITS+PASS1_BITS+3+1)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 2x2 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  long tmp0, tmp10, z1;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE*2];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; inptr++, quantptr++, wsptr++, ctr--) {
+    /* Don't bother to process columns 2,4,6 */
+    if (ctr == DCTSIZE-2 || ctr == DCTSIZE-4 || ctr == DCTSIZE-6)
+      continue;
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*3] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero; we need not examine terms 2,4,6 for 2x2 output */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      
+      continue;
+    }
+    
+    /* Even part */
+    
+    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp10 = z1 << (CONST_BITS+2);
+    
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp0 = MULTIPLY(z1, - FIX_0_720959822); /* sqrt(2) * (c7-c5+c3-c1) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp0 += MULTIPLY(z1, FIX_0_850430095); /* sqrt(2) * (-c1+c3+c5+c7) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp0 += MULTIPLY(z1, - FIX_1_272758580); /* sqrt(2) * (-c1+c3-c5-c7) */
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp0 += MULTIPLY(z1, FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
+
+    /* Final output stage */
+    
+    wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp0, CONST_BITS-PASS1_BITS+2);
+    wsptr[DCTSIZE*1] = (int) DESCALE(tmp10 - tmp0, CONST_BITS-PASS1_BITS+2);
+  }
+  
+  /* Pass 2: process 2 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 2; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* It's not clear whether a zero row test is worthwhile here ... */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[3] == 0 && wsptr[5] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((long) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+    
+    tmp10 = ((long) wsptr[0]) << (CONST_BITS+2);
+    
+    /* Odd part */
+
+    tmp0 = MULTIPLY((long) wsptr[7], - FIX_0_720959822) /* sqrt(2) * (c7-c5+c3-c1) */
+	 + MULTIPLY((long) wsptr[5], FIX_0_850430095) /* sqrt(2) * (-c1+c3+c5+c7) */
+	 + MULTIPLY((long) wsptr[3], - FIX_1_272758580) /* sqrt(2) * (-c1+c3-c5-c7) */
+	 + MULTIPLY((long) wsptr[1], FIX_3_624509785); /* sqrt(2) * (c1+c3+c5+c7) */
+
+    /* Final output stage */
+    
+    outptr[0] = range_limit[(int) DESCALE(tmp10 + tmp0,
+					  CONST_BITS+PASS1_BITS+3+2)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) DESCALE(tmp10 - tmp0,
+					  CONST_BITS+PASS1_BITS+3+2)
+			    & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 1x1 output block.
+ */
+
+GLOBAL(void)
+jpeg_idct_1x1 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  int dcval;
+  ISLOW_MULT_TYPE * quantptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  SHIFT_TEMPS
+
+  /* We hardly need an inverse DCT routine for this: just take the
+   * average pixel value, which is one-eighth of the DC coefficient.
+   */
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  dcval = DEQUANTIZE(coef_block[0], quantptr[0]);
+  dcval = (int) DESCALE((long) dcval, 3);
+
+  output_buf[0][output_col] = range_limit[dcval & RANGE_MASK];
+}
+
+#endif /* IDCT_SCALING_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jinclude.h b/ml/dlib/dlib/external/libjpeg/jinclude.h
new file mode 100644
index 000000000..0a4f15146
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jinclude.h
@@ -0,0 +1,91 @@
+/*
+ * jinclude.h
+ *
+ * Copyright (C) 1991-1994, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file exists to provide a single place to fix any problems with
+ * including the wrong system include files.  (Common problems are taken
+ * care of by the standard jconfig symbols, but on really weird systems
+ * you may have to edit this file.)
+ *
+ * NOTE: this file is NOT intended to be included by applications using the
+ * JPEG library.  Most applications need only include jpeglib.h.
+ */
+
+
+/* Include auto-config file to find out which system include files we need. */
+
+#include "jconfig.h"		/* auto configuration options */
+#define JCONFIG_INCLUDED	/* so that jpeglib.h doesn't do it again */
+
+/*
+ * We need the NULL macro and size_t typedef.
+ * On an ANSI-conforming system it is sufficient to include <stddef.h>.
+ * Otherwise, we get them from <stdlib.h> or <stdio.h>; we may have to
+ * pull in <sys/types.h> as well.
+ * Note that the core JPEG library does not require <stdio.h>;
+ * only the default error handler and data source/destination modules do.
+ * But we must pull it in because of the references to FILE in jpeglib.h.
+ * You can remove those references if you want to compile without <stdio.h>.
+ */
+
+#ifdef HAVE_STDDEF_H
+#include <stddef.h>
+#endif
+
+#ifdef HAVE_STDLIB_H
+#include <stdlib.h>
+#endif
+
+#ifdef NEED_SYS_TYPES_H
+#include <sys/types.h>
+#endif
+
+#include <stdio.h>
+
+/*
+ * We need memory copying and zeroing functions, plus strncpy().
+ * ANSI and System V implementations declare these in <string.h>.
+ * BSD doesn't have the mem() functions, but it does have bcopy()/bzero().
+ * Some systems may declare memset and memcpy in <memory.h>.
+ *
+ * NOTE: we assume the size parameters to these functions are of type size_t.
+ * Change the casts in these macros if not!
+ */
+
+#ifdef NEED_BSD_STRINGS
+
+#include <strings.h>
+#define MEMZERO(target,size)	bzero((void *)(target), (size_t)(size))
+#define MEMCOPY(dest,src,size)	bcopy((const void *)(src), (void *)(dest), (size_t)(size))
+
+#else /* not BSD, assume ANSI/SysV string lib */
+
+#include <string.h>
+#define MEMZERO(target,size)	memset((void *)(target), 0, (size_t)(size))
+#define MEMCOPY(dest,src,size)	memcpy((void *)(dest), (const void *)(src), (size_t)(size))
+
+#endif
+
+/*
+ * In ANSI C, and indeed any rational implementation, size_t is also the
+ * type returned by sizeof().  However, it seems there are some irrational
+ * implementations out there, in which sizeof() returns an int even though
+ * size_t is defined as long or unsigned long.  To ensure consistent results
+ * we always use this SIZEOF() macro in place of using sizeof() directly.
+ */
+
+#define SIZEOF(object)	((size_t) sizeof(object))
+
+/*
+ * The modules that use fread() and fwrite() always invoke them through
+ * these macros.  On some systems you may need to twiddle the argument casts.
+ * CAUTION: argument order is different from underlying functions!
+ */
+
+#define JFREAD(file,buf,sizeofbuf)  \
+  ((size_t) fread((void *) (buf), (size_t) 1, (size_t) (sizeofbuf), (file)))
+#define JFWRITE(file,buf,sizeofbuf)  \
+  ((size_t) fwrite((const void *) (buf), (size_t) 1, (size_t) (sizeofbuf), (file)))
diff --git a/ml/dlib/dlib/external/libjpeg/jmemmgr.cpp b/ml/dlib/dlib/external/libjpeg/jmemmgr.cpp
new file mode 100644
index 000000000..3a2e61955
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jmemmgr.cpp
@@ -0,0 +1,1118 @@
+/*
+ * jmemmgr.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the JPEG system-independent memory management
+ * routines.  This code is usable across a wide variety of machines; most
+ * of the system dependencies have been isolated in a separate file.
+ * The major functions provided here are:
+ *   * pool-based allocation and freeing of memory;
+ *   * policy decisions about how to divide available memory among the
+ *     virtual arrays;
+ *   * control logic for swapping virtual arrays between main memory and
+ *     backing storage.
+ * The separate system-dependent file provides the actual backing-storage
+ * access code, and it contains the policy decision about how much total
+ * main memory to use.
+ * This file is system-dependent in the sense that some of its functions
+ * are unnecessary in some systems.  For example, if there is enough virtual
+ * memory so that backing storage will never be used, much of the virtual
+ * array control logic could be removed.  (Of course, if you have that much
+ * memory then you shouldn't care about a little bit of unused code...)
+ */
+
+#define JPEG_INTERNALS
+#define AM_MEMORY_MANAGER	/* we define jvirt_Xarray_control structs */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jmemsys.h"		/* import the system-dependent declarations */
+
+#ifndef NO_GETENV
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare getenv() */
+extern char * getenv JPP((const char * name));
+#endif
+#endif
+
+
+/*
+ * Some important notes:
+ *   The allocation routines provided here must never return NULL.
+ *   They should exit to error_exit if unsuccessful.
+ *
+ *   It's not a good idea to try to merge the sarray and barray routines,
+ *   even though they are textually almost the same, because samples are
+ *   usually stored as bytes while coefficients are shorts or ints.  Thus,
+ *   in machines where byte pointers have a different representation from
+ *   word pointers, the resulting machine code could not be the same.
+ */
+
+
+/*
+ * Many machines require storage alignment: longs must start on 4-byte
+ * boundaries, doubles on 8-byte boundaries, etc.  On such machines, malloc()
+ * always returns pointers that are multiples of the worst-case alignment
+ * requirement, and we had better do so too.
+ * There isn't any really portable way to determine the worst-case alignment
+ * requirement.  This module assumes that the alignment requirement is
+ * multiples of sizeof(ALIGN_TYPE).
+ * By default, we define ALIGN_TYPE as double.  This is necessary on some
+ * workstations (where doubles really do need 8-byte alignment) and will work
+ * fine on nearly everything.  If your machine has lesser alignment needs,
+ * you can save a few bytes by making ALIGN_TYPE smaller.
+ * The only place I know of where this will NOT work is certain Macintosh
+ * 680x0 compilers that define double as a 10-byte IEEE extended float.
+ * Doing 10-byte alignment is counterproductive because longwords won't be
+ * aligned well.  Put "#define ALIGN_TYPE long" in jconfig.h if you have
+ * such a compiler.
+ */
+
+#ifndef ALIGN_TYPE		/* so can override from jconfig.h */
+#define ALIGN_TYPE  double
+#endif
+
+
+/*
+ * We allocate objects from "pools", where each pool is gotten with a single
+ * request to jpeg_get_small() or jpeg_get_large().  There is no per-object
+ * overhead within a pool, except for alignment padding.  Each pool has a
+ * header with a link to the next pool of the same class.
+ * Small and large pool headers are identical except that the latter's
+ * link pointer must be FAR on 80x86 machines.
+ * Notice that the "real" header fields are union'ed with a dummy ALIGN_TYPE
+ * field.  This forces the compiler to make SIZEOF(small_pool_hdr) a multiple
+ * of the alignment requirement of ALIGN_TYPE.
+ */
+
+typedef union small_pool_struct * small_pool_ptr;
+
+typedef union small_pool_struct {
+  struct {
+    small_pool_ptr next;	/* next in list of pools */
+    size_t bytes_used;		/* how many bytes already used within pool */
+    size_t bytes_left;		/* bytes still available in this pool */
+  } hdr;
+  ALIGN_TYPE dummy;		/* included in union to ensure alignment */
+} small_pool_hdr;
+
+typedef union large_pool_struct FAR * large_pool_ptr;
+
+typedef union large_pool_struct {
+  struct {
+    large_pool_ptr next;	/* next in list of pools */
+    size_t bytes_used;		/* how many bytes already used within pool */
+    size_t bytes_left;		/* bytes still available in this pool */
+  } hdr;
+  ALIGN_TYPE dummy;		/* included in union to ensure alignment */
+} large_pool_hdr;
+
+
+/*
+ * Here is the full definition of a memory manager object.
+ */
+
+typedef struct {
+  struct jpeg_memory_mgr pub;	/* public fields */
+
+  /* Each pool identifier (lifetime class) names a linked list of pools. */
+  small_pool_ptr small_list[JPOOL_NUMPOOLS];
+  large_pool_ptr large_list[JPOOL_NUMPOOLS];
+
+  /* Since we only have one lifetime class of virtual arrays, only one
+   * linked list is necessary (for each datatype).  Note that the virtual
+   * array control blocks being linked together are actually stored somewhere
+   * in the small-pool list.
+   */
+  jvirt_sarray_ptr virt_sarray_list;
+  jvirt_barray_ptr virt_barray_list;
+
+  /* This counts total space obtained from jpeg_get_small/large */
+  long total_space_allocated;
+
+  /* alloc_sarray and alloc_barray set this value for use by virtual
+   * array routines.
+   */
+  JDIMENSION last_rowsperchunk;	/* from most recent alloc_sarray/barray */
+} my_memory_mgr;
+
+typedef my_memory_mgr * my_mem_ptr;
+
+
+/*
+ * The control blocks for virtual arrays.
+ * Note that these blocks are allocated in the "small" pool area.
+ * System-dependent info for the associated backing store (if any) is hidden
+ * inside the backing_store_info struct.
+ */
+
+struct jvirt_sarray_control {
+  JSAMPARRAY mem_buffer;	/* => the in-memory buffer */
+  JDIMENSION rows_in_array;	/* total virtual array height */
+  JDIMENSION samplesperrow;	/* width of array (and of memory buffer) */
+  JDIMENSION maxaccess;		/* max rows accessed by access_virt_sarray */
+  JDIMENSION rows_in_mem;	/* height of memory buffer */
+  JDIMENSION rowsperchunk;	/* allocation chunk size in mem_buffer */
+  JDIMENSION cur_start_row;	/* first logical row # in the buffer */
+  JDIMENSION first_undef_row;	/* row # of first uninitialized row */
+  int pre_zero;		/* pre-zero mode requested? */
+  int dirty;		/* do current buffer contents need written? */
+  int b_s_open;		/* is backing-store data valid? */
+  jvirt_sarray_ptr next;	/* link to next virtual sarray control block */
+  backing_store_info b_s_info;	/* System-dependent control info */
+};
+
+struct jvirt_barray_control {
+  JBLOCKARRAY mem_buffer;	/* => the in-memory buffer */
+  JDIMENSION rows_in_array;	/* total virtual array height */
+  JDIMENSION blocksperrow;	/* width of array (and of memory buffer) */
+  JDIMENSION maxaccess;		/* max rows accessed by access_virt_barray */
+  JDIMENSION rows_in_mem;	/* height of memory buffer */
+  JDIMENSION rowsperchunk;	/* allocation chunk size in mem_buffer */
+  JDIMENSION cur_start_row;	/* first logical row # in the buffer */
+  JDIMENSION first_undef_row;	/* row # of first uninitialized row */
+  int pre_zero;		/* pre-zero mode requested? */
+  int dirty;		/* do current buffer contents need written? */
+  int b_s_open;		/* is backing-store data valid? */
+  jvirt_barray_ptr next;	/* link to next virtual barray control block */
+  backing_store_info b_s_info;	/* System-dependent control info */
+};
+
+
+#ifdef MEM_STATS		/* optional extra stuff for statistics */
+
+LOCAL(void)
+print_mem_stats (j_common_ptr cinfo, int pool_id)
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr shdr_ptr;
+  large_pool_ptr lhdr_ptr;
+
+  /* Since this is only a debugging stub, we can cheat a little by using
+   * fprintf directly rather than going through the trace message code.
+   * This is helpful because message parm array can't handle longs.
+   */
+  fprintf(stderr, "Freeing pool %d, total space = %ld\n",
+	  pool_id, mem->total_space_allocated);
+
+  for (lhdr_ptr = mem->large_list[pool_id]; lhdr_ptr != NULL;
+       lhdr_ptr = lhdr_ptr->hdr.next) {
+    fprintf(stderr, "  Large chunk used %ld\n",
+	    (long) lhdr_ptr->hdr.bytes_used);
+  }
+
+  for (shdr_ptr = mem->small_list[pool_id]; shdr_ptr != NULL;
+       shdr_ptr = shdr_ptr->hdr.next) {
+    fprintf(stderr, "  Small chunk used %ld free %ld\n",
+	    (long) shdr_ptr->hdr.bytes_used,
+	    (long) shdr_ptr->hdr.bytes_left);
+  }
+}
+
+#endif /* MEM_STATS */
+
+
+LOCAL(void)
+out_of_memory (j_common_ptr cinfo, int which)
+/* Report an out-of-memory error and stop execution */
+/* If we compiled MEM_STATS support, report alloc requests before dying */
+{
+#ifdef MEM_STATS
+  cinfo->err->trace_level = 2;	/* force self_destruct to report stats */
+#endif
+  ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, which);
+}
+
+
+/*
+ * Allocation of "small" objects.
+ *
+ * For these, we use pooled storage.  When a new pool must be created,
+ * we try to get enough space for the current request plus a "slop" factor,
+ * where the slop will be the amount of leftover space in the new pool.
+ * The speed vs. space tradeoff is largely determined by the slop values.
+ * A different slop value is provided for each pool class (lifetime),
+ * and we also distinguish the first pool of a class from later ones.
+ * NOTE: the values given work fairly well on both 16- and 32-bit-int
+ * machines, but may be too small if longs are 64 bits or more.
+ */
+
+static const size_t first_pool_slop[JPOOL_NUMPOOLS] = 
+{
+	1600,			/* first PERMANENT pool */
+	16000			/* first IMAGE pool */
+};
+
+static const size_t extra_pool_slop[JPOOL_NUMPOOLS] = 
+{
+	0,			/* additional PERMANENT pools */
+	5000			/* additional IMAGE pools */
+};
+
+#define MIN_SLOP  50		/* greater than 0 to avoid futile looping */
+
+
+METHODDEF(void *)
+alloc_small (j_common_ptr cinfo, int pool_id, size_t sizeofobject)
+/* Allocate a "small" object */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr hdr_ptr, prev_hdr_ptr;
+  char * data_ptr;
+  size_t odd_bytes, min_request, slop;
+
+  /* Check for unsatisfiable request (do now to ensure no overflow below) */
+  if (sizeofobject > (size_t) (MAX_ALLOC_CHUNK-SIZEOF(small_pool_hdr)))
+    out_of_memory(cinfo, 1);	/* request exceeds malloc's ability */
+
+  /* Round up the requested size to a multiple of SIZEOF(ALIGN_TYPE) */
+  odd_bytes = sizeofobject % SIZEOF(ALIGN_TYPE);
+  if (odd_bytes > 0)
+    sizeofobject += SIZEOF(ALIGN_TYPE) - odd_bytes;
+
+  /* See if space is available in any existing pool */
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+  prev_hdr_ptr = NULL;
+  hdr_ptr = mem->small_list[pool_id];
+  while (hdr_ptr != NULL) {
+    if (hdr_ptr->hdr.bytes_left >= sizeofobject)
+      break;			/* found pool with enough space */
+    prev_hdr_ptr = hdr_ptr;
+    hdr_ptr = hdr_ptr->hdr.next;
+  }
+
+  /* Time to make a new pool? */
+  if (hdr_ptr == NULL) {
+    /* min_request is what we need now, slop is what will be leftover */
+    min_request = sizeofobject + SIZEOF(small_pool_hdr);
+    if (prev_hdr_ptr == NULL)	/* first pool in class? */
+      slop = first_pool_slop[pool_id];
+    else
+      slop = extra_pool_slop[pool_id];
+    /* Don't ask for more than MAX_ALLOC_CHUNK */
+    if (slop > (size_t) (MAX_ALLOC_CHUNK-min_request))
+      slop = (size_t) (MAX_ALLOC_CHUNK-min_request);
+    /* Try to get space, if fail reduce slop and try again */
+    for (;;) {
+      hdr_ptr = (small_pool_ptr) jpeg_get_small(cinfo, min_request + slop);
+      if (hdr_ptr != NULL)
+	break;
+      slop /= 2;
+      if (slop < MIN_SLOP)	/* give up when it gets real small */
+	out_of_memory(cinfo, 2); /* jpeg_get_small failed */
+    }
+    mem->total_space_allocated += (long)(min_request + slop);
+    /* Success, initialize the new pool header and add to end of list */
+    hdr_ptr->hdr.next = NULL;
+    hdr_ptr->hdr.bytes_used = 0;
+    hdr_ptr->hdr.bytes_left = sizeofobject + slop;
+    if (prev_hdr_ptr == NULL)	/* first pool in class? */
+      mem->small_list[pool_id] = hdr_ptr;
+    else
+      prev_hdr_ptr->hdr.next = hdr_ptr;
+  }
+
+  /* OK, allocate the object from the current pool */
+  data_ptr = (char *) (hdr_ptr + 1); /* point to first data byte in pool */
+  data_ptr += hdr_ptr->hdr.bytes_used; /* point to place for object */
+  hdr_ptr->hdr.bytes_used += sizeofobject;
+  hdr_ptr->hdr.bytes_left -= sizeofobject;
+
+  return (void *) data_ptr;
+}
+
+
+/*
+ * Allocation of "large" objects.
+ *
+ * The external semantics of these are the same as "small" objects,
+ * except that FAR pointers are used on 80x86.  However the pool
+ * management heuristics are quite different.  We assume that each
+ * request is large enough that it may as well be passed directly to
+ * jpeg_get_large; the pool management just links everything together
+ * so that we can free it all on demand.
+ * Note: the major use of "large" objects is in JSAMPARRAY and JBLOCKARRAY
+ * structures.  The routines that create these structures (see below)
+ * deliberately bunch rows together to ensure a large request size.
+ */
+
+METHODDEF(void FAR *)
+alloc_large (j_common_ptr cinfo, int pool_id, size_t sizeofobject)
+/* Allocate a "large" object */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  large_pool_ptr hdr_ptr;
+  size_t odd_bytes;
+
+  /* Check for unsatisfiable request (do now to ensure no overflow below) */
+  if (sizeofobject > (size_t) (MAX_ALLOC_CHUNK-SIZEOF(large_pool_hdr)))
+    out_of_memory(cinfo, 3);	/* request exceeds malloc's ability */
+
+  /* Round up the requested size to a multiple of SIZEOF(ALIGN_TYPE) */
+  odd_bytes = sizeofobject % SIZEOF(ALIGN_TYPE);
+  if (odd_bytes > 0)
+    sizeofobject += SIZEOF(ALIGN_TYPE) - odd_bytes;
+
+  /* Always make a new pool */
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  hdr_ptr = (large_pool_ptr) jpeg_get_large(cinfo, sizeofobject +
+					    SIZEOF(large_pool_hdr));
+  if (hdr_ptr == NULL)
+    out_of_memory(cinfo, 4);	/* jpeg_get_large failed */
+  mem->total_space_allocated += (long)(sizeofobject + SIZEOF(large_pool_hdr));
+
+  /* Success, initialize the new pool header and add to list */
+  hdr_ptr->hdr.next = mem->large_list[pool_id];
+  /* We maintain space counts in each pool header for statistical purposes,
+   * even though they are not needed for allocation.
+   */
+  hdr_ptr->hdr.bytes_used = sizeofobject;
+  hdr_ptr->hdr.bytes_left = 0;
+  mem->large_list[pool_id] = hdr_ptr;
+
+  return (void FAR *) (hdr_ptr + 1); /* point to first data byte in pool */
+}
+
+
+/*
+ * Creation of 2-D sample arrays.
+ * The pointers are in near heap, the samples themselves in FAR heap.
+ *
+ * To minimize allocation overhead and to allow I/O of large contiguous
+ * blocks, we allocate the sample rows in groups of as many rows as possible
+ * without exceeding MAX_ALLOC_CHUNK total bytes per allocation request.
+ * NB: the virtual array control routines, later in this file, know about
+ * this chunking of rows.  The rowsperchunk value is left in the mem manager
+ * object so that it can be saved away if this sarray is the workspace for
+ * a virtual array.
+ */
+
+METHODDEF(JSAMPARRAY)
+alloc_sarray (j_common_ptr cinfo, int pool_id,
+	      JDIMENSION samplesperrow, JDIMENSION numrows)
+/* Allocate a 2-D sample array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  JSAMPARRAY result;
+  JSAMPROW workspace;
+  JDIMENSION rowsperchunk, currow, i;
+  long ltemp;
+
+  /* Calculate max # of rows allowed in one allocation chunk */
+  ltemp = (MAX_ALLOC_CHUNK-SIZEOF(large_pool_hdr)) /
+	  ((long) samplesperrow * SIZEOF(JSAMPLE));
+  if (ltemp <= 0)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+  if (ltemp < (long) numrows)
+    rowsperchunk = (JDIMENSION) ltemp;
+  else
+    rowsperchunk = numrows;
+  mem->last_rowsperchunk = rowsperchunk;
+
+  /* Get space for row pointers (small object) */
+  result = (JSAMPARRAY) alloc_small(cinfo, pool_id,
+				    (size_t) (numrows * SIZEOF(JSAMPROW)));
+
+  /* Get the rows themselves (large objects) */
+  currow = 0;
+  while (currow < numrows) {
+    rowsperchunk = MIN(rowsperchunk, numrows - currow);
+    workspace = (JSAMPROW) alloc_large(cinfo, pool_id,
+	(size_t) ((size_t) rowsperchunk * (size_t) samplesperrow
+		  * SIZEOF(JSAMPLE)));
+    for (i = rowsperchunk; i > 0; i--) {
+      result[currow++] = workspace;
+      workspace += samplesperrow;
+    }
+  }
+
+  return result;
+}
+
+
+/*
+ * Creation of 2-D coefficient-block arrays.
+ * This is essentially the same as the code for sample arrays, above.
+ */
+
+METHODDEF(JBLOCKARRAY)
+alloc_barray (j_common_ptr cinfo, int pool_id,
+	      JDIMENSION blocksperrow, JDIMENSION numrows)
+/* Allocate a 2-D coefficient-block array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  JBLOCKARRAY result;
+  JBLOCKROW workspace;
+  JDIMENSION rowsperchunk, currow, i;
+  long ltemp;
+
+  /* Calculate max # of rows allowed in one allocation chunk */
+  ltemp = (MAX_ALLOC_CHUNK-SIZEOF(large_pool_hdr)) /
+	  ((long) blocksperrow * SIZEOF(JBLOCK));
+  if (ltemp <= 0)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+  if (ltemp < (long) numrows)
+    rowsperchunk = (JDIMENSION) ltemp;
+  else
+    rowsperchunk = numrows;
+  mem->last_rowsperchunk = rowsperchunk;
+
+  /* Get space for row pointers (small object) */
+  result = (JBLOCKARRAY) alloc_small(cinfo, pool_id,
+				     (size_t) (numrows * SIZEOF(JBLOCKROW)));
+
+  /* Get the rows themselves (large objects) */
+  currow = 0;
+  while (currow < numrows) {
+    rowsperchunk = MIN(rowsperchunk, numrows - currow);
+    workspace = (JBLOCKROW) alloc_large(cinfo, pool_id,
+	(size_t) ((size_t) rowsperchunk * (size_t) blocksperrow
+		  * SIZEOF(JBLOCK)));
+    for (i = rowsperchunk; i > 0; i--) {
+      result[currow++] = workspace;
+      workspace += blocksperrow;
+    }
+  }
+
+  return result;
+}
+
+
+/*
+ * About virtual array management:
+ *
+ * The above "normal" array routines are only used to allocate strip buffers
+ * (as wide as the image, but just a few rows high).  Full-image-sized buffers
+ * are handled as "virtual" arrays.  The array is still accessed a strip at a
+ * time, but the memory manager must save the whole array for repeated
+ * accesses.  The intended implementation is that there is a strip buffer in
+ * memory (as high as is possible given the desired memory limit), plus a
+ * backing file that holds the rest of the array.
+ *
+ * The request_virt_array routines are told the total size of the image and
+ * the maximum number of rows that will be accessed at once.  The in-memory
+ * buffer must be at least as large as the maxaccess value.
+ *
+ * The request routines create control blocks but not the in-memory buffers.
+ * That is postponed until realize_virt_arrays is called.  At that time the
+ * total amount of space needed is known (approximately, anyway), so free
+ * memory can be divided up fairly.
+ *
+ * The access_virt_array routines are responsible for making a specific strip
+ * area accessible (after reading or writing the backing file, if necessary).
+ * Note that the access routines are told whether the caller intends to modify
+ * the accessed strip; during a read-only pass this saves having to rewrite
+ * data to disk.  The access routines are also responsible for pre-zeroing
+ * any newly accessed rows, if pre-zeroing was requested.
+ *
+ * In current usage, the access requests are usually for nonoverlapping
+ * strips; that is, successive access start_row numbers differ by exactly
+ * num_rows = maxaccess.  This means we can get good performance with simple
+ * buffer dump/reload logic, by making the in-memory buffer be a multiple
+ * of the access height; then there will never be accesses across bufferload
+ * boundaries.  The code will still work with overlapping access requests,
+ * but it doesn't handle bufferload overlaps very efficiently.
+ */
+
+
+METHODDEF(jvirt_sarray_ptr)
+request_virt_sarray (j_common_ptr cinfo, int pool_id, int pre_zero,
+		     JDIMENSION samplesperrow, JDIMENSION numrows,
+		     JDIMENSION maxaccess)
+/* Request a virtual 2-D sample array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  jvirt_sarray_ptr result;
+
+  /* Only IMAGE-lifetime virtual arrays are currently supported */
+  if (pool_id != JPOOL_IMAGE)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  /* get control block */
+  result = (jvirt_sarray_ptr) alloc_small(cinfo, pool_id,
+					  SIZEOF(struct jvirt_sarray_control));
+
+  result->mem_buffer = NULL;	/* marks array not yet realized */
+  result->rows_in_array = numrows;
+  result->samplesperrow = samplesperrow;
+  result->maxaccess = maxaccess;
+  result->pre_zero = pre_zero;
+  result->b_s_open = FALSE;	/* no associated backing-store object */
+  result->next = mem->virt_sarray_list; /* add to list of virtual arrays */
+  mem->virt_sarray_list = result;
+
+  return result;
+}
+
+
+METHODDEF(jvirt_barray_ptr)
+request_virt_barray (j_common_ptr cinfo, int pool_id, int pre_zero,
+		     JDIMENSION blocksperrow, JDIMENSION numrows,
+		     JDIMENSION maxaccess)
+/* Request a virtual 2-D coefficient-block array */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  jvirt_barray_ptr result;
+
+  /* Only IMAGE-lifetime virtual arrays are currently supported */
+  if (pool_id != JPOOL_IMAGE)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+  /* get control block */
+  result = (jvirt_barray_ptr) alloc_small(cinfo, pool_id,
+					  SIZEOF(struct jvirt_barray_control));
+
+  result->mem_buffer = NULL;	/* marks array not yet realized */
+  result->rows_in_array = numrows;
+  result->blocksperrow = blocksperrow;
+  result->maxaccess = maxaccess;
+  result->pre_zero = pre_zero;
+  result->b_s_open = FALSE;	/* no associated backing-store object */
+  result->next = mem->virt_barray_list; /* add to list of virtual arrays */
+  mem->virt_barray_list = result;
+
+  return result;
+}
+
+
+METHODDEF(void)
+realize_virt_arrays (j_common_ptr cinfo)
+/* Allocate the in-memory buffers for any unrealized virtual arrays */
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  long space_per_minheight, maximum_space, avail_mem;
+  long minheights, max_minheights;
+  jvirt_sarray_ptr sptr;
+  jvirt_barray_ptr bptr;
+
+  /* Compute the minimum space needed (maxaccess rows in each buffer)
+   * and the maximum space needed (full image height in each buffer).
+   * These may be of use to the system-dependent jpeg_mem_available routine.
+   */
+  space_per_minheight = 0;
+  maximum_space = 0;
+  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+    if (sptr->mem_buffer == NULL) { /* if not realized yet */
+      space_per_minheight += (long) sptr->maxaccess *
+			     (long) sptr->samplesperrow * SIZEOF(JSAMPLE);
+      maximum_space += (long) sptr->rows_in_array *
+		       (long) sptr->samplesperrow * SIZEOF(JSAMPLE);
+    }
+  }
+  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+    if (bptr->mem_buffer == NULL) { /* if not realized yet */
+      space_per_minheight += (long) bptr->maxaccess *
+			     (long) bptr->blocksperrow * SIZEOF(JBLOCK);
+      maximum_space += (long) bptr->rows_in_array *
+		       (long) bptr->blocksperrow * SIZEOF(JBLOCK);
+    }
+  }
+
+  if (space_per_minheight <= 0)
+    return;			/* no unrealized arrays, no work */
+
+  /* Determine amount of memory to actually use; this is system-dependent. */
+  avail_mem = jpeg_mem_available(cinfo, space_per_minheight, maximum_space,
+				 mem->total_space_allocated);
+
+  /* If the maximum space needed is available, make all the buffers full
+   * height; otherwise parcel it out with the same number of minheights
+   * in each buffer.
+   */
+  if (avail_mem >= maximum_space)
+    max_minheights = 1000000000L;
+  else {
+    max_minheights = avail_mem / space_per_minheight;
+    /* If there doesn't seem to be enough space, try to get the minimum
+     * anyway.  This allows a "stub" implementation of jpeg_mem_available().
+     */
+    if (max_minheights <= 0)
+      max_minheights = 1;
+  }
+
+  /* Allocate the in-memory buffers and initialize backing store as needed. */
+
+  for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+    if (sptr->mem_buffer == NULL) { /* if not realized yet */
+      minheights = ((long) sptr->rows_in_array - 1L) / sptr->maxaccess + 1L;
+      if (minheights <= max_minheights) {
+	/* This buffer fits in memory */
+	sptr->rows_in_mem = sptr->rows_in_array;
+      } else {
+	/* It doesn't fit in memory, create backing store. */
+	sptr->rows_in_mem = (JDIMENSION) (max_minheights * sptr->maxaccess);
+	jpeg_open_backing_store(cinfo, & sptr->b_s_info,
+				(long) sptr->rows_in_array *
+				(long) sptr->samplesperrow *
+				(long) SIZEOF(JSAMPLE));
+	sptr->b_s_open = TRUE;
+      }
+      sptr->mem_buffer = alloc_sarray(cinfo, JPOOL_IMAGE,
+				      sptr->samplesperrow, sptr->rows_in_mem);
+      sptr->rowsperchunk = mem->last_rowsperchunk;
+      sptr->cur_start_row = 0;
+      sptr->first_undef_row = 0;
+      sptr->dirty = FALSE;
+    }
+  }
+
+  for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+    if (bptr->mem_buffer == NULL) { /* if not realized yet */
+      minheights = ((long) bptr->rows_in_array - 1L) / bptr->maxaccess + 1L;
+      if (minheights <= max_minheights) {
+	/* This buffer fits in memory */
+	bptr->rows_in_mem = bptr->rows_in_array;
+      } else {
+	/* It doesn't fit in memory, create backing store. */
+	bptr->rows_in_mem = (JDIMENSION) (max_minheights * bptr->maxaccess);
+	jpeg_open_backing_store(cinfo, & bptr->b_s_info,
+				(long) bptr->rows_in_array *
+				(long) bptr->blocksperrow *
+				(long) SIZEOF(JBLOCK));
+	bptr->b_s_open = TRUE;
+      }
+      bptr->mem_buffer = alloc_barray(cinfo, JPOOL_IMAGE,
+				      bptr->blocksperrow, bptr->rows_in_mem);
+      bptr->rowsperchunk = mem->last_rowsperchunk;
+      bptr->cur_start_row = 0;
+      bptr->first_undef_row = 0;
+      bptr->dirty = FALSE;
+    }
+  }
+}
+
+
+LOCAL(void)
+do_sarray_io (j_common_ptr cinfo, jvirt_sarray_ptr ptr, int writing)
+/* Do backing store read or write of a virtual sample array */
+{
+  long bytesperrow, file_offset, byte_count, rows, thisrow, i;
+
+  bytesperrow = (long) ptr->samplesperrow * SIZEOF(JSAMPLE);
+  file_offset = ptr->cur_start_row * bytesperrow;
+  /* Loop to read or write each allocation chunk in mem_buffer */
+  for (i = 0; i < (long) ptr->rows_in_mem; i += ptr->rowsperchunk) {
+    /* One chunk, but check for short chunk at end of buffer */
+    rows = MIN((long) ptr->rowsperchunk, (long) ptr->rows_in_mem - i);
+    /* Transfer no more than is currently defined */
+    thisrow = (long) ptr->cur_start_row + i;
+    rows = MIN(rows, (long) ptr->first_undef_row - thisrow);
+    /* Transfer no more than fits in file */
+    rows = MIN(rows, (long) ptr->rows_in_array - thisrow);
+    if (rows <= 0)		/* this chunk might be past end of file! */
+      break;
+    byte_count = rows * bytesperrow;
+    if (writing)
+      (*ptr->b_s_info.write_backing_store) (cinfo, & ptr->b_s_info,
+					    (void FAR *) ptr->mem_buffer[i],
+					    file_offset, byte_count);
+    else
+      (*ptr->b_s_info.read_backing_store) (cinfo, & ptr->b_s_info,
+					   (void FAR *) ptr->mem_buffer[i],
+					   file_offset, byte_count);
+    file_offset += byte_count;
+  }
+}
+
+
+LOCAL(void)
+do_barray_io (j_common_ptr cinfo, jvirt_barray_ptr ptr, int writing)
+/* Do backing store read or write of a virtual coefficient-block array */
+{
+  long bytesperrow, file_offset, byte_count, rows, thisrow, i;
+
+  bytesperrow = (long) ptr->blocksperrow * SIZEOF(JBLOCK);
+  file_offset = ptr->cur_start_row * bytesperrow;
+  /* Loop to read or write each allocation chunk in mem_buffer */
+  for (i = 0; i < (long) ptr->rows_in_mem; i += ptr->rowsperchunk) {
+    /* One chunk, but check for short chunk at end of buffer */
+    rows = MIN((long) ptr->rowsperchunk, (long) ptr->rows_in_mem - i);
+    /* Transfer no more than is currently defined */
+    thisrow = (long) ptr->cur_start_row + i;
+    rows = MIN(rows, (long) ptr->first_undef_row - thisrow);
+    /* Transfer no more than fits in file */
+    rows = MIN(rows, (long) ptr->rows_in_array - thisrow);
+    if (rows <= 0)		/* this chunk might be past end of file! */
+      break;
+    byte_count = rows * bytesperrow;
+    if (writing)
+      (*ptr->b_s_info.write_backing_store) (cinfo, & ptr->b_s_info,
+					    (void FAR *) ptr->mem_buffer[i],
+					    file_offset, byte_count);
+    else
+      (*ptr->b_s_info.read_backing_store) (cinfo, & ptr->b_s_info,
+					   (void FAR *) ptr->mem_buffer[i],
+					   file_offset, byte_count);
+    file_offset += byte_count;
+  }
+}
+
+
+METHODDEF(JSAMPARRAY)
+access_virt_sarray (j_common_ptr cinfo, jvirt_sarray_ptr ptr,
+		    JDIMENSION start_row, JDIMENSION num_rows,
+		    int writable)
+/* Access the part of a virtual sample array starting at start_row */
+/* and extending for num_rows rows.  writable is true if  */
+/* caller intends to modify the accessed area. */
+{
+  JDIMENSION end_row = start_row + num_rows;
+  JDIMENSION undef_row;
+
+  /* debugging check */
+  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
+      ptr->mem_buffer == NULL)
+    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+
+  /* Make the desired part of the virtual array accessible */
+  if (start_row < ptr->cur_start_row ||
+      end_row > ptr->cur_start_row+ptr->rows_in_mem) {
+    if (! ptr->b_s_open)
+      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
+    /* Flush old buffer contents if necessary */
+    if (ptr->dirty) {
+      do_sarray_io(cinfo, ptr, TRUE);
+      ptr->dirty = FALSE;
+    }
+    /* Decide what part of virtual array to access.
+     * Algorithm: if target address > current window, assume forward scan,
+     * load starting at target address.  If target address < current window,
+     * assume backward scan, load so that target area is top of window.
+     * Note that when switching from forward write to forward read, will have
+     * start_row = 0, so the limiting case applies and we load from 0 anyway.
+     */
+    if (start_row > ptr->cur_start_row) {
+      ptr->cur_start_row = start_row;
+    } else {
+      /* use long arithmetic here to avoid overflow & unsigned problems */
+      long ltemp;
+
+      ltemp = (long) end_row - (long) ptr->rows_in_mem;
+      if (ltemp < 0)
+	ltemp = 0;		/* don't fall off front end of file */
+      ptr->cur_start_row = (JDIMENSION) ltemp;
+    }
+    /* Read in the selected part of the array.
+     * During the initial write pass, we will do no actual read
+     * because the selected part is all undefined.
+     */
+    do_sarray_io(cinfo, ptr, FALSE);
+  }
+  /* Ensure the accessed part of the array is defined; prezero if needed.
+   * To improve locality of access, we only prezero the part of the array
+   * that the caller is about to access, not the entire in-memory array.
+   */
+  if (ptr->first_undef_row < end_row) {
+    if (ptr->first_undef_row < start_row) {
+      if (writable)		/* writer skipped over a section of array */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+      undef_row = start_row;	/* but reader is allowed to read ahead */
+    } else {
+      undef_row = ptr->first_undef_row;
+    }
+    if (writable)
+      ptr->first_undef_row = end_row;
+    if (ptr->pre_zero) {
+      size_t bytesperrow = (size_t) ptr->samplesperrow * SIZEOF(JSAMPLE);
+      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
+      end_row -= ptr->cur_start_row;
+      while (undef_row < end_row) {
+	jzero_far((void FAR *) ptr->mem_buffer[undef_row], bytesperrow);
+	undef_row++;
+      }
+    } else {
+      if (! writable)		/* reader looking at undefined data */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+    }
+  }
+  /* Flag the buffer dirty if caller will write in it */
+  if (writable)
+    ptr->dirty = TRUE;
+  /* Return address of proper part of the buffer */
+  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
+}
+
+
+METHODDEF(JBLOCKARRAY)
+access_virt_barray (j_common_ptr cinfo, jvirt_barray_ptr ptr,
+		    JDIMENSION start_row, JDIMENSION num_rows,
+		    int writable)
+/* Access the part of a virtual block array starting at start_row */
+/* and extending for num_rows rows.  writable is true if  */
+/* caller intends to modify the accessed area. */
+{
+  JDIMENSION end_row = start_row + num_rows;
+  JDIMENSION undef_row;
+
+  /* debugging check */
+  if (end_row > ptr->rows_in_array || num_rows > ptr->maxaccess ||
+      ptr->mem_buffer == NULL)
+    ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+
+  /* Make the desired part of the virtual array accessible */
+  if (start_row < ptr->cur_start_row ||
+      end_row > ptr->cur_start_row+ptr->rows_in_mem) {
+    if (! ptr->b_s_open)
+      ERREXIT(cinfo, JERR_VIRTUAL_BUG);
+    /* Flush old buffer contents if necessary */
+    if (ptr->dirty) {
+      do_barray_io(cinfo, ptr, TRUE);
+      ptr->dirty = FALSE;
+    }
+    /* Decide what part of virtual array to access.
+     * Algorithm: if target address > current window, assume forward scan,
+     * load starting at target address.  If target address < current window,
+     * assume backward scan, load so that target area is top of window.
+     * Note that when switching from forward write to forward read, will have
+     * start_row = 0, so the limiting case applies and we load from 0 anyway.
+     */
+    if (start_row > ptr->cur_start_row) {
+      ptr->cur_start_row = start_row;
+    } else {
+      /* use long arithmetic here to avoid overflow & unsigned problems */
+      long ltemp;
+
+      ltemp = (long) end_row - (long) ptr->rows_in_mem;
+      if (ltemp < 0)
+	ltemp = 0;		/* don't fall off front end of file */
+      ptr->cur_start_row = (JDIMENSION) ltemp;
+    }
+    /* Read in the selected part of the array.
+     * During the initial write pass, we will do no actual read
+     * because the selected part is all undefined.
+     */
+    do_barray_io(cinfo, ptr, FALSE);
+  }
+  /* Ensure the accessed part of the array is defined; prezero if needed.
+   * To improve locality of access, we only prezero the part of the array
+   * that the caller is about to access, not the entire in-memory array.
+   */
+  if (ptr->first_undef_row < end_row) {
+    if (ptr->first_undef_row < start_row) {
+      if (writable)		/* writer skipped over a section of array */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+      undef_row = start_row;	/* but reader is allowed to read ahead */
+    } else {
+      undef_row = ptr->first_undef_row;
+    }
+    if (writable)
+      ptr->first_undef_row = end_row;
+    if (ptr->pre_zero) {
+      size_t bytesperrow = (size_t) ptr->blocksperrow * SIZEOF(JBLOCK);
+      undef_row -= ptr->cur_start_row; /* make indexes relative to buffer */
+      end_row -= ptr->cur_start_row;
+      while (undef_row < end_row) {
+	jzero_far((void FAR *) ptr->mem_buffer[undef_row], bytesperrow);
+	undef_row++;
+      }
+    } else {
+      if (! writable)		/* reader looking at undefined data */
+	ERREXIT(cinfo, JERR_BAD_VIRTUAL_ACCESS);
+    }
+  }
+  /* Flag the buffer dirty if caller will write in it */
+  if (writable)
+    ptr->dirty = TRUE;
+  /* Return address of proper part of the buffer */
+  return ptr->mem_buffer + (start_row - ptr->cur_start_row);
+}
+
+
+/*
+ * Release all objects belonging to a specified pool.
+ */
+
+METHODDEF(void)
+free_pool (j_common_ptr cinfo, int pool_id)
+{
+  my_mem_ptr mem = (my_mem_ptr) cinfo->mem;
+  small_pool_ptr shdr_ptr;
+  large_pool_ptr lhdr_ptr;
+  size_t space_freed;
+
+  if (pool_id < 0 || pool_id >= JPOOL_NUMPOOLS)
+    ERREXIT1(cinfo, JERR_BAD_POOL_ID, pool_id);	/* safety check */
+
+#ifdef MEM_STATS
+  if (cinfo->err->trace_level > 1)
+    print_mem_stats(cinfo, pool_id); /* print pool's memory usage statistics */
+#endif
+
+  /* If freeing IMAGE pool, close any virtual arrays first */
+  if (pool_id == JPOOL_IMAGE) {
+    jvirt_sarray_ptr sptr;
+    jvirt_barray_ptr bptr;
+
+    for (sptr = mem->virt_sarray_list; sptr != NULL; sptr = sptr->next) {
+      if (sptr->b_s_open) {	/* there may be no backing store */
+	sptr->b_s_open = FALSE;	/* prevent recursive close if error */
+	(*sptr->b_s_info.close_backing_store) (cinfo, & sptr->b_s_info);
+      }
+    }
+    mem->virt_sarray_list = NULL;
+    for (bptr = mem->virt_barray_list; bptr != NULL; bptr = bptr->next) {
+      if (bptr->b_s_open) {	/* there may be no backing store */
+	bptr->b_s_open = FALSE;	/* prevent recursive close if error */
+	(*bptr->b_s_info.close_backing_store) (cinfo, & bptr->b_s_info);
+      }
+    }
+    mem->virt_barray_list = NULL;
+  }
+
+  /* Release large objects */
+  lhdr_ptr = mem->large_list[pool_id];
+  mem->large_list[pool_id] = NULL;
+
+  while (lhdr_ptr != NULL) {
+    large_pool_ptr next_lhdr_ptr = lhdr_ptr->hdr.next;
+    space_freed = lhdr_ptr->hdr.bytes_used +
+		  lhdr_ptr->hdr.bytes_left +
+		  SIZEOF(large_pool_hdr);
+    jpeg_free_large(cinfo, (void FAR *) lhdr_ptr, space_freed);
+    mem->total_space_allocated -= (long)space_freed;
+    lhdr_ptr = next_lhdr_ptr;
+  }
+
+  /* Release small objects */
+  shdr_ptr = mem->small_list[pool_id];
+  mem->small_list[pool_id] = NULL;
+
+  while (shdr_ptr != NULL) {
+    small_pool_ptr next_shdr_ptr = shdr_ptr->hdr.next;
+    space_freed = shdr_ptr->hdr.bytes_used +
+		  shdr_ptr->hdr.bytes_left +
+		  SIZEOF(small_pool_hdr);
+    jpeg_free_small(cinfo, (void *) shdr_ptr, space_freed);
+    mem->total_space_allocated -= (long)space_freed;
+    shdr_ptr = next_shdr_ptr;
+  }
+}
+
+
+/*
+ * Close up shop entirely.
+ * Note that this cannot be called unless cinfo->mem is non-NULL.
+ */
+
+METHODDEF(void)
+self_destruct (j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Close all backing store, release all memory.
+   * Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS-1; pool >= JPOOL_PERMANENT; pool--) {
+    free_pool(cinfo, pool);
+  }
+
+  /* Release the memory manager control block too. */
+  jpeg_free_small(cinfo, (void *) cinfo->mem, SIZEOF(my_memory_mgr));
+  cinfo->mem = NULL;		/* ensures I will be called only once */
+
+  jpeg_mem_term(cinfo);		/* system-dependent cleanup */
+}
+
+
+/*
+ * Memory manager initialization.
+ * When this is called, only the error manager pointer is valid in cinfo!
+ */
+
+GLOBAL(void)
+jinit_memory_mgr (j_common_ptr cinfo)
+{
+  my_mem_ptr mem;
+  long max_to_use;
+  int pool;
+  size_t test_mac;
+
+  cinfo->mem = NULL;		/* for safety if init fails */
+
+  /* Check for configuration errors.
+   * SIZEOF(ALIGN_TYPE) should be a power of 2; otherwise, it probably
+   * doesn't reflect any real hardware alignment requirement.
+   * The test is a little tricky: for X>0, X and X-1 have no one-bits
+   * in common if and only if X is a power of 2, ie has only one one-bit.
+   * Some compilers may give an "unreachable code" warning here; ignore it.
+   */
+  if ((SIZEOF(ALIGN_TYPE) & (SIZEOF(ALIGN_TYPE)-1)) != 0)
+    ERREXIT(cinfo, JERR_BAD_ALIGN_TYPE);
+  /* MAX_ALLOC_CHUNK must be representable as type size_t, and must be
+   * a multiple of SIZEOF(ALIGN_TYPE).
+   * Again, an "unreachable code" warning may be ignored here.
+   * But a "constant too large" warning means you need to fix MAX_ALLOC_CHUNK.
+   */
+  test_mac = (size_t) MAX_ALLOC_CHUNK;
+  if ((long) test_mac != MAX_ALLOC_CHUNK ||
+      (MAX_ALLOC_CHUNK % SIZEOF(ALIGN_TYPE)) != 0)
+    ERREXIT(cinfo, JERR_BAD_ALLOC_CHUNK);
+
+  max_to_use = jpeg_mem_init(cinfo); /* system-dependent initialization */
+
+  /* Attempt to allocate memory manager's control block */
+  mem = (my_mem_ptr) jpeg_get_small(cinfo, SIZEOF(my_memory_mgr));
+
+  if (mem == NULL) {
+    jpeg_mem_term(cinfo);	/* system-dependent cleanup */
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 0);
+  }
+
+  /* OK, fill in the method pointers */
+  mem->pub.alloc_small = alloc_small;
+  mem->pub.alloc_large = alloc_large;
+  mem->pub.alloc_sarray = alloc_sarray;
+  mem->pub.alloc_barray = alloc_barray;
+  mem->pub.request_virt_sarray = request_virt_sarray;
+  mem->pub.request_virt_barray = request_virt_barray;
+  mem->pub.realize_virt_arrays = realize_virt_arrays;
+  mem->pub.access_virt_sarray = access_virt_sarray;
+  mem->pub.access_virt_barray = access_virt_barray;
+  mem->pub.free_pool = free_pool;
+  mem->pub.self_destruct = self_destruct;
+
+  /* Make MAX_ALLOC_CHUNK accessible to other modules */
+  mem->pub.max_alloc_chunk = MAX_ALLOC_CHUNK;
+
+  /* Initialize working state */
+  mem->pub.max_memory_to_use = max_to_use;
+
+  for (pool = JPOOL_NUMPOOLS-1; pool >= JPOOL_PERMANENT; pool--) {
+    mem->small_list[pool] = NULL;
+    mem->large_list[pool] = NULL;
+  }
+  mem->virt_sarray_list = NULL;
+  mem->virt_barray_list = NULL;
+
+  mem->total_space_allocated = SIZEOF(my_memory_mgr);
+
+  /* Declare ourselves open for business */
+  cinfo->mem = & mem->pub;
+
+  /* Check for an environment variable JPEGMEM; if found, override the
+   * default max_memory setting from jpeg_mem_init.  Note that the
+   * surrounding application may again override this value.
+   * If your system doesn't support getenv(), define NO_GETENV to disable
+   * this feature.
+   */
+#ifndef NO_GETENV
+  { char * memenv;
+
+    if ((memenv = getenv("JPEGMEM")) != NULL) {
+      char ch = 'x';
+
+      if (sscanf(memenv, "%ld%c", &max_to_use, &ch) > 0) {
+	if (ch == 'm' || ch == 'M')
+	  max_to_use *= 1000L;
+	mem->pub.max_memory_to_use = max_to_use * 1000L;
+      }
+    }
+  }
+#endif
+
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jmemnobs.cpp b/ml/dlib/dlib/external/libjpeg/jmemnobs.cpp
new file mode 100644
index 000000000..27fe6c457
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jmemnobs.cpp
@@ -0,0 +1,109 @@
+/*
+ * jmemnobs.c
+ *
+ * Copyright (C) 1992-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file provides a really simple implementation of the system-
+ * dependent portion of the JPEG memory manager.  This implementation
+ * assumes that no backing-store files are needed: all required space
+ * can be obtained from malloc().
+ * This is very portable in the sense that it'll compile on almost anything,
+ * but you'd better have lots of main memory (or virtual memory) if you want
+ * to process big images.
+ * Note that the max_memory_to_use option is ignored by this implementation.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jmemsys.h"		/* import the system-dependent declarations */
+
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare malloc(),free() */
+extern void * malloc JPP((size_t size));
+extern void free JPP((void *ptr));
+#endif
+
+
+/*
+ * Memory allocation and freeing are controlled by the regular library
+ * routines malloc() and free().
+ */
+
+GLOBAL(void *)
+jpeg_get_small (j_common_ptr , size_t sizeofobject)
+{
+  return (void *) malloc(sizeofobject);
+}
+
+GLOBAL(void)
+jpeg_free_small (j_common_ptr , void * object, size_t )
+{
+  free(object);
+}
+
+
+/*
+ * "Large" objects are treated the same as "small" ones.
+ * NB: although we include FAR keywords in the routine declarations,
+ * this file won't actually work in 80x86 small/medium model; at least,
+ * you probably won't be able to process useful-size images in only 64KB.
+ */
+
+GLOBAL(void FAR *)
+jpeg_get_large (j_common_ptr , size_t sizeofobject)
+{
+  return (void FAR *) malloc(sizeofobject);
+}
+
+GLOBAL(void)
+jpeg_free_large (j_common_ptr , void FAR * object, size_t )
+{
+  free(object);
+}
+
+
+/*
+ * This routine computes the total memory space available for allocation.
+ * Here we always say, "we got all you want bud!"
+ */
+
+GLOBAL(long)
+jpeg_mem_available (j_common_ptr , long ,
+		    long max_bytes_needed, long )
+{
+  return max_bytes_needed;
+}
+
+
+/*
+ * Backing store (temporary file) management.
+ * Since jpeg_mem_available always promised the moon,
+ * this should never be called and we can just error out.
+ */
+
+GLOBAL(void)
+jpeg_open_backing_store (j_common_ptr cinfo, backing_store_ptr ,
+			 long )
+{
+  ERREXIT(cinfo, JERR_NO_BACKING_STORE);
+}
+
+
+/*
+ * These routines take care of any system-dependent initialization and
+ * cleanup required.  Here, there isn't any.
+ */
+
+GLOBAL(long)
+jpeg_mem_init (j_common_ptr )
+{
+  return 0;			/* just set max_memory_to_use to 0 */
+}
+
+GLOBAL(void)
+jpeg_mem_term (j_common_ptr )
+{
+  /* no work */
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jmemsys.h b/ml/dlib/dlib/external/libjpeg/jmemsys.h
new file mode 100644
index 000000000..6c3c6d348
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jmemsys.h
@@ -0,0 +1,198 @@
+/*
+ * jmemsys.h
+ *
+ * Copyright (C) 1992-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This include file defines the interface between the system-independent
+ * and system-dependent portions of the JPEG memory manager.  No other
+ * modules need include it.  (The system-independent portion is jmemmgr.c;
+ * there are several different versions of the system-dependent portion.)
+ *
+ * This file works as-is for the system-dependent memory managers supplied
+ * in the IJG distribution.  You may need to modify it if you write a
+ * custom memory manager.  If system-dependent changes are needed in
+ * this file, the best method is to #ifdef them based on a configuration
+ * symbol supplied in jconfig.h, as we have done with USE_MSDOS_MEMMGR
+ * and USE_MAC_MEMMGR.
+ */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_get_small		jGetSmall
+#define jpeg_free_small		jFreeSmall
+#define jpeg_get_large		jGetLarge
+#define jpeg_free_large		jFreeLarge
+#define jpeg_mem_available	jMemAvail
+#define jpeg_open_backing_store	jOpenBackStore
+#define jpeg_mem_init		jMemInit
+#define jpeg_mem_term		jMemTerm
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/*
+ * These two functions are used to allocate and release small chunks of
+ * memory.  (Typically the total amount requested through jpeg_get_small is
+ * no more than 20K or so; this will be requested in chunks of a few K each.)
+ * Behavior should be the same as for the standard library functions malloc
+ * and free; in particular, jpeg_get_small must return NULL on failure.
+ * On most systems, these ARE malloc and free.  jpeg_free_small is passed the
+ * size of the object being freed, just in case it's needed.
+ * On an 80x86 machine using small-data memory model, these manage near heap.
+ */
+
+EXTERN(void *) jpeg_get_small JPP((j_common_ptr cinfo, size_t sizeofobject));
+EXTERN(void) jpeg_free_small JPP((j_common_ptr cinfo, void * object,
+				  size_t sizeofobject));
+
+/*
+ * These two functions are used to allocate and release large chunks of
+ * memory (up to the total free space designated by jpeg_mem_available).
+ * The interface is the same as above, except that on an 80x86 machine,
+ * far pointers are used.  On most other machines these are identical to
+ * the jpeg_get/free_small routines; but we keep them separate anyway,
+ * in case a different allocation strategy is desirable for large chunks.
+ */
+
+EXTERN(void FAR *) jpeg_get_large JPP((j_common_ptr cinfo,
+				       size_t sizeofobject));
+EXTERN(void) jpeg_free_large JPP((j_common_ptr cinfo, void FAR * object,
+				  size_t sizeofobject));
+
+/*
+ * The macro MAX_ALLOC_CHUNK designates the maximum number of bytes that may
+ * be requested in a single call to jpeg_get_large (and jpeg_get_small for that
+ * matter, but that case should never come into play).  This macro is needed
+ * to model the 64Kb-segment-size limit of far addressing on 80x86 machines.
+ * On those machines, we expect that jconfig.h will provide a proper value.
+ * On machines with 32-bit flat address spaces, any large constant may be used.
+ *
+ * NB: jmemmgr.c expects that MAX_ALLOC_CHUNK will be representable as type
+ * size_t and will be a multiple of sizeof(align_type).
+ */
+
+#ifndef MAX_ALLOC_CHUNK		/* may be overridden in jconfig.h */
+#define MAX_ALLOC_CHUNK  1000000000L
+#endif
+
+/*
+ * This routine computes the total space still available for allocation by
+ * jpeg_get_large.  If more space than this is needed, backing store will be
+ * used.  NOTE: any memory already allocated must not be counted.
+ *
+ * There is a minimum space requirement, corresponding to the minimum
+ * feasible buffer sizes; jmemmgr.c will request that much space even if
+ * jpeg_mem_available returns zero.  The maximum space needed, enough to hold
+ * all working storage in memory, is also passed in case it is useful.
+ * Finally, the total space already allocated is passed.  If no better
+ * method is available, cinfo->mem->max_memory_to_use - already_allocated
+ * is often a suitable calculation.
+ *
+ * It is OK for jpeg_mem_available to underestimate the space available
+ * (that'll just lead to more backing-store access than is really necessary).
+ * However, an overestimate will lead to failure.  Hence it's wise to subtract
+ * a slop factor from the true available space.  5% should be enough.
+ *
+ * On machines with lots of virtual memory, any large constant may be returned.
+ * Conversely, zero may be returned to always use the minimum amount of memory.
+ */
+
+EXTERN(long) jpeg_mem_available JPP((j_common_ptr cinfo,
+				     long min_bytes_needed,
+				     long max_bytes_needed,
+				     long already_allocated));
+
+
+/*
+ * This structure holds whatever state is needed to access a single
+ * backing-store object.  The read/write/close method pointers are called
+ * by jmemmgr.c to manipulate the backing-store object; all other fields
+ * are private to the system-dependent backing store routines.
+ */
+
+#define TEMP_NAME_LENGTH   64	/* max length of a temporary file's name */
+
+
+#ifdef USE_MSDOS_MEMMGR		/* DOS-specific junk */
+
+typedef unsigned short XMSH;	/* type of extended-memory handles */
+typedef unsigned short EMSH;	/* type of expanded-memory handles */
+
+typedef union {
+  short file_handle;		/* DOS file handle if it's a temp file */
+  XMSH xms_handle;		/* handle if it's a chunk of XMS */
+  EMSH ems_handle;		/* handle if it's a chunk of EMS */
+} handle_union;
+
+#endif /* USE_MSDOS_MEMMGR */
+
+#ifdef USE_MAC_MEMMGR		/* Mac-specific junk */
+#include <Files.h>
+#endif /* USE_MAC_MEMMGR */
+
+
+typedef struct backing_store_struct * backing_store_ptr;
+
+typedef struct backing_store_struct {
+  /* Methods for reading/writing/closing this backing-store object */
+  JMETHOD(void, read_backing_store, (j_common_ptr cinfo,
+				     backing_store_ptr info,
+				     void FAR * buffer_address,
+				     long file_offset, long byte_count));
+  JMETHOD(void, write_backing_store, (j_common_ptr cinfo,
+				      backing_store_ptr info,
+				      void FAR * buffer_address,
+				      long file_offset, long byte_count));
+  JMETHOD(void, close_backing_store, (j_common_ptr cinfo,
+				      backing_store_ptr info));
+
+  /* Private fields for system-dependent backing-store management */
+#ifdef USE_MSDOS_MEMMGR
+  /* For the MS-DOS manager (jmemdos.c), we need: */
+  handle_union handle;		/* reference to backing-store storage object */
+  char temp_name[TEMP_NAME_LENGTH]; /* name if it's a file */
+#else
+#ifdef USE_MAC_MEMMGR
+  /* For the Mac manager (jmemmac.c), we need: */
+  short temp_file;		/* file reference number to temp file */
+  FSSpec tempSpec;		/* the FSSpec for the temp file */
+  char temp_name[TEMP_NAME_LENGTH]; /* name if it's a file */
+#else
+  /* For a typical implementation with temp files, we need: */
+  FILE * temp_file;		/* stdio reference to temp file */
+  char temp_name[TEMP_NAME_LENGTH]; /* name of temp file */
+#endif
+#endif
+} backing_store_info;
+
+
+/*
+ * Initial opening of a backing-store object.  This must fill in the
+ * read/write/close pointers in the object.  The read/write routines
+ * may take an error exit if the specified maximum file size is exceeded.
+ * (If jpeg_mem_available always returns a large value, this routine can
+ * just take an error exit.)
+ */
+
+EXTERN(void) jpeg_open_backing_store JPP((j_common_ptr cinfo,
+					  backing_store_ptr info,
+					  long total_bytes_needed));
+
+
+/*
+ * These routines take care of any system-dependent initialization and
+ * cleanup required.  jpeg_mem_init will be called before anything is
+ * allocated (and, therefore, nothing in cinfo is of use except the error
+ * manager pointer).  It should return a suitable default value for
+ * max_memory_to_use; this may subsequently be overridden by the surrounding
+ * application.  (Note that max_memory_to_use is only important if
+ * jpeg_mem_available chooses to consult it ... no one else will.)
+ * jpeg_mem_term may assume that all requested memory has been freed and that
+ * all opened backing-store objects have been closed.
+ */
+
+EXTERN(long) jpeg_mem_init JPP((j_common_ptr cinfo));
+EXTERN(void) jpeg_mem_term JPP((j_common_ptr cinfo));
diff --git a/ml/dlib/dlib/external/libjpeg/jmorecfg.h b/ml/dlib/dlib/external/libjpeg/jmorecfg.h
new file mode 100644
index 000000000..6082f069a
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jmorecfg.h
@@ -0,0 +1,356 @@
+/*
+ * jmorecfg.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains additional configuration options that customize the
+ * JPEG software for special applications or support machine-dependent
+ * optimizations.  Most users will not need to touch this file.
+ */
+
+
+/*
+ * Define BITS_IN_JSAMPLE as either
+ *   8   for 8-bit sample values (the usual setting)
+ *   12  for 12-bit sample values
+ * Only 8 and 12 are legal data precisions for lossy JPEG according to the
+ * JPEG standard, and the IJG code does not support anything else!
+ * We do not support run-time selection of data precision, sorry.
+ */
+
+#define BITS_IN_JSAMPLE  8	/* use 8 or 12 */
+
+
+/*
+ * Maximum number of components (color channels) allowed in JPEG image.
+ * To meet the letter of the JPEG spec, set this to 255.  However, darn
+ * few applications need more than 4 channels (maybe 5 for CMYK + alpha
+ * mask).  We recommend 10 as a reasonable compromise; use 4 if you are
+ * really short on memory.  (Each allowed component costs a hundred or so
+ * bytes of storage, whether actually used in an image or not.)
+ */
+
+#define MAX_COMPONENTS  10	/* maximum number of image components */
+
+
+/*
+ * Basic data types.
+ * You may need to change these if you have a machine with unusual data
+ * type sizes; for example, "char" not 8 bits, "short" not 16 bits,
+ * or "long" not 32 bits.  We don't care whether "int" is 16 or 32 bits,
+ * but it had better be at least 16.
+ */
+
+/* Representation of a single sample (pixel element value).
+ * We frequently allocate large arrays of these, so it's important to keep
+ * them small.  But if you have memory to burn and access to char or short
+ * arrays is very slow on your hardware, you might want to change these.
+ */
+
+
+#ifdef _MSC_VER
+// Disable the following warnings for Visual Studio
+// This is a warning you get from visual studio 2005 about things in the standard C++
+// library being "deprecated."  I checked the C++ standard and it doesn't say jack 
+// about any of them (I checked the searchable PDF).   So this warning is total Bunk.
+#pragma warning(disable : 4996)
+#endif
+
+
+#if BITS_IN_JSAMPLE == 8
+/* JSAMPLE should be the smallest type that will hold the values 0..255.
+ * You can use a signed char by having GETJSAMPLE mask it with 0xFF.
+ */
+
+#ifdef HAVE_UNSIGNED_CHAR
+
+typedef unsigned char JSAMPLE;
+#define GETJSAMPLE(value)  ((int) (value))
+
+#else /* not HAVE_UNSIGNED_CHAR */
+
+typedef char JSAMPLE;
+#ifdef CHAR_IS_UNSIGNED
+#define GETJSAMPLE(value)  ((int) (value))
+#else
+#define GETJSAMPLE(value)  ((int) (value) & 0xFF)
+#endif /* CHAR_IS_UNSIGNED */
+
+#endif /* HAVE_UNSIGNED_CHAR */
+
+#define MAXJSAMPLE	255
+#define CENTERJSAMPLE	128
+
+#endif /* BITS_IN_JSAMPLE == 8 */
+
+
+#if BITS_IN_JSAMPLE == 12
+/* JSAMPLE should be the smallest type that will hold the values 0..4095.
+ * On nearly all machines "short" will do nicely.
+ */
+
+typedef short JSAMPLE;
+#define GETJSAMPLE(value)  ((int) (value))
+
+#define MAXJSAMPLE	4095
+#define CENTERJSAMPLE	2048
+
+#endif /* BITS_IN_JSAMPLE == 12 */
+
+
+/* Representation of a DCT frequency coefficient.
+ * This should be a signed value of at least 16 bits; "short" is usually OK.
+ * Again, we allocate large arrays of these, but you can change to int
+ * if you have memory to burn and "short" is really slow.
+ */
+
+typedef short JCOEF;
+
+
+/* Compressed datastreams are represented as arrays of JOCTET.
+ * These must be EXACTLY 8 bits wide, at least once they are written to
+ * external storage.  Note that when using the stdio data source/destination
+ * managers, this is also the data type passed to fread/fwrite.
+ */
+
+#ifdef HAVE_UNSIGNED_CHAR
+
+typedef unsigned char JOCTET;
+#define GETJOCTET(value)  (value)
+
+#else /* not HAVE_UNSIGNED_CHAR */
+
+typedef char JOCTET;
+#ifdef CHAR_IS_UNSIGNED
+#define GETJOCTET(value)  (value)
+#else
+#define GETJOCTET(value)  ((value) & 0xFF)
+#endif /* CHAR_IS_UNSIGNED */
+
+#endif /* HAVE_UNSIGNED_CHAR */
+
+
+/* These typedefs are used for various table entries and so forth.
+ * They must be at least as wide as specified; but making them too big
+ * won't cost a huge amount of memory, so we don't provide special
+ * extraction code like we did for JSAMPLE.  (In other words, these
+ * typedefs live at a different point on the speed/space tradeoff curve.)
+ */
+
+/* unsigned char must hold at least the values 0..255. */
+
+
+/* unsigned short must hold at least the values 0..65535. */
+
+
+
+/* Datatype used for image dimensions.  The JPEG standard only supports
+ * images up to 64K*64K due to 16-bit fields in SOF markers.  Therefore
+ * "unsigned int" is sufficient on all machines.  However, if you need to
+ * handle larger images and you don't mind deviating from the spec, you
+ * can change this datatype.
+ */
+
+typedef unsigned int JDIMENSION;
+
+#define JPEG_MAX_DIMENSION  65500L  /* a tad under 64K to prevent overflows */
+
+
+/* These macros are used in all function definitions and extern declarations.
+ * You could modify them if you need to change function linkage conventions;
+ * in particular, you'll need to do that to make the library a Windows DLL.
+ * Another application is to make all functions global for use with debuggers
+ * or code profilers that require it.
+ */
+
+/* a function called through method pointers: */
+#define METHODDEF(type)		static type
+/* a function used only in its module: */
+#define LOCAL(type)		static type
+/* a function referenced thru EXTERNs: */
+#define GLOBAL(type)		type
+/*  
+    Use C linking unless we are supposed to be compiling our own copy of
+    libjpeg.  Then let it use C++ linking so that we are less likely to get
+    linker name conflicts with other libraries that happen to statically include
+    libjpeg as well.
+*/
+#if defined(__cplusplus) && !defined(DLIB_JPEG_STATIC)
+#define EXTERN(type)		extern "C" type
+#else
+#define EXTERN(type)		extern type
+#endif
+
+
+/* This macro is used to declare a "method", that is, a function pointer.
+ * We want to supply prototype parameters if the compiler can cope.
+ * Note that the arglist parameter must be parenthesized!
+ * Again, you can customize this if you need special linkage keywords.
+ */
+
+#ifdef HAVE_PROTOTYPES
+#define JMETHOD(type,methodname,arglist)  type (*methodname) arglist
+#else
+#define JMETHOD(type,methodname,arglist)  type (*methodname) ()
+#endif
+
+
+/* Here is the pseudo-keyword for declaring pointers that must be "far"
+ * on 80x86 machines.  Most of the specialized coding for 80x86 is handled
+ * by just saying "FAR *" where such a pointer is needed.  In a few places
+ * explicit coding is needed; see uses of the NEED_FAR_POINTERS symbol.
+ */
+
+#ifdef NEED_FAR_POINTERS
+#define FAR  far
+#else
+#ifndef FAR 
+    #define FAR
+#endif
+#endif
+
+
+/*
+ * On a few systems, type boolean and/or its values FALSE, TRUE may appear
+ * in standard header files.  Or you may have conflicts with application-
+ * specific header files that you want to include together with these files.
+ * Defining HAVE_BOOLEAN before including jpeglib.h should make it work.
+ */
+
+#ifndef FALSE			/* in case these macros already exist */
+#define FALSE	0		/* values of boolean */
+#endif
+#ifndef TRUE
+#define TRUE	1
+#endif
+
+
+/*
+ * The remaining options affect code selection within the JPEG library,
+ * but they don't need to be visible to most applications using the library.
+ * To minimize application namespace pollution, the symbols won't be
+ * defined unless JPEG_INTERNALS or JPEG_INTERNAL_OPTIONS has been defined.
+ */
+
+#ifdef JPEG_INTERNALS
+#define JPEG_INTERNAL_OPTIONS
+#endif
+
+#ifdef JPEG_INTERNAL_OPTIONS
+
+
+/*
+ * These defines indicate whether to include various optional functions.
+ * Undefining some of these symbols will produce a smaller but less capable
+ * library.  Note that you can leave certain source files out of the
+ * compilation/linking process if you've #undef'd the corresponding symbols.
+ * (You may HAVE to do that if your compiler doesn't like null source files.)
+ */
+
+/* Arithmetic coding is unsupported for legal reasons.  Complaints to IBM. */
+
+/* Capability options common to encoder and decoder: */
+
+#define DCT_ISLOW_SUPPORTED	/* slow but accurate integer algorithm */
+#define DCT_IFAST_SUPPORTED	/* faster, less accurate integer method */
+#define DCT_FLOAT_SUPPORTED	/* floating-point: accurate, fast on fast HW */
+
+/* Encoder capability options: */
+
+#undef  C_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define C_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
+#define C_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
+#define ENTROPY_OPT_SUPPORTED	    /* Optimization of entropy coding parms? */
+/* Note: if you selected 12-bit data precision, it is dangerous to turn off
+ * ENTROPY_OPT_SUPPORTED.  The standard Huffman tables are only good for 8-bit
+ * precision, so jchuff.c normally uses entropy optimization to compute
+ * usable tables for higher precision.  If you don't want to do optimization,
+ * you'll have to supply different default Huffman tables.
+ * The exact same statements apply for progressive JPEG: the default tables
+ * don't work for progressive mode.  (This may get fixed, however.)
+ */
+#define INPUT_SMOOTHING_SUPPORTED   /* Input image smoothing option? */
+
+/* Decoder capability options: */
+
+#undef  D_ARITH_CODING_SUPPORTED    /* Arithmetic coding back end? */
+#define D_MULTISCAN_FILES_SUPPORTED /* Multiple-scan JPEG files? */
+#define D_PROGRESSIVE_SUPPORTED	    /* Progressive JPEG? (Requires MULTISCAN)*/
+#define SAVE_MARKERS_SUPPORTED	    /* jpeg_save_markers() needed? */
+#define BLOCK_SMOOTHING_SUPPORTED   /* Block smoothing? (Progressive only) */
+#define IDCT_SCALING_SUPPORTED	    /* Output rescaling via IDCT? */
+#undef  UPSAMPLE_SCALING_SUPPORTED  /* Output rescaling at upsample stage? */
+#define UPSAMPLE_MERGING_SUPPORTED  /* Fast path for sloppy upsampling? */
+#define QUANT_1PASS_SUPPORTED	    /* 1-pass color quantization? */
+#define QUANT_2PASS_SUPPORTED	    /* 2-pass color quantization? */
+
+/* more capability options later, no doubt */
+
+
+/*
+ * Ordering of RGB data in scanlines passed to or from the application.
+ * If your application wants to deal with data in the order B,G,R, just
+ * change these macros.  You can also deal with formats such as R,G,B,X
+ * (one extra byte per pixel) by changing RGB_PIXELSIZE.  Note that changing
+ * the offsets will also change the order in which colormap data is organized.
+ * RESTRICTIONS:
+ * 1. The sample applications cjpeg,djpeg do NOT support modified RGB formats.
+ * 2. These macros only affect RGB<=>YCbCr color conversion, so they are not
+ *    useful if you are using JPEG color spaces other than YCbCr or grayscale.
+ * 3. The color quantizer modules will not behave desirably if RGB_PIXELSIZE
+ *    is not 3 (they don't understand about dummy color components!).  So you
+ *    can't use color quantization if you change that value.
+ */
+
+#define RGB_RED		0	/* Offset of Red in an RGB scanline element */
+#define RGB_GREEN	1	/* Offset of Green */
+#define RGB_BLUE	2	/* Offset of Blue */
+#define RGB_PIXELSIZE	3	/* JSAMPLEs per RGB scanline element */
+
+
+/* Definitions for speed-related optimizations. */
+
+
+/* If your compiler supports inline functions, define INLINE
+ * as the inline keyword; otherwise define it as empty.
+ */
+
+#ifndef INLINE
+#ifdef __GNUC__			/* for instance, GNU C knows about inline */
+#define INLINE __inline__
+#endif
+#ifndef INLINE
+#define INLINE			/* default is to define it as empty */
+#endif
+#endif
+
+
+/* On some machines (notably 68000 series) "int" is 32 bits, but multiplying
+ * two 16-bit shorts is faster than multiplying two ints.  Define MULTIPLIER
+ * as short on such a machine.  MULTIPLIER must be at least 16 bits wide.
+ */
+
+#ifndef MULTIPLIER
+#define MULTIPLIER  int		/* type for fastest integer multiply */
+#endif
+
+
+/* FAST_FLOAT should be either float or double, whichever is done faster
+ * by your compiler.  (Note that this type is only used in the floating point
+ * DCT routines, so it only matters if you've defined DCT_FLOAT_SUPPORTED.)
+ * Typically, float is faster in ANSI C compilers, while double is faster in
+ * pre-ANSI compilers (because they insist on converting to double anyway).
+ * The code below therefore chooses float if we have ANSI-style prototypes.
+ */
+
+#ifndef FAST_FLOAT
+#ifdef HAVE_PROTOTYPES
+#define FAST_FLOAT  float
+#else
+#define FAST_FLOAT  double
+#endif
+#endif
+
+#endif /* JPEG_INTERNAL_OPTIONS */
diff --git a/ml/dlib/dlib/external/libjpeg/jpegint.h b/ml/dlib/dlib/external/libjpeg/jpegint.h
new file mode 100644
index 000000000..654067965
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jpegint.h
@@ -0,0 +1,392 @@
+/*
+ * jpegint.h
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file provides common declarations for the various JPEG modules.
+ * These declarations are considered internal to the JPEG library; most
+ * applications using the library shouldn't need to include this file.
+ */
+
+
+/* Declarations for both compression & decompression */
+
+typedef enum {			/* Operating modes for buffer controllers */
+	JBUF_PASS_THRU,		/* Plain stripwise operation */
+	/* Remaining modes require a full-image buffer to have been created */
+	JBUF_SAVE_SOURCE,	/* Run source subobject only, save output */
+	JBUF_CRANK_DEST,	/* Run dest subobject only, using saved data */
+	JBUF_SAVE_AND_PASS	/* Run both subobjects, save output */
+} J_BUF_MODE;
+
+/* Values of global_state field (jdapi.c has some dependencies on ordering!) */
+#define CSTATE_START	100	/* after create_compress */
+#define CSTATE_SCANNING	101	/* start_compress done, write_scanlines OK */
+#define CSTATE_RAW_OK	102	/* start_compress done, write_raw_data OK */
+#define CSTATE_WRCOEFS	103	/* jpeg_write_coefficients done */
+#define DSTATE_START	200	/* after create_decompress */
+#define DSTATE_INHEADER	201	/* reading header markers, no SOS yet */
+#define DSTATE_READY	202	/* found SOS, ready for start_decompress */
+#define DSTATE_PRELOAD	203	/* reading multiscan file in start_decompress*/
+#define DSTATE_PRESCAN	204	/* performing dummy pass for 2-pass quant */
+#define DSTATE_SCANNING	205	/* start_decompress done, read_scanlines OK */
+#define DSTATE_RAW_OK	206	/* start_decompress done, read_raw_data OK */
+#define DSTATE_BUFIMAGE	207	/* expecting jpeg_start_output */
+#define DSTATE_BUFPOST	208	/* looking for SOS/EOI in jpeg_finish_output */
+#define DSTATE_RDCOEFS	209	/* reading file in jpeg_read_coefficients */
+#define DSTATE_STOPPING	210	/* looking for EOI in jpeg_finish_decompress */
+
+
+/* Declarations for compression modules */
+
+/* Master control module */
+struct jpeg_comp_master {
+  JMETHOD(void, prepare_for_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, pass_startup, (j_compress_ptr cinfo));
+  JMETHOD(void, finish_pass, (j_compress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  int call_pass_startup;	/* True if pass_startup must be called */
+  int is_last_pass;		/* True during last pass */
+};
+
+/* Main buffer control (downsampled-data buffer) */
+struct jpeg_c_main_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, process_data, (j_compress_ptr cinfo,
+			       JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			       JDIMENSION in_rows_avail));
+};
+
+/* Compression preprocessing (downsampling input buffer control) */
+struct jpeg_c_prep_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, pre_process_data, (j_compress_ptr cinfo,
+				   JSAMPARRAY input_buf,
+				   JDIMENSION *in_row_ctr,
+				   JDIMENSION in_rows_avail,
+				   JSAMPIMAGE output_buf,
+				   JDIMENSION *out_row_group_ctr,
+				   JDIMENSION out_row_groups_avail));
+};
+
+/* Coefficient buffer control */
+struct jpeg_c_coef_controller {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(int, compress_data, (j_compress_ptr cinfo,
+				   JSAMPIMAGE input_buf));
+};
+
+/* Colorspace conversion */
+struct jpeg_color_converter {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, color_convert, (j_compress_ptr cinfo,
+				JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+				JDIMENSION output_row, int num_rows));
+};
+
+/* Downsampling */
+struct jpeg_downsampler {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  JMETHOD(void, downsample, (j_compress_ptr cinfo,
+			     JSAMPIMAGE input_buf, JDIMENSION in_row_index,
+			     JSAMPIMAGE output_buf,
+			     JDIMENSION out_row_group_index));
+
+  int need_context_rows;	/* TRUE if need rows above & below */
+};
+
+/* Forward DCT (also controls coefficient quantization) */
+struct jpeg_forward_dct {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo));
+  /* perhaps this should be an array??? */
+  JMETHOD(void, forward_DCT, (j_compress_ptr cinfo,
+			      jpeg_component_info * compptr,
+			      JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+			      JDIMENSION start_row, JDIMENSION start_col,
+			      JDIMENSION num_blocks));
+};
+
+/* Entropy encoding */
+struct jpeg_entropy_encoder {
+  JMETHOD(void, start_pass, (j_compress_ptr cinfo, int gather_statistics));
+  JMETHOD(int, encode_mcu, (j_compress_ptr cinfo, JBLOCKROW *MCU_data));
+  JMETHOD(void, finish_pass, (j_compress_ptr cinfo));
+};
+
+/* Marker writing */
+struct jpeg_marker_writer {
+  JMETHOD(void, write_file_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_frame_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_scan_header, (j_compress_ptr cinfo));
+  JMETHOD(void, write_file_trailer, (j_compress_ptr cinfo));
+  JMETHOD(void, write_tables_only, (j_compress_ptr cinfo));
+  /* These routines are exported to allow insertion of extra markers */
+  /* Probably only COM and APPn markers should be written this way */
+  JMETHOD(void, write_marker_header, (j_compress_ptr cinfo, int marker,
+				      unsigned int datalen));
+  JMETHOD(void, write_marker_byte, (j_compress_ptr cinfo, int val));
+};
+
+
+/* Declarations for decompression modules */
+
+/* Master control module */
+struct jpeg_decomp_master {
+  JMETHOD(void, prepare_for_output_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, finish_output_pass, (j_decompress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  int is_dummy_pass;	/* True during 1st pass for 2-pass quant */
+};
+
+/* Input control module */
+struct jpeg_input_controller {
+  JMETHOD(int, consume_input, (j_decompress_ptr cinfo));
+  JMETHOD(void, reset_input_controller, (j_decompress_ptr cinfo));
+  JMETHOD(void, start_input_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, finish_input_pass, (j_decompress_ptr cinfo));
+
+  /* State variables made visible to other modules */
+  int has_multiple_scans;	/* True if file has multiple scans */
+  int eoi_reached;		/* True when EOI has been consumed */
+};
+
+/* Main buffer control (downsampled-data buffer) */
+struct jpeg_d_main_controller {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, process_data, (j_decompress_ptr cinfo,
+			       JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			       JDIMENSION out_rows_avail));
+};
+
+/* Coefficient buffer control */
+struct jpeg_d_coef_controller {
+  JMETHOD(void, start_input_pass, (j_decompress_ptr cinfo));
+  JMETHOD(int, consume_data, (j_decompress_ptr cinfo));
+  JMETHOD(void, start_output_pass, (j_decompress_ptr cinfo));
+  JMETHOD(int, decompress_data, (j_decompress_ptr cinfo,
+				 JSAMPIMAGE output_buf));
+  /* Pointer to array of coefficient virtual arrays, or NULL if none */
+  jvirt_barray_ptr *coef_arrays;
+};
+
+/* Decompression postprocessing (color quantization buffer control) */
+struct jpeg_d_post_controller {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, J_BUF_MODE pass_mode));
+  JMETHOD(void, post_process_data, (j_decompress_ptr cinfo,
+				    JSAMPIMAGE input_buf,
+				    JDIMENSION *in_row_group_ctr,
+				    JDIMENSION in_row_groups_avail,
+				    JSAMPARRAY output_buf,
+				    JDIMENSION *out_row_ctr,
+				    JDIMENSION out_rows_avail));
+};
+
+/* Marker reading & parsing */
+struct jpeg_marker_reader {
+  JMETHOD(void, reset_marker_reader, (j_decompress_ptr cinfo));
+  /* Read markers until SOS or EOI.
+   * Returns same codes as are defined for jpeg_consume_input:
+   * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+   */
+  JMETHOD(int, read_markers, (j_decompress_ptr cinfo));
+  /* Read a restart marker --- exported for use by entropy decoder only */
+  jpeg_marker_parser_method read_restart_marker;
+
+  /* State of marker reader --- nominally internal, but applications
+   * supplying COM or APPn handlers might like to know the state.
+   */
+  int saw_SOI;		/* found SOI? */
+  int saw_SOF;		/* found SOF? */
+  int next_restart_num;		/* next restart number expected (0-7) */
+  unsigned int discarded_bytes;	/* # of bytes skipped looking for a marker */
+};
+
+/* Entropy decoding */
+struct jpeg_entropy_decoder {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(int, decode_mcu, (j_decompress_ptr cinfo,
+				JBLOCKROW *MCU_data));
+
+  /* This is here to share code between baseline and progressive decoders; */
+  /* other modules probably should not use it */
+  int insufficient_data;	/* set TRUE after emitting warning */
+};
+
+/* Inverse DCT (also performs dequantization) */
+typedef JMETHOD(void, inverse_DCT_method_ptr,
+		(j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col));
+
+struct jpeg_inverse_dct {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  /* It is useful to allow each component to have a separate IDCT method. */
+  inverse_DCT_method_ptr inverse_DCT[MAX_COMPONENTS];
+};
+
+/* Upsampling (note that upsampler must also call color converter) */
+struct jpeg_upsampler {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, upsample, (j_decompress_ptr cinfo,
+			   JSAMPIMAGE input_buf,
+			   JDIMENSION *in_row_group_ctr,
+			   JDIMENSION in_row_groups_avail,
+			   JSAMPARRAY output_buf,
+			   JDIMENSION *out_row_ctr,
+			   JDIMENSION out_rows_avail));
+
+  int need_context_rows;	/* TRUE if need rows above & below */
+};
+
+/* Colorspace conversion */
+struct jpeg_color_deconverter {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, color_convert, (j_decompress_ptr cinfo,
+				JSAMPIMAGE input_buf, JDIMENSION input_row,
+				JSAMPARRAY output_buf, int num_rows));
+};
+
+/* Color quantization or color precision reduction */
+struct jpeg_color_quantizer {
+  JMETHOD(void, start_pass, (j_decompress_ptr cinfo, int is_pre_scan));
+  JMETHOD(void, color_quantize, (j_decompress_ptr cinfo,
+				 JSAMPARRAY input_buf, JSAMPARRAY output_buf,
+				 int num_rows));
+  JMETHOD(void, finish_pass, (j_decompress_ptr cinfo));
+  JMETHOD(void, new_color_map, (j_decompress_ptr cinfo));
+};
+
+
+/* Miscellaneous useful macros */
+
+#undef MAX
+#define MAX(a,b)	((a) > (b) ? (a) : (b))
+#undef MIN
+#define MIN(a,b)	((a) < (b) ? (a) : (b))
+
+
+/* We assume that right shift corresponds to signed division by 2 with
+ * rounding towards minus infinity.  This is correct for typical "arithmetic
+ * shift" instructions that shift in copies of the sign bit.  But some
+ * C compilers implement >> with an unsigned shift.  For these machines you
+ * must define RIGHT_SHIFT_IS_UNSIGNED.
+ * RIGHT_SHIFT provides a proper signed right shift of an long quantity.
+ * It is only applied with constant shift counts.  SHIFT_TEMPS must be
+ * included in the variables of any routine using RIGHT_SHIFT.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define SHIFT_TEMPS	long shift_temp;
+#define RIGHT_SHIFT(x,shft)  \
+	((shift_temp = (x)) < 0 ? \
+	 (shift_temp >> (shft)) | ((~((long) 0)) << (32-(shft))) : \
+	 (shift_temp >> (shft)))
+#else
+#define SHIFT_TEMPS
+#define RIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jinit_compress_master	jICompress
+#define jinit_c_master_control	jICMaster
+#define jinit_c_main_controller	jICMainC
+#define jinit_c_prep_controller	jICPrepC
+#define jinit_c_coef_controller	jICCoefC
+#define jinit_color_converter	jICColor
+#define jinit_downsampler	jIDownsampler
+#define jinit_forward_dct	jIFDCT
+#define jinit_huff_encoder	jIHEncoder
+#define jinit_phuff_encoder	jIPHEncoder
+#define jinit_marker_writer	jIMWriter
+#define jinit_master_decompress	jIDMaster
+#define jinit_d_main_controller	jIDMainC
+#define jinit_d_coef_controller	jIDCoefC
+#define jinit_d_post_controller	jIDPostC
+#define jinit_input_controller	jIInCtlr
+#define jinit_marker_reader	jIMReader
+#define jinit_huff_decoder	jIHDecoder
+#define jinit_phuff_decoder	jIPHDecoder
+#define jinit_inverse_dct	jIIDCT
+#define jinit_upsampler		jIUpsampler
+#define jinit_color_deconverter	jIDColor
+#define jinit_1pass_quantizer	jI1Quant
+#define jinit_2pass_quantizer	jI2Quant
+#define jinit_merged_upsampler	jIMUpsampler
+#define jinit_memory_mgr	jIMemMgr
+#define jdiv_round_up		jDivRound
+#define jround_up		jRound
+#define jcopy_sample_rows	jCopySamples
+#define jcopy_block_row		jCopyBlocks
+#define jzero_far		jZeroFar
+#define jpeg_zigzag_order	jZIGTable
+#define jpeg_natural_order	jZAGTable
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Compression module initialization routines */
+EXTERN(void) jinit_compress_master JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_c_master_control JPP((j_compress_ptr cinfo,
+					 int transcode_only));
+EXTERN(void) jinit_c_main_controller JPP((j_compress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_c_prep_controller JPP((j_compress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_c_coef_controller JPP((j_compress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_color_converter JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_downsampler JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_forward_dct JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_huff_encoder JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_phuff_encoder JPP((j_compress_ptr cinfo));
+EXTERN(void) jinit_marker_writer JPP((j_compress_ptr cinfo));
+/* Decompression module initialization routines */
+EXTERN(void) jinit_master_decompress JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_d_main_controller JPP((j_decompress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_d_coef_controller JPP((j_decompress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_d_post_controller JPP((j_decompress_ptr cinfo,
+					  int need_full_buffer));
+EXTERN(void) jinit_input_controller JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_marker_reader JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_huff_decoder JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_phuff_decoder JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_inverse_dct JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_upsampler JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_color_deconverter JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_1pass_quantizer JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_2pass_quantizer JPP((j_decompress_ptr cinfo));
+EXTERN(void) jinit_merged_upsampler JPP((j_decompress_ptr cinfo));
+/* Memory manager initialization */
+EXTERN(void) jinit_memory_mgr JPP((j_common_ptr cinfo));
+
+/* Utility routines in jutils.c */
+EXTERN(long) jdiv_round_up JPP((long a, long b));
+EXTERN(long) jround_up JPP((long a, long b));
+EXTERN(void) jcopy_sample_rows JPP((JSAMPARRAY input_array, int source_row,
+				    JSAMPARRAY output_array, int dest_row,
+				    int num_rows, JDIMENSION num_cols));
+EXTERN(void) jcopy_block_row JPP((JBLOCKROW input_row, JBLOCKROW output_row,
+				  JDIMENSION num_blocks));
+EXTERN(void) jzero_far JPP((void FAR * target, size_t bytestozero));
+/* Constant tables in jutils.c */
+#if 0				/* This table is not actually needed in v6a */
+extern const int jpeg_zigzag_order[]; /* natural coef order to zigzag order */
+#endif
+extern const int jpeg_natural_order[]; /* zigzag coef order to natural order */
+
+/* Suppress undefined-structure complaints if necessary. */
+
+#ifdef INCOMPLETE_TYPES_BROKEN
+#ifndef AM_MEMORY_MANAGER	/* only jmemmgr.c defines these */
+struct jvirt_sarray_control { long dummy; };
+struct jvirt_barray_control { long dummy; };
+#endif
+#endif /* INCOMPLETE_TYPES_BROKEN */
diff --git a/ml/dlib/dlib/external/libjpeg/jpeglib.h b/ml/dlib/dlib/external/libjpeg/jpeglib.h
new file mode 100644
index 000000000..e611602d2
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jpeglib.h
@@ -0,0 +1,1096 @@
+/*
+ * jpeglib.h
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the application interface for the JPEG library.
+ * Most applications using the library need only include this file,
+ * and perhaps jerror.h if they want to know the exact error codes.
+ */
+
+#ifndef JPEGLIB_H
+#define JPEGLIB_H
+
+/*
+ * First we include the configuration files that record how this
+ * installation of the JPEG library is set up.  jconfig.h can be
+ * generated automatically for many systems.  jmorecfg.h contains
+ * manual configuration options that most people need not worry about.
+ */
+
+#ifndef JCONFIG_INCLUDED	/* in case jinclude.h already did */
+#include "jconfig.h"		/* widely used configuration options */
+#endif
+#include "jmorecfg.h"		/* seldom changed options */
+
+
+/* Version ID for the JPEG library.
+ * Might be useful for tests like "#if JPEG_LIB_VERSION >= 60".
+ */
+
+#define JPEG_LIB_VERSION  62	/* Version 6b */
+
+
+/* Various constants determining the sizes of things.
+ * All of these are specified by the JPEG standard, so don't change them
+ * if you want to be compatible.
+ */
+
+#define DCTSIZE		    8	/* The basic DCT block is 8x8 samples */
+#define DCTSIZE2	    64	/* DCTSIZE squared; # of elements in a block */
+#define NUM_QUANT_TBLS      4	/* Quantization tables are numbered 0..3 */
+#define NUM_HUFF_TBLS       4	/* Huffman tables are numbered 0..3 */
+#define NUM_ARITH_TBLS      16	/* Arith-coding tables are numbered 0..15 */
+#define MAX_COMPS_IN_SCAN   4	/* JPEG limit on # of components in one scan */
+#define MAX_SAMP_FACTOR     4	/* JPEG limit on sampling factors */
+/* Unfortunately, some bozo at Adobe saw no reason to be bound by the standard;
+ * the PostScript DCT filter can emit files with many more than 10 blocks/MCU.
+ * If you happen to run across such a file, you can up D_MAX_BLOCKS_IN_MCU
+ * to handle it.  We even let you do this from the jconfig.h file.  However,
+ * we strongly discourage changing C_MAX_BLOCKS_IN_MCU; just because Adobe
+ * sometimes emits noncompliant files doesn't mean you should too.
+ */
+#define C_MAX_BLOCKS_IN_MCU   10 /* compressor's limit on blocks per MCU */
+#ifndef D_MAX_BLOCKS_IN_MCU
+#define D_MAX_BLOCKS_IN_MCU   10 /* decompressor's limit on blocks per MCU */
+#endif
+
+
+/* Data structures for images (arrays of samples and of DCT coefficients).
+ * On 80x86 machines, the image arrays are too big for near pointers,
+ * but the pointer arrays can fit in near memory.
+ */
+
+typedef JSAMPLE FAR *JSAMPROW;	/* ptr to one image row of pixel samples. */
+typedef JSAMPROW *JSAMPARRAY;	/* ptr to some rows (a 2-D sample array) */
+typedef JSAMPARRAY *JSAMPIMAGE;	/* a 3-D sample array: top index is color */
+
+typedef JCOEF JBLOCK[DCTSIZE2];	/* one block of coefficients */
+typedef JBLOCK FAR *JBLOCKROW;	/* pointer to one row of coefficient blocks */
+typedef JBLOCKROW *JBLOCKARRAY;		/* a 2-D array of coefficient blocks */
+typedef JBLOCKARRAY *JBLOCKIMAGE;	/* a 3-D array of coefficient blocks */
+
+typedef JCOEF FAR *JCOEFPTR;	/* useful in a couple of places */
+
+
+/* Types for JPEG compression parameters and working tables. */
+
+
+/* DCT coefficient quantization tables. */
+
+typedef struct {
+  /* This array gives the coefficient quantizers in natural array order
+   * (not the zigzag order in which they are stored in a JPEG DQT marker).
+   * CAUTION: IJG versions prior to v6a kept this array in zigzag order.
+   */
+  unsigned short quantval[DCTSIZE2];	/* quantization step for each coefficient */
+  /* This field is used only during compression.  It's initialized FALSE when
+   * the table is created, and set TRUE when it's been output to the file.
+   * You could suppress output of a table by setting this to TRUE.
+   * (See jpeg_suppress_tables for an example.)
+   */
+  int sent_table;		/* TRUE when table has been output */
+} JQUANT_TBL;
+
+
+/* Huffman coding tables. */
+
+typedef struct {
+  /* These two fields directly represent the contents of a JPEG DHT marker */
+  unsigned char bits[17];		/* bits[k] = # of symbols with codes of */
+				/* length k bits; bits[0] is unused */
+  unsigned char huffval[256];		/* The symbols, in order of incr code length */
+  /* This field is used only during compression.  It's initialized FALSE when
+   * the table is created, and set TRUE when it's been output to the file.
+   * You could suppress output of a table by setting this to TRUE.
+   * (See jpeg_suppress_tables for an example.)
+   */
+  int sent_table;		/* TRUE when table has been output */
+} JHUFF_TBL;
+
+
+/* Basic info about one component (color channel). */
+
+typedef struct {
+  /* These values are fixed over the whole image. */
+  /* For compression, they must be supplied by parameter setup; */
+  /* for decompression, they are read from the SOF marker. */
+  int component_id;		/* identifier for this component (0..255) */
+  int component_index;		/* its index in SOF or cinfo->comp_info[] */
+  int h_samp_factor;		/* horizontal sampling factor (1..4) */
+  int v_samp_factor;		/* vertical sampling factor (1..4) */
+  int quant_tbl_no;		/* quantization table selector (0..3) */
+  /* These values may vary between scans. */
+  /* For compression, they must be supplied by parameter setup; */
+  /* for decompression, they are read from the SOS marker. */
+  /* The decompressor output side may not use these variables. */
+  int dc_tbl_no;		/* DC entropy table selector (0..3) */
+  int ac_tbl_no;		/* AC entropy table selector (0..3) */
+  
+  /* Remaining fields should be treated as private by applications. */
+  
+  /* These values are computed during compression or decompression startup: */
+  /* Component's size in DCT blocks.
+   * Any dummy blocks added to complete an MCU are not counted; therefore
+   * these values do not depend on whether a scan is interleaved or not.
+   */
+  JDIMENSION width_in_blocks;
+  JDIMENSION height_in_blocks;
+  /* Size of a DCT block in samples.  Always DCTSIZE for compression.
+   * For decompression this is the size of the output from one DCT block,
+   * reflecting any scaling we choose to apply during the IDCT step.
+   * Values of 1,2,4,8 are likely to be supported.  Note that different
+   * components may receive different IDCT scalings.
+   */
+  int DCT_scaled_size;
+  /* The downsampled dimensions are the component's actual, unpadded number
+   * of samples at the main buffer (preprocessing/compression interface), thus
+   * downsampled_width = ceil(image_width * Hi/Hmax)
+   * and similarly for height.  For decompression, IDCT scaling is included, so
+   * downsampled_width = ceil(image_width * Hi/Hmax * DCT_scaled_size/DCTSIZE)
+   */
+  JDIMENSION downsampled_width;	 /* actual width in samples */
+  JDIMENSION downsampled_height; /* actual height in samples */
+  /* This flag is used only for decompression.  In cases where some of the
+   * components will be ignored (eg grayscale output from YCbCr image),
+   * we can skip most computations for the unused components.
+   */
+  int component_needed;	/* do we need the value of this component? */
+
+  /* These values are computed before starting a scan of the component. */
+  /* The decompressor output side may not use these variables. */
+  int MCU_width;		/* number of blocks per MCU, horizontally */
+  int MCU_height;		/* number of blocks per MCU, vertically */
+  int MCU_blocks;		/* MCU_width * MCU_height */
+  int MCU_sample_width;		/* MCU width in samples, MCU_width*DCT_scaled_size */
+  int last_col_width;		/* # of non-dummy blocks across in last MCU */
+  int last_row_height;		/* # of non-dummy blocks down in last MCU */
+
+  /* Saved quantization table for component; NULL if none yet saved.
+   * See jdinput.c comments about the need for this information.
+   * This field is currently used only for decompression.
+   */
+  JQUANT_TBL * quant_table;
+
+  /* Private per-component storage for DCT or IDCT subsystem. */
+  void * dct_table;
+} jpeg_component_info;
+
+
+/* The script for encoding a multiple-scan file is an array of these: */
+
+typedef struct {
+  int comps_in_scan;		/* number of components encoded in this scan */
+  int component_index[MAX_COMPS_IN_SCAN]; /* their SOF/comp_info[] indexes */
+  int Ss, Se;			/* progressive JPEG spectral selection parms */
+  int Ah, Al;			/* progressive JPEG successive approx. parms */
+} jpeg_scan_info;
+
+/* The decompressor can save APPn and COM markers in a list of these: */
+
+typedef struct jpeg_marker_struct FAR * jpeg_saved_marker_ptr;
+
+struct jpeg_marker_struct {
+  jpeg_saved_marker_ptr next;	/* next in list, or NULL */
+  unsigned char marker;			/* marker code: JPEG_COM, or JPEG_APP0+n */
+  unsigned int original_length;	/* # bytes of data in the file */
+  unsigned int data_length;	/* # bytes of data saved at data[] */
+  JOCTET FAR * data;		/* the data contained in the marker */
+  /* the marker length word is not counted in data_length or original_length */
+};
+
+/* Known color spaces. */
+
+typedef enum {
+	JCS_UNKNOWN,		/* error/unspecified */
+	JCS_GRAYSCALE,		/* monochrome */
+	JCS_RGB,		/* red/green/blue */
+	JCS_YCbCr,		/* Y/Cb/Cr (also known as YUV) */
+	JCS_CMYK,		/* C/M/Y/K */
+	JCS_YCCK		/* Y/Cb/Cr/K */
+} J_COLOR_SPACE;
+
+/* DCT/IDCT algorithm options. */
+
+typedef enum {
+	JDCT_ISLOW,		/* slow but accurate integer algorithm */
+	JDCT_IFAST,		/* faster, less accurate integer method */
+	JDCT_FLOAT		/* floating-point: accurate, fast on fast HW */
+} J_DCT_METHOD;
+
+#ifndef JDCT_DEFAULT		/* may be overridden in jconfig.h */
+#define JDCT_DEFAULT  JDCT_ISLOW
+#endif
+#ifndef JDCT_FASTEST		/* may be overridden in jconfig.h */
+#define JDCT_FASTEST  JDCT_IFAST
+#endif
+
+/* Dithering options for decompression. */
+
+typedef enum {
+	JDITHER_NONE,		/* no dithering */
+	JDITHER_ORDERED,	/* simple ordered dither */
+	JDITHER_FS		/* Floyd-Steinberg error diffusion dither */
+} J_DITHER_MODE;
+
+
+/* Common fields between JPEG compression and decompression master structs. */
+
+#define jpeg_common_fields \
+  struct jpeg_error_mgr * err;	/* Error handler module */\
+  struct jpeg_memory_mgr * mem;	/* Memory manager module */\
+  struct jpeg_progress_mgr * progress; /* Progress monitor, or NULL if none */\
+  void * client_data;		/* Available for use by application */\
+  int is_decompressor;	/* So common code can tell which is which */\
+  int global_state		/* For checking call sequence validity */
+
+/* Routines that are to be used by both halves of the library are declared
+ * to receive a pointer to this structure.  There are no actual instances of
+ * jpeg_common_struct, only of jpeg_compress_struct and jpeg_decompress_struct.
+ */
+struct jpeg_common_struct {
+  jpeg_common_fields;		/* Fields common to both master struct types */
+  /* Additional fields follow in an actual jpeg_compress_struct or
+   * jpeg_decompress_struct.  All three structs must agree on these
+   * initial fields!  (This would be a lot cleaner in C++.)
+   */
+};
+
+typedef struct jpeg_common_struct * j_common_ptr;
+typedef struct jpeg_compress_struct * j_compress_ptr;
+typedef struct jpeg_decompress_struct * j_decompress_ptr;
+
+
+/* Master record for a compression instance */
+
+struct jpeg_compress_struct {
+  jpeg_common_fields;		/* Fields shared with jpeg_decompress_struct */
+
+  /* Destination for compressed data */
+  struct jpeg_destination_mgr * dest;
+
+  /* Description of source image --- these fields must be filled in by
+   * outer application before starting compression.  in_color_space must
+   * be correct before you can even call jpeg_set_defaults().
+   */
+
+  JDIMENSION image_width;	/* input image width */
+  JDIMENSION image_height;	/* input image height */
+  int input_components;		/* # of color components in input image */
+  J_COLOR_SPACE in_color_space;	/* colorspace of input image */
+
+  double input_gamma;		/* image gamma of input image */
+
+  /* Compression parameters --- these fields must be set before calling
+   * jpeg_start_compress().  We recommend calling jpeg_set_defaults() to
+   * initialize everything to reasonable defaults, then changing anything
+   * the application specifically wants to change.  That way you won't get
+   * burnt when new parameters are added.  Also note that there are several
+   * helper routines to simplify changing parameters.
+   */
+
+  int data_precision;		/* bits of precision in image data */
+
+  int num_components;		/* # of color components in JPEG image */
+  J_COLOR_SPACE jpeg_color_space; /* colorspace of JPEG image */
+
+  jpeg_component_info * comp_info;
+  /* comp_info[i] describes component that appears i'th in SOF */
+  
+  JQUANT_TBL * quant_tbl_ptrs[NUM_QUANT_TBLS];
+  /* ptrs to coefficient quantization tables, or NULL if not defined */
+  
+  JHUFF_TBL * dc_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  JHUFF_TBL * ac_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  /* ptrs to Huffman coding tables, or NULL if not defined */
+  
+  unsigned char arith_dc_L[NUM_ARITH_TBLS]; /* L values for DC arith-coding tables */
+  unsigned char arith_dc_U[NUM_ARITH_TBLS]; /* U values for DC arith-coding tables */
+  unsigned char arith_ac_K[NUM_ARITH_TBLS]; /* Kx values for AC arith-coding tables */
+
+  int num_scans;		/* # of entries in scan_info array */
+  const jpeg_scan_info * scan_info; /* script for multi-scan file, or NULL */
+  /* The default value of scan_info is NULL, which causes a single-scan
+   * sequential JPEG file to be emitted.  To create a multi-scan file,
+   * set num_scans and scan_info to point to an array of scan definitions.
+   */
+
+  int raw_data_in;		/* TRUE=caller supplies downsampled data */
+  int arith_code;		/* TRUE=arithmetic coding, FALSE=Huffman */
+  int optimize_coding;	/* TRUE=optimize entropy encoding parms */
+  int CCIR601_sampling;	/* TRUE=first samples are cosited */
+  int smoothing_factor;		/* 1..100, or 0 for no input smoothing */
+  J_DCT_METHOD dct_method;	/* DCT algorithm selector */
+
+  /* The restart interval can be specified in absolute MCUs by setting
+   * restart_interval, or in MCU rows by setting restart_in_rows
+   * (in which case the correct restart_interval will be figured
+   * for each scan).
+   */
+  unsigned int restart_interval; /* MCUs per restart, or 0 for no restart */
+  int restart_in_rows;		/* if > 0, MCU rows per restart interval */
+
+  /* Parameters controlling emission of special markers. */
+
+  int write_JFIF_header;	/* should a JFIF marker be written? */
+  unsigned char JFIF_major_version;	/* What to write for the JFIF version number */
+  unsigned char JFIF_minor_version;
+  /* These three values are not used by the JPEG code, merely copied */
+  /* into the JFIF APP0 marker.  density_unit can be 0 for unknown, */
+  /* 1 for dots/inch, or 2 for dots/cm.  Note that the pixel aspect */
+  /* ratio is defined by X_density/Y_density even when density_unit=0. */
+  unsigned char density_unit;		/* JFIF code for pixel size units */
+  unsigned short X_density;		/* Horizontal pixel density */
+  unsigned short Y_density;		/* Vertical pixel density */
+  int write_Adobe_marker;	/* should an Adobe marker be written? */
+  
+  /* State variable: index of next scanline to be written to
+   * jpeg_write_scanlines().  Application may use this to control its
+   * processing loop, e.g., "while (next_scanline < image_height)".
+   */
+
+  JDIMENSION next_scanline;	/* 0 .. image_height-1  */
+
+  /* Remaining fields are known throughout compressor, but generally
+   * should not be touched by a surrounding application.
+   */
+
+  /*
+   * These fields are computed during compression startup
+   */
+  int progressive_mode;	/* TRUE if scan script uses progressive mode */
+  int max_h_samp_factor;	/* largest h_samp_factor */
+  int max_v_samp_factor;	/* largest v_samp_factor */
+
+  JDIMENSION total_iMCU_rows;	/* # of iMCU rows to be input to coef ctlr */
+  /* The coefficient controller receives data in units of MCU rows as defined
+   * for fully interleaved scans (whether the JPEG file is interleaved or not).
+   * There are v_samp_factor * DCTSIZE sample rows of each component in an
+   * "iMCU" (interleaved MCU) row.
+   */
+  
+  /*
+   * These fields are valid during any one scan.
+   * They describe the components and MCUs actually appearing in the scan.
+   */
+  int comps_in_scan;		/* # of JPEG components in this scan */
+  jpeg_component_info * cur_comp_info[MAX_COMPS_IN_SCAN];
+  /* *cur_comp_info[i] describes component that appears i'th in SOS */
+  
+  JDIMENSION MCUs_per_row;	/* # of MCUs across the image */
+  JDIMENSION MCU_rows_in_scan;	/* # of MCU rows in the image */
+  
+  int blocks_in_MCU;		/* # of DCT blocks per MCU */
+  int MCU_membership[C_MAX_BLOCKS_IN_MCU];
+  /* MCU_membership[i] is index in cur_comp_info of component owning */
+  /* i'th block in an MCU */
+
+  int Ss, Se, Ah, Al;		/* progressive JPEG parameters for scan */
+
+  /*
+   * Links to compression subobjects (methods and private variables of modules)
+   */
+  struct jpeg_comp_master * master;
+  struct jpeg_c_main_controller * main;
+  struct jpeg_c_prep_controller * prep;
+  struct jpeg_c_coef_controller * coef;
+  struct jpeg_marker_writer * marker;
+  struct jpeg_color_converter * cconvert;
+  struct jpeg_downsampler * downsample;
+  struct jpeg_forward_dct * fdct;
+  struct jpeg_entropy_encoder * entropy;
+  jpeg_scan_info * script_space; /* workspace for jpeg_simple_progression */
+  int script_space_size;
+};
+
+
+/* Master record for a decompression instance */
+
+struct jpeg_decompress_struct {
+  jpeg_common_fields;		/* Fields shared with jpeg_compress_struct */
+
+  /* Source of compressed data */
+  struct jpeg_source_mgr * src;
+
+  /* Basic description of image --- filled in by jpeg_read_header(). */
+  /* Application may inspect these values to decide how to process image. */
+
+  JDIMENSION image_width;	/* nominal image width (from SOF marker) */
+  JDIMENSION image_height;	/* nominal image height */
+  int num_components;		/* # of color components in JPEG image */
+  J_COLOR_SPACE jpeg_color_space; /* colorspace of JPEG image */
+
+  /* Decompression processing parameters --- these fields must be set before
+   * calling jpeg_start_decompress().  Note that jpeg_read_header() initializes
+   * them to default values.
+   */
+
+  J_COLOR_SPACE out_color_space; /* colorspace for output */
+
+  unsigned int scale_num, scale_denom; /* fraction by which to scale image */
+
+  double output_gamma;		/* image gamma wanted in output */
+
+  int buffered_image;	/* TRUE=multiple output passes */
+  int raw_data_out;		/* TRUE=downsampled data wanted */
+
+  J_DCT_METHOD dct_method;	/* IDCT algorithm selector */
+  int do_fancy_upsampling;	/* TRUE=apply fancy upsampling */
+  int do_block_smoothing;	/* TRUE=apply interblock smoothing */
+
+  int quantize_colors;	/* TRUE=colormapped output wanted */
+  /* the following are ignored if not quantize_colors: */
+  J_DITHER_MODE dither_mode;	/* type of color dithering to use */
+  int two_pass_quantize;	/* TRUE=use two-pass color quantization */
+  int desired_number_of_colors;	/* max # colors to use in created colormap */
+  /* these are significant only in buffered-image mode: */
+  int enable_1pass_quant;	/* enable future use of 1-pass quantizer */
+  int enable_external_quant;/* enable future use of external colormap */
+  int enable_2pass_quant;	/* enable future use of 2-pass quantizer */
+
+  /* Description of actual output image that will be returned to application.
+   * These fields are computed by jpeg_start_decompress().
+   * You can also use jpeg_calc_output_dimensions() to determine these values
+   * in advance of calling jpeg_start_decompress().
+   */
+
+  JDIMENSION output_width;	/* scaled image width */
+  JDIMENSION output_height;	/* scaled image height */
+  int out_color_components;	/* # of color components in out_color_space */
+  int output_components;	/* # of color components returned */
+  /* output_components is 1 (a colormap index) when quantizing colors;
+   * otherwise it equals out_color_components.
+   */
+  int rec_outbuf_height;	/* min recommended height of scanline buffer */
+  /* If the buffer passed to jpeg_read_scanlines() is less than this many rows
+   * high, space and time will be wasted due to unnecessary data copying.
+   * Usually rec_outbuf_height will be 1 or 2, at most 4.
+   */
+
+  /* When quantizing colors, the output colormap is described by these fields.
+   * The application can supply a colormap by setting colormap non-NULL before
+   * calling jpeg_start_decompress; otherwise a colormap is created during
+   * jpeg_start_decompress or jpeg_start_output.
+   * The map has out_color_components rows and actual_number_of_colors columns.
+   */
+  int actual_number_of_colors;	/* number of entries in use */
+  JSAMPARRAY colormap;		/* The color map as a 2-D pixel array */
+
+  /* State variables: these variables indicate the progress of decompression.
+   * The application may examine these but must not modify them.
+   */
+
+  /* Row index of next scanline to be read from jpeg_read_scanlines().
+   * Application may use this to control its processing loop, e.g.,
+   * "while (output_scanline < output_height)".
+   */
+  JDIMENSION output_scanline;	/* 0 .. output_height-1  */
+
+  /* Current input scan number and number of iMCU rows completed in scan.
+   * These indicate the progress of the decompressor input side.
+   */
+  int input_scan_number;	/* Number of SOS markers seen so far */
+  JDIMENSION input_iMCU_row;	/* Number of iMCU rows completed */
+
+  /* The "output scan number" is the notional scan being displayed by the
+   * output side.  The decompressor will not allow output scan/row number
+   * to get ahead of input scan/row, but it can fall arbitrarily far behind.
+   */
+  int output_scan_number;	/* Nominal scan number being displayed */
+  JDIMENSION output_iMCU_row;	/* Number of iMCU rows read */
+
+  /* Current progression status.  coef_bits[c][i] indicates the precision
+   * with which component c's DCT coefficient i (in zigzag order) is known.
+   * It is -1 when no data has yet been received, otherwise it is the point
+   * transform (shift) value for the most recent scan of the coefficient
+   * (thus, 0 at completion of the progression).
+   * This pointer is NULL when reading a non-progressive file.
+   */
+  int (*coef_bits)[DCTSIZE2];	/* -1 or current Al value for each coef */
+
+  /* Internal JPEG parameters --- the application usually need not look at
+   * these fields.  Note that the decompressor output side may not use
+   * any parameters that can change between scans.
+   */
+
+  /* Quantization and Huffman tables are carried forward across input
+   * datastreams when processing abbreviated JPEG datastreams.
+   */
+
+  JQUANT_TBL * quant_tbl_ptrs[NUM_QUANT_TBLS];
+  /* ptrs to coefficient quantization tables, or NULL if not defined */
+
+  JHUFF_TBL * dc_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  JHUFF_TBL * ac_huff_tbl_ptrs[NUM_HUFF_TBLS];
+  /* ptrs to Huffman coding tables, or NULL if not defined */
+
+  /* These parameters are never carried across datastreams, since they
+   * are given in SOF/SOS markers or defined to be reset by SOI.
+   */
+
+  int data_precision;		/* bits of precision in image data */
+
+  jpeg_component_info * comp_info;
+  /* comp_info[i] describes component that appears i'th in SOF */
+
+  int progressive_mode;	/* TRUE if SOFn specifies progressive mode */
+  int arith_code;		/* TRUE=arithmetic coding, FALSE=Huffman */
+
+  unsigned char arith_dc_L[NUM_ARITH_TBLS]; /* L values for DC arith-coding tables */
+  unsigned char arith_dc_U[NUM_ARITH_TBLS]; /* U values for DC arith-coding tables */
+  unsigned char arith_ac_K[NUM_ARITH_TBLS]; /* Kx values for AC arith-coding tables */
+
+  unsigned int restart_interval; /* MCUs per restart interval, or 0 for no restart */
+
+  /* These fields record data obtained from optional markers recognized by
+   * the JPEG library.
+   */
+  int saw_JFIF_marker;	/* TRUE iff a JFIF APP0 marker was found */
+  /* Data copied from JFIF marker; only valid if saw_JFIF_marker is TRUE: */
+  unsigned char JFIF_major_version;	/* JFIF version number */
+  unsigned char JFIF_minor_version;
+  unsigned char density_unit;		/* JFIF code for pixel size units */
+  unsigned short X_density;		/* Horizontal pixel density */
+  unsigned short Y_density;		/* Vertical pixel density */
+  int saw_Adobe_marker;	/* TRUE iff an Adobe APP14 marker was found */
+  unsigned char Adobe_transform;	/* Color transform code from Adobe marker */
+
+  int CCIR601_sampling;	/* TRUE=first samples are cosited */
+
+  /* Aside from the specific data retained from APPn markers known to the
+   * library, the uninterpreted contents of any or all APPn and COM markers
+   * can be saved in a list for examination by the application.
+   */
+  jpeg_saved_marker_ptr marker_list; /* Head of list of saved markers */
+
+  /* Remaining fields are known throughout decompressor, but generally
+   * should not be touched by a surrounding application.
+   */
+
+  /*
+   * These fields are computed during decompression startup
+   */
+  int max_h_samp_factor;	/* largest h_samp_factor */
+  int max_v_samp_factor;	/* largest v_samp_factor */
+
+  int min_DCT_scaled_size;	/* smallest DCT_scaled_size of any component */
+
+  JDIMENSION total_iMCU_rows;	/* # of iMCU rows in image */
+  /* The coefficient controller's input and output progress is measured in
+   * units of "iMCU" (interleaved MCU) rows.  These are the same as MCU rows
+   * in fully interleaved JPEG scans, but are used whether the scan is
+   * interleaved or not.  We define an iMCU row as v_samp_factor DCT block
+   * rows of each component.  Therefore, the IDCT output contains
+   * v_samp_factor*DCT_scaled_size sample rows of a component per iMCU row.
+   */
+
+  JSAMPLE * sample_range_limit; /* table for fast range-limiting */
+
+  /*
+   * These fields are valid during any one scan.
+   * They describe the components and MCUs actually appearing in the scan.
+   * Note that the decompressor output side must not use these fields.
+   */
+  int comps_in_scan;		/* # of JPEG components in this scan */
+  jpeg_component_info * cur_comp_info[MAX_COMPS_IN_SCAN];
+  /* *cur_comp_info[i] describes component that appears i'th in SOS */
+
+  JDIMENSION MCUs_per_row;	/* # of MCUs across the image */
+  JDIMENSION MCU_rows_in_scan;	/* # of MCU rows in the image */
+
+  int blocks_in_MCU;		/* # of DCT blocks per MCU */
+  int MCU_membership[D_MAX_BLOCKS_IN_MCU];
+  /* MCU_membership[i] is index in cur_comp_info of component owning */
+  /* i'th block in an MCU */
+
+  int Ss, Se, Ah, Al;		/* progressive JPEG parameters for scan */
+
+  /* This field is shared between entropy decoder and marker parser.
+   * It is either zero or the code of a JPEG marker that has been
+   * read from the data source, but has not yet been processed.
+   */
+  int unread_marker;
+
+  /*
+   * Links to decompression subobjects (methods, private variables of modules)
+   */
+  struct jpeg_decomp_master * master;
+  struct jpeg_d_main_controller * main;
+  struct jpeg_d_coef_controller * coef;
+  struct jpeg_d_post_controller * post;
+  struct jpeg_input_controller * inputctl;
+  struct jpeg_marker_reader * marker;
+  struct jpeg_entropy_decoder * entropy;
+  struct jpeg_inverse_dct * idct;
+  struct jpeg_upsampler * upsample;
+  struct jpeg_color_deconverter * cconvert;
+  struct jpeg_color_quantizer * cquantize;
+};
+
+
+/* "Object" declarations for JPEG modules that may be supplied or called
+ * directly by the surrounding application.
+ * As with all objects in the JPEG library, these structs only define the
+ * publicly visible methods and state variables of a module.  Additional
+ * private fields may exist after the public ones.
+ */
+
+
+/* Error handler object */
+
+struct jpeg_error_mgr {
+  /* Error exit handler: does not return to caller */
+  JMETHOD(void, error_exit, (j_common_ptr cinfo));
+  /* Conditionally emit a trace or warning message */
+  JMETHOD(void, emit_message, (j_common_ptr cinfo, int msg_level));
+  /* Routine that actually outputs a trace or error message */
+  JMETHOD(void, output_message, (j_common_ptr cinfo));
+  /* Format a message string for the most recent JPEG error or message */
+  JMETHOD(void, format_message, (j_common_ptr cinfo, char * buffer));
+#define JMSG_LENGTH_MAX  200	/* recommended size of format_message buffer */
+  /* Reset error state variables at start of a new image */
+  JMETHOD(void, reset_error_mgr, (j_common_ptr cinfo));
+  
+  /* The message ID code and any parameters are saved here.
+   * A message can have one string parameter or up to 8 int parameters.
+   */
+  int msg_code;
+#define JMSG_STR_PARM_MAX  80
+  union {
+    int i[8];
+    char s[JMSG_STR_PARM_MAX];
+  } msg_parm;
+  
+  /* Standard state variables for error facility */
+  
+  int trace_level;		/* max msg_level that will be displayed */
+  
+  /* For recoverable corrupt-data errors, we emit a warning message,
+   * but keep going unless emit_message chooses to abort.  emit_message
+   * should count warnings in num_warnings.  The surrounding application
+   * can check for bad data by seeing if num_warnings is nonzero at the
+   * end of processing.
+   */
+  long num_warnings;		/* number of corrupt-data warnings */
+
+  /* These fields point to the table(s) of error message strings.
+   * An application can change the table pointer to switch to a different
+   * message list (typically, to change the language in which errors are
+   * reported).  Some applications may wish to add additional error codes
+   * that will be handled by the JPEG library error mechanism; the second
+   * table pointer is used for this purpose.
+   *
+   * First table includes all errors generated by JPEG library itself.
+   * Error code 0 is reserved for a "no such error string" message.
+   */
+  const char * const * jpeg_message_table; /* Library errors */
+  int last_jpeg_message;    /* Table contains strings 0..last_jpeg_message */
+  /* Second table can be added by application (see cjpeg/djpeg for example).
+   * It contains strings numbered first_addon_message..last_addon_message.
+   */
+  const char * const * addon_message_table; /* Non-library errors */
+  int first_addon_message;	/* code for first string in addon table */
+  int last_addon_message;	/* code for last string in addon table */
+};
+
+
+/* Progress monitor object */
+
+struct jpeg_progress_mgr {
+  JMETHOD(void, progress_monitor, (j_common_ptr cinfo));
+
+  long pass_counter;		/* work units completed in this pass */
+  long pass_limit;		/* total number of work units in this pass */
+  int completed_passes;		/* passes completed so far */
+  int total_passes;		/* total number of passes expected */
+};
+
+
+/* Data destination object for compression */
+
+struct jpeg_destination_mgr {
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+
+  JMETHOD(void, init_destination, (j_compress_ptr cinfo));
+  JMETHOD(int, empty_output_buffer, (j_compress_ptr cinfo));
+  JMETHOD(void, term_destination, (j_compress_ptr cinfo));
+};
+
+
+/* Data source object for decompression */
+
+struct jpeg_source_mgr {
+  const JOCTET * next_input_byte; /* => next byte to read from buffer */
+  size_t bytes_in_buffer;	/* # of bytes remaining in buffer */
+
+  JMETHOD(void, init_source, (j_decompress_ptr cinfo));
+  JMETHOD(int, fill_input_buffer, (j_decompress_ptr cinfo));
+  JMETHOD(void, skip_input_data, (j_decompress_ptr cinfo, long num_bytes));
+  JMETHOD(int, resync_to_restart, (j_decompress_ptr cinfo, int desired));
+  JMETHOD(void, term_source, (j_decompress_ptr cinfo));
+};
+
+
+/* Memory manager object.
+ * Allocates "small" objects (a few K total), "large" objects (tens of K),
+ * and "really big" objects (virtual arrays with backing store if needed).
+ * The memory manager does not allow individual objects to be freed; rather,
+ * each created object is assigned to a pool, and whole pools can be freed
+ * at once.  This is faster and more convenient than remembering exactly what
+ * to free, especially where malloc()/free() are not too speedy.
+ * NB: alloc routines never return NULL.  They exit to error_exit if not
+ * successful.
+ */
+
+#define JPOOL_PERMANENT	0	/* lasts until master record is destroyed */
+#define JPOOL_IMAGE	1	/* lasts until done with image/datastream */
+#define JPOOL_NUMPOOLS	2
+
+typedef struct jvirt_sarray_control * jvirt_sarray_ptr;
+typedef struct jvirt_barray_control * jvirt_barray_ptr;
+
+
+struct jpeg_memory_mgr {
+  /* Method pointers */
+  JMETHOD(void *, alloc_small, (j_common_ptr cinfo, int pool_id,
+				size_t sizeofobject));
+  JMETHOD(void FAR *, alloc_large, (j_common_ptr cinfo, int pool_id,
+				     size_t sizeofobject));
+  JMETHOD(JSAMPARRAY, alloc_sarray, (j_common_ptr cinfo, int pool_id,
+				     JDIMENSION samplesperrow,
+				     JDIMENSION numrows));
+  JMETHOD(JBLOCKARRAY, alloc_barray, (j_common_ptr cinfo, int pool_id,
+				      JDIMENSION blocksperrow,
+				      JDIMENSION numrows));
+  JMETHOD(jvirt_sarray_ptr, request_virt_sarray, (j_common_ptr cinfo,
+						  int pool_id,
+						  int pre_zero,
+						  JDIMENSION samplesperrow,
+						  JDIMENSION numrows,
+						  JDIMENSION maxaccess));
+  JMETHOD(jvirt_barray_ptr, request_virt_barray, (j_common_ptr cinfo,
+						  int pool_id,
+						  int pre_zero,
+						  JDIMENSION blocksperrow,
+						  JDIMENSION numrows,
+						  JDIMENSION maxaccess));
+  JMETHOD(void, realize_virt_arrays, (j_common_ptr cinfo));
+  JMETHOD(JSAMPARRAY, access_virt_sarray, (j_common_ptr cinfo,
+					   jvirt_sarray_ptr ptr,
+					   JDIMENSION start_row,
+					   JDIMENSION num_rows,
+					   int writable));
+  JMETHOD(JBLOCKARRAY, access_virt_barray, (j_common_ptr cinfo,
+					    jvirt_barray_ptr ptr,
+					    JDIMENSION start_row,
+					    JDIMENSION num_rows,
+					    int writable));
+  JMETHOD(void, free_pool, (j_common_ptr cinfo, int pool_id));
+  JMETHOD(void, self_destruct, (j_common_ptr cinfo));
+
+  /* Limit on memory allocation for this JPEG object.  (Note that this is
+   * merely advisory, not a guaranteed maximum; it only affects the space
+   * used for virtual-array buffers.)  May be changed by outer application
+   * after creating the JPEG object.
+   */
+  long max_memory_to_use;
+
+  /* Maximum allocation request accepted by alloc_large. */
+  long max_alloc_chunk;
+};
+
+
+/* Routine signature for application-supplied marker processing methods.
+ * Need not pass marker code since it is stored in cinfo->unread_marker.
+ */
+typedef JMETHOD(int, jpeg_marker_parser_method, (j_decompress_ptr cinfo));
+
+
+/* Declarations for routines called by application.
+ * The JPP macro hides prototype parameters from compilers that can't cope.
+ * Note JPP requires double parentheses.
+ */
+
+#ifdef HAVE_PROTOTYPES
+#define JPP(arglist)	arglist
+#else
+#define JPP(arglist)	()
+#endif
+
+
+/* Short forms of external names for systems with brain-damaged linkers.
+ * We shorten external names to be unique in the first six letters, which
+ * is good enough for all known systems.
+ * (If your compiler itself needs names to be unique in less than 15 
+ * characters, you are out of luck.  Get a better compiler.)
+ */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_std_error		jStdError
+#define jpeg_CreateCompress	jCreaCompress
+#define jpeg_CreateDecompress	jCreaDecompress
+#define jpeg_destroy_compress	jDestCompress
+#define jpeg_destroy_decompress	jDestDecompress
+#define jpeg_stdio_dest		jStdDest
+#define jpeg_stdio_src		jStdSrc
+#define jpeg_set_defaults	jSetDefaults
+#define jpeg_set_colorspace	jSetColorspace
+#define jpeg_default_colorspace	jDefColorspace
+#define jpeg_set_quality	jSetQuality
+#define jpeg_set_linear_quality	jSetLQuality
+#define jpeg_add_quant_table	jAddQuantTable
+#define jpeg_quality_scaling	jQualityScaling
+#define jpeg_simple_progression	jSimProgress
+#define jpeg_suppress_tables	jSuppressTables
+#define jpeg_alloc_quant_table	jAlcQTable
+#define jpeg_alloc_huff_table	jAlcHTable
+#define jpeg_start_compress	jStrtCompress
+#define jpeg_write_scanlines	jWrtScanlines
+#define jpeg_finish_compress	jFinCompress
+#define jpeg_write_raw_data	jWrtRawData
+#define jpeg_write_marker	jWrtMarker
+#define jpeg_write_m_header	jWrtMHeader
+#define jpeg_write_m_byte	jWrtMByte
+#define jpeg_write_tables	jWrtTables
+#define jpeg_read_header	jReadHeader
+#define jpeg_start_decompress	jStrtDecompress
+#define jpeg_read_scanlines	jReadScanlines
+#define jpeg_finish_decompress	jFinDecompress
+#define jpeg_read_raw_data	jReadRawData
+#define jpeg_has_multiple_scans	jHasMultScn
+#define jpeg_start_output	jStrtOutput
+#define jpeg_finish_output	jFinOutput
+#define jpeg_input_complete	jInComplete
+#define jpeg_new_colormap	jNewCMap
+#define jpeg_consume_input	jConsumeInput
+#define jpeg_calc_output_dimensions	jCalcDimensions
+#define jpeg_save_markers	jSaveMarkers
+#define jpeg_set_marker_processor	jSetMarker
+#define jpeg_read_coefficients	jReadCoefs
+#define jpeg_write_coefficients	jWrtCoefs
+#define jpeg_copy_critical_parameters	jCopyCrit
+#define jpeg_abort_compress	jAbrtCompress
+#define jpeg_abort_decompress	jAbrtDecompress
+#define jpeg_abort		jAbort
+#define jpeg_destroy		jDestroy
+#define jpeg_resync_to_restart	jResyncRestart
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+
+/* Default error-management setup */
+EXTERN(struct jpeg_error_mgr *) jpeg_std_error
+	JPP((struct jpeg_error_mgr * err));
+
+/* Initialization of JPEG compression objects.
+ * jpeg_create_compress() and jpeg_create_decompress() are the exported
+ * names that applications should call.  These expand to calls on
+ * jpeg_CreateCompress and jpeg_CreateDecompress with additional information
+ * passed for version mismatch checking.
+ * NB: you must set up the error-manager BEFORE calling jpeg_create_xxx.
+ */
+#define jpeg_create_compress(cinfo) \
+    jpeg_CreateCompress((cinfo), JPEG_LIB_VERSION, \
+			(size_t) sizeof(struct jpeg_compress_struct))
+#define jpeg_create_decompress(cinfo) \
+    jpeg_CreateDecompress((cinfo), JPEG_LIB_VERSION, \
+			  (size_t) sizeof(struct jpeg_decompress_struct))
+EXTERN(void) jpeg_CreateCompress JPP((j_compress_ptr cinfo,
+				      int version, size_t structsize));
+EXTERN(void) jpeg_CreateDecompress JPP((j_decompress_ptr cinfo,
+					int version, size_t structsize));
+/* Destruction of JPEG compression objects */
+EXTERN(void) jpeg_destroy_compress JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_destroy_decompress JPP((j_decompress_ptr cinfo));
+
+/* Standard data source and destination managers: stdio streams. */
+/* Caller is responsible for opening the file before and closing after. */
+EXTERN(void) jpeg_stdio_dest JPP((j_compress_ptr cinfo, FILE * outfile));
+EXTERN(void) jpeg_stdio_src JPP((j_decompress_ptr cinfo, FILE * infile));
+
+/* Default parameter setup for compression */
+EXTERN(void) jpeg_set_defaults JPP((j_compress_ptr cinfo));
+/* Compression parameter setup aids */
+EXTERN(void) jpeg_set_colorspace JPP((j_compress_ptr cinfo,
+				      J_COLOR_SPACE colorspace));
+EXTERN(void) jpeg_default_colorspace JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_set_quality JPP((j_compress_ptr cinfo, int quality,
+				   int force_baseline));
+EXTERN(void) jpeg_set_linear_quality JPP((j_compress_ptr cinfo,
+					  int scale_factor,
+					  int force_baseline));
+EXTERN(void) jpeg_add_quant_table JPP((j_compress_ptr cinfo, int which_tbl,
+				       const unsigned int *basic_table,
+				       int scale_factor,
+				       int force_baseline));
+EXTERN(int) jpeg_quality_scaling JPP((int quality));
+EXTERN(void) jpeg_simple_progression JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_suppress_tables JPP((j_compress_ptr cinfo,
+				       int suppress));
+EXTERN(JQUANT_TBL *) jpeg_alloc_quant_table JPP((j_common_ptr cinfo));
+EXTERN(JHUFF_TBL *) jpeg_alloc_huff_table JPP((j_common_ptr cinfo));
+
+/* Main entry points for compression */
+EXTERN(void) jpeg_start_compress JPP((j_compress_ptr cinfo,
+				      int write_all_tables));
+EXTERN(JDIMENSION) jpeg_write_scanlines JPP((j_compress_ptr cinfo,
+					     JSAMPARRAY scanlines,
+					     JDIMENSION num_lines));
+EXTERN(void) jpeg_finish_compress JPP((j_compress_ptr cinfo));
+
+/* Replaces jpeg_write_scanlines when writing raw downsampled data. */
+EXTERN(JDIMENSION) jpeg_write_raw_data JPP((j_compress_ptr cinfo,
+					    JSAMPIMAGE data,
+					    JDIMENSION num_lines));
+
+/* Write a special marker.  See libjpeg.doc concerning safe usage. */
+EXTERN(void) jpeg_write_marker
+	JPP((j_compress_ptr cinfo, int marker,
+	     const JOCTET * dataptr, unsigned int datalen));
+/* Same, but piecemeal. */
+EXTERN(void) jpeg_write_m_header
+	JPP((j_compress_ptr cinfo, int marker, unsigned int datalen));
+EXTERN(void) jpeg_write_m_byte
+	JPP((j_compress_ptr cinfo, int val));
+
+/* Alternate compression function: just write an abbreviated table file */
+EXTERN(void) jpeg_write_tables JPP((j_compress_ptr cinfo));
+
+/* Decompression startup: read start of JPEG datastream to see what's there */
+EXTERN(int) jpeg_read_header JPP((j_decompress_ptr cinfo,
+				  int require_image));
+/* Return value is one of: */
+#define JPEG_SUSPENDED		0 /* Suspended due to lack of input data */
+#define JPEG_HEADER_OK		1 /* Found valid image datastream */
+#define JPEG_HEADER_TABLES_ONLY	2 /* Found valid table-specs-only datastream */
+/* If you pass require_image = TRUE (normal case), you need not check for
+ * a TABLES_ONLY return code; an abbreviated file will cause an error exit.
+ * JPEG_SUSPENDED is only possible if you use a data source module that can
+ * give a suspension return (the stdio source module doesn't).
+ */
+
+/* Main entry points for decompression */
+EXTERN(int) jpeg_start_decompress JPP((j_decompress_ptr cinfo));
+EXTERN(JDIMENSION) jpeg_read_scanlines JPP((j_decompress_ptr cinfo,
+					    JSAMPARRAY scanlines,
+					    JDIMENSION max_lines));
+EXTERN(int) jpeg_finish_decompress JPP((j_decompress_ptr cinfo));
+
+/* Replaces jpeg_read_scanlines when reading raw downsampled data. */
+EXTERN(JDIMENSION) jpeg_read_raw_data JPP((j_decompress_ptr cinfo,
+					   JSAMPIMAGE data,
+					   JDIMENSION max_lines));
+
+/* Additional entry points for buffered-image mode. */
+EXTERN(int) jpeg_has_multiple_scans JPP((j_decompress_ptr cinfo));
+EXTERN(int) jpeg_start_output JPP((j_decompress_ptr cinfo,
+				       int scan_number));
+EXTERN(int) jpeg_finish_output JPP((j_decompress_ptr cinfo));
+EXTERN(int) jpeg_input_complete JPP((j_decompress_ptr cinfo));
+EXTERN(void) jpeg_new_colormap JPP((j_decompress_ptr cinfo));
+EXTERN(int) jpeg_consume_input JPP((j_decompress_ptr cinfo));
+/* Return value is one of: */
+/* #define JPEG_SUSPENDED	0    Suspended due to lack of input data */
+#define JPEG_REACHED_SOS	1 /* Reached start of new scan */
+#define JPEG_REACHED_EOI	2 /* Reached end of image */
+#define JPEG_ROW_COMPLETED	3 /* Completed one iMCU row */
+#define JPEG_SCAN_COMPLETED	4 /* Completed last iMCU row of a scan */
+
+/* Precalculate output dimensions for current decompression parameters. */
+EXTERN(void) jpeg_calc_output_dimensions JPP((j_decompress_ptr cinfo));
+
+/* Control saving of COM and APPn markers into marker_list. */
+EXTERN(void) jpeg_save_markers
+	JPP((j_decompress_ptr cinfo, int marker_code,
+	     unsigned int length_limit));
+
+/* Install a special processing method for COM or APPn markers. */
+EXTERN(void) jpeg_set_marker_processor
+	JPP((j_decompress_ptr cinfo, int marker_code,
+	     jpeg_marker_parser_method routine));
+
+/* Read or write raw DCT coefficients --- useful for lossless transcoding. */
+EXTERN(jvirt_barray_ptr *) jpeg_read_coefficients JPP((j_decompress_ptr cinfo));
+EXTERN(void) jpeg_write_coefficients JPP((j_compress_ptr cinfo,
+					  jvirt_barray_ptr * coef_arrays));
+EXTERN(void) jpeg_copy_critical_parameters JPP((j_decompress_ptr srcinfo,
+						j_compress_ptr dstinfo));
+
+/* If you choose to abort compression or decompression before completing
+ * jpeg_finish_(de)compress, then you need to clean up to release memory,
+ * temporary files, etc.  You can just call jpeg_destroy_(de)compress
+ * if you're done with the JPEG object, but if you want to clean it up and
+ * reuse it, call this:
+ */
+EXTERN(void) jpeg_abort_compress JPP((j_compress_ptr cinfo));
+EXTERN(void) jpeg_abort_decompress JPP((j_decompress_ptr cinfo));
+
+/* Generic versions of jpeg_abort and jpeg_destroy that work on either
+ * flavor of JPEG object.  These may be more convenient in some places.
+ */
+EXTERN(void) jpeg_abort JPP((j_common_ptr cinfo));
+EXTERN(void) jpeg_destroy JPP((j_common_ptr cinfo));
+
+/* Default restart-marker-resync procedure for use by data source modules */
+EXTERN(int) jpeg_resync_to_restart JPP((j_decompress_ptr cinfo,
+					    int desired));
+
+
+/* These marker codes are exported since applications and data source modules
+ * are likely to want to use them.
+ */
+
+#define JPEG_RST0	0xD0	/* RST0 marker code */
+#define JPEG_EOI	0xD9	/* EOI marker code */
+#define JPEG_APP0	0xE0	/* APP0 marker code */
+#define JPEG_COM	0xFE	/* COM marker code */
+
+
+/* If we have a brain-damaged compiler that emits warnings (or worse, errors)
+ * for structure definitions that are never filled in, keep it quiet by
+ * supplying dummy definitions for the various substructures.
+ */
+
+#ifdef INCOMPLETE_TYPES_BROKEN
+#ifndef JPEG_INTERNALS		/* will be defined in jpegint.h */
+struct jvirt_sarray_control { long dummy; };
+struct jvirt_barray_control { long dummy; };
+struct jpeg_comp_master { long dummy; };
+struct jpeg_c_main_controller { long dummy; };
+struct jpeg_c_prep_controller { long dummy; };
+struct jpeg_c_coef_controller { long dummy; };
+struct jpeg_marker_writer { long dummy; };
+struct jpeg_color_converter { long dummy; };
+struct jpeg_downsampler { long dummy; };
+struct jpeg_forward_dct { long dummy; };
+struct jpeg_entropy_encoder { long dummy; };
+struct jpeg_decomp_master { long dummy; };
+struct jpeg_d_main_controller { long dummy; };
+struct jpeg_d_coef_controller { long dummy; };
+struct jpeg_d_post_controller { long dummy; };
+struct jpeg_input_controller { long dummy; };
+struct jpeg_marker_reader { long dummy; };
+struct jpeg_entropy_decoder { long dummy; };
+struct jpeg_inverse_dct { long dummy; };
+struct jpeg_upsampler { long dummy; };
+struct jpeg_color_deconverter { long dummy; };
+struct jpeg_color_quantizer { long dummy; };
+#endif /* JPEG_INTERNALS */
+#endif /* INCOMPLETE_TYPES_BROKEN */
+
+
+/*
+ * The JPEG library modules define JPEG_INTERNALS before including this file.
+ * The internal structure declarations are read only when that is true.
+ * Applications using the library should not include jpegint.h, but may wish
+ * to include jerror.h.
+ */
+
+#ifdef JPEG_INTERNALS
+#include "jpegint.h"		/* fetch private declarations */
+#include "jerror.h"		/* fetch error codes too */
+#endif
+
+#endif /* JPEGLIB_H */
diff --git a/ml/dlib/dlib/external/libjpeg/jquant1.cpp b/ml/dlib/dlib/external/libjpeg/jquant1.cpp
new file mode 100644
index 000000000..7582015ad
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jquant1.cpp
@@ -0,0 +1,856 @@
+/*
+ * jquant1.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains 1-pass color quantization (color mapping) routines.
+ * These routines provide mapping to a fixed color map using equally spaced
+ * color values.  Optional Floyd-Steinberg or ordered dithering is available.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef QUANT_1PASS_SUPPORTED
+
+
+/*
+ * The main purpose of 1-pass quantization is to provide a fast, if not very
+ * high quality, colormapped output capability.  A 2-pass quantizer usually
+ * gives better visual quality; however, for quantized grayscale output this
+ * quantizer is perfectly adequate.  Dithering is highly recommended with this
+ * quantizer, though you can turn it off if you really want to.
+ *
+ * In 1-pass quantization the colormap must be chosen in advance of seeing the
+ * image.  We use a map consisting of all combinations of Ncolors[i] color
+ * values for the i'th component.  The Ncolors[] values are chosen so that
+ * their product, the total number of colors, is no more than that requested.
+ * (In most cases, the product will be somewhat less.)
+ *
+ * Since the colormap is orthogonal, the representative value for each color
+ * component can be determined without considering the other components;
+ * then these indexes can be combined into a colormap index by a standard
+ * N-dimensional-array-subscript calculation.  Most of the arithmetic involved
+ * can be precalculated and stored in the lookup table colorindex[].
+ * colorindex[i][j] maps pixel value j in component i to the nearest
+ * representative value (grid plane) for that component; this index is
+ * multiplied by the array stride for component i, so that the
+ * index of the colormap entry closest to a given pixel value is just
+ *    sum( colorindex[component-number][pixel-component-value] )
+ * Aside from being fast, this scheme allows for variable spacing between
+ * representative values with no additional lookup cost.
+ *
+ * If gamma correction has been applied in color conversion, it might be wise
+ * to adjust the color grid spacing so that the representative colors are
+ * equidistant in linear space.  At this writing, gamma correction is not
+ * implemented by jdcolor, so nothing is done here.
+ */
+
+
+/* Declarations for ordered dithering.
+ *
+ * We use a standard 16x16 ordered dither array.  The basic concept of ordered
+ * dithering is described in many references, for instance Dale Schumacher's
+ * chapter II.2 of Graphics Gems II (James Arvo, ed. Academic Press, 1991).
+ * In place of Schumacher's comparisons against a "threshold" value, we add a
+ * "dither" value to the input pixel and then round the result to the nearest
+ * output value.  The dither value is equivalent to (0.5 - threshold) times
+ * the distance between output values.  For ordered dithering, we assume that
+ * the output colors are equally spaced; if not, results will probably be
+ * worse, since the dither may be too much or too little at a given point.
+ *
+ * The normal calculation would be to form pixel value + dither, range-limit
+ * this to 0..MAXJSAMPLE, and then index into the colorindex table as usual.
+ * We can skip the separate range-limiting step by extending the colorindex
+ * table in both directions.
+ */
+
+#define ODITHER_SIZE  16	/* dimension of dither matrix */
+/* NB: if ODITHER_SIZE is not a power of 2, ODITHER_MASK uses will break */
+#define ODITHER_CELLS (ODITHER_SIZE*ODITHER_SIZE)	/* # cells in matrix */
+#define ODITHER_MASK  (ODITHER_SIZE-1) /* mask for wrapping around counters */
+
+typedef int ODITHER_MATRIX[ODITHER_SIZE][ODITHER_SIZE];
+typedef int (*ODITHER_MATRIX_PTR)[ODITHER_SIZE];
+
+static const unsigned char base_dither_matrix[ODITHER_SIZE][ODITHER_SIZE] = {
+  /* Bayer's order-4 dither array.  Generated by the code given in
+   * Stephen Hawley's article "Ordered Dithering" in Graphics Gems I.
+   * The values in this array must range from 0 to ODITHER_CELLS-1.
+   */
+  {   0,192, 48,240, 12,204, 60,252,  3,195, 51,243, 15,207, 63,255 },
+  { 128, 64,176,112,140, 76,188,124,131, 67,179,115,143, 79,191,127 },
+  {  32,224, 16,208, 44,236, 28,220, 35,227, 19,211, 47,239, 31,223 },
+  { 160, 96,144, 80,172,108,156, 92,163, 99,147, 83,175,111,159, 95 },
+  {   8,200, 56,248,  4,196, 52,244, 11,203, 59,251,  7,199, 55,247 },
+  { 136, 72,184,120,132, 68,180,116,139, 75,187,123,135, 71,183,119 },
+  {  40,232, 24,216, 36,228, 20,212, 43,235, 27,219, 39,231, 23,215 },
+  { 168,104,152, 88,164,100,148, 84,171,107,155, 91,167,103,151, 87 },
+  {   2,194, 50,242, 14,206, 62,254,  1,193, 49,241, 13,205, 61,253 },
+  { 130, 66,178,114,142, 78,190,126,129, 65,177,113,141, 77,189,125 },
+  {  34,226, 18,210, 46,238, 30,222, 33,225, 17,209, 45,237, 29,221 },
+  { 162, 98,146, 82,174,110,158, 94,161, 97,145, 81,173,109,157, 93 },
+  {  10,202, 58,250,  6,198, 54,246,  9,201, 57,249,  5,197, 53,245 },
+  { 138, 74,186,122,134, 70,182,118,137, 73,185,121,133, 69,181,117 },
+  {  42,234, 26,218, 38,230, 22,214, 41,233, 25,217, 37,229, 21,213 },
+  { 170,106,154, 90,166,102,150, 86,169,105,153, 89,165,101,149, 85 }
+};
+
+
+/* Declarations for Floyd-Steinberg dithering.
+ *
+ * Errors are accumulated into the array fserrors[], at a resolution of
+ * 1/16th of a pixel count.  The error at a given pixel is propagated
+ * to its not-yet-processed neighbors using the standard F-S fractions,
+ *		...	(here)	7/16
+ *		3/16	5/16	1/16
+ * We work left-to-right on even rows, right-to-left on odd rows.
+ *
+ * We can get away with a single array (holding one row's worth of errors)
+ * by using it to store the current row's errors at pixel columns not yet
+ * processed, but the next row's errors at columns already processed.  We
+ * need only a few extra variables to hold the errors immediately around the
+ * current column.  (If we are lucky, those variables are in registers, but
+ * even if not, they're probably cheaper to access than array elements are.)
+ *
+ * The fserrors[] array is indexed [component#][position].
+ * We provide (#columns + 2) entries per component; the extra entry at each
+ * end saves us from special-casing the first and last pixels.
+ *
+ * Note: on a wide image, we might not have enough room in a PC's near data
+ * segment to hold the error array; so it is allocated with alloc_large.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef short FSERROR;		/* 16 bits should be enough */
+typedef int LOCFSERROR;		/* use 'int' for calculation temps */
+#else
+typedef long FSERROR;		/* may need more than 16 bits */
+typedef long LOCFSERROR;	/* be sure calculation temps are big enough */
+#endif
+
+typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
+
+
+/* Private subobject */
+
+#define MAX_Q_COMPS 4		/* max components I can handle */
+
+typedef struct {
+  struct jpeg_color_quantizer pub; /* public fields */
+
+  /* Initially allocated colormap is saved here */
+  JSAMPARRAY sv_colormap;	/* The color map as a 2-D pixel array */
+  int sv_actual;		/* number of entries in use */
+
+  JSAMPARRAY colorindex;	/* Precomputed mapping for speed */
+  /* colorindex[i][j] = index of color closest to pixel value j in component i,
+   * premultiplied as described above.  Since colormap indexes must fit into
+   * JSAMPLEs, the entries of this array will too.
+   */
+  int is_padded;		/* is the colorindex padded for odither? */
+
+  int Ncolors[MAX_Q_COMPS];	/* # of values alloced to each component */
+
+  /* Variables for ordered dithering */
+  int row_index;		/* cur row's vertical index in dither matrix */
+  ODITHER_MATRIX_PTR odither[MAX_Q_COMPS]; /* one dither array per component */
+
+  /* Variables for Floyd-Steinberg dithering */
+  FSERRPTR fserrors[MAX_Q_COMPS]; /* accumulated errors */
+  int on_odd_row;		/* flag to remember which row we are on */
+} my_cquantizer;
+
+typedef my_cquantizer * my_cquantize_ptr;
+
+
+/*
+ * Policy-making subroutines for create_colormap and create_colorindex.
+ * These routines determine the colormap to be used.  The rest of the module
+ * only assumes that the colormap is orthogonal.
+ *
+ *  * select_ncolors decides how to divvy up the available colors
+ *    among the components.
+ *  * output_value defines the set of representative values for a component.
+ *  * largest_input_value defines the mapping from input values to
+ *    representative values for a component.
+ * Note that the latter two routines may impose different policies for
+ * different components, though this is not currently done.
+ */
+
+
+LOCAL(int)
+select_ncolors (j_decompress_ptr cinfo, int Ncolors[])
+/* Determine allocation of desired colors to components, */
+/* and fill in Ncolors[] array to indicate choice. */
+/* Return value is total number of colors (product of Ncolors[] values). */
+{
+  int nc = cinfo->out_color_components; /* number of color components */
+  int max_colors = cinfo->desired_number_of_colors;
+  int total_colors, iroot, i, j;
+  int changed;
+  long temp;
+  static const int RGB_order[3] = { RGB_GREEN, RGB_RED, RGB_BLUE };
+
+  /* We can allocate at least the nc'th root of max_colors per component. */
+  /* Compute floor(nc'th root of max_colors). */
+  iroot = 1;
+  do {
+    iroot++;
+    temp = iroot;		/* set temp = iroot ** nc */
+    for (i = 1; i < nc; i++)
+      temp *= iroot;
+  } while (temp <= (long) max_colors); /* repeat till iroot exceeds root */
+  iroot--;			/* now iroot = floor(root) */
+
+  /* Must have at least 2 color values per component */
+  if (iroot < 2)
+    ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, (int) temp);
+
+  /* Initialize to iroot color values for each component */
+  total_colors = 1;
+  for (i = 0; i < nc; i++) {
+    Ncolors[i] = iroot;
+    total_colors *= iroot;
+  }
+  /* We may be able to increment the count for one or more components without
+   * exceeding max_colors, though we know not all can be incremented.
+   * Sometimes, the first component can be incremented more than once!
+   * (Example: for 16 colors, we start at 2*2*2, go to 3*2*2, then 4*2*2.)
+   * In RGB colorspace, try to increment G first, then R, then B.
+   */
+  do {
+    changed = FALSE;
+    for (i = 0; i < nc; i++) {
+      j = (cinfo->out_color_space == JCS_RGB ? RGB_order[i] : i);
+      /* calculate new total_colors if Ncolors[j] is incremented */
+      temp = total_colors / Ncolors[j];
+      temp *= Ncolors[j]+1;	/* done in long arith to avoid oflo */
+      if (temp > (long) max_colors)
+	break;			/* won't fit, done with this pass */
+      Ncolors[j]++;		/* OK, apply the increment */
+      total_colors = (int) temp;
+      changed = TRUE;
+    }
+  } while (changed);
+
+  return total_colors;
+}
+
+
+LOCAL(int)
+output_value (j_decompress_ptr , int , int j, int maxj)
+/* Return j'th output value, where j will range from 0 to maxj */
+/* The output values must fall in 0..MAXJSAMPLE in increasing order */
+{
+  /* We always provide values 0 and MAXJSAMPLE for each component;
+   * any additional values are equally spaced between these limits.
+   * (Forcing the upper and lower values to the limits ensures that
+   * dithering can't produce a color outside the selected gamut.)
+   */
+  return (int) (((long) j * MAXJSAMPLE + maxj/2) / maxj);
+}
+
+
+LOCAL(int)
+largest_input_value (j_decompress_ptr , int , int j, int maxj)
+/* Return largest input value that should map to j'th output value */
+/* Must have largest(j=0) >= 0, and largest(j=maxj) >= MAXJSAMPLE */
+{
+  /* Breakpoints are halfway between values returned by output_value */
+  return (int) (((long) (2*j + 1) * MAXJSAMPLE + maxj) / (2*maxj));
+}
+
+
+/*
+ * Create the colormap.
+ */
+
+LOCAL(void)
+create_colormap (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPARRAY colormap;		/* Created colormap */
+  int total_colors;		/* Number of distinct output colors */
+  int i,j,k, nci, blksize, blkdist, ptr, val;
+
+  /* Select number of colors for each component */
+  total_colors = select_ncolors(cinfo, cquantize->Ncolors);
+
+  /* Report selected color counts */
+  if (cinfo->out_color_components == 3)
+    TRACEMS4(cinfo, 1, JTRC_QUANT_3_NCOLORS,
+	     total_colors, cquantize->Ncolors[0],
+	     cquantize->Ncolors[1], cquantize->Ncolors[2]);
+  else
+    TRACEMS1(cinfo, 1, JTRC_QUANT_NCOLORS, total_colors);
+
+  /* Allocate and fill in the colormap. */
+  /* The colors are ordered in the map in standard row-major order, */
+  /* i.e. rightmost (highest-indexed) color changes most rapidly. */
+
+  colormap = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) total_colors, (JDIMENSION) cinfo->out_color_components);
+
+  /* blksize is number of adjacent repeated entries for a component */
+  /* blkdist is distance between groups of identical entries for a component */
+  blkdist = total_colors;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    /* fill in colormap entries for i'th color component */
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    blksize = blkdist / nci;
+    for (j = 0; j < nci; j++) {
+      /* Compute j'th output value (out of nci) for component */
+      val = output_value(cinfo, i, j, nci-1);
+      /* Fill in all colormap entries that have this value of this component */
+      for (ptr = j * blksize; ptr < total_colors; ptr += blkdist) {
+	/* fill in blksize entries beginning at ptr */
+	for (k = 0; k < blksize; k++)
+	  colormap[i][ptr+k] = (JSAMPLE) val;
+      }
+    }
+    blkdist = blksize;		/* blksize of this color is blkdist of next */
+  }
+
+  /* Save the colormap in private storage,
+   * where it will survive color quantization mode changes.
+   */
+  cquantize->sv_colormap = colormap;
+  cquantize->sv_actual = total_colors;
+}
+
+
+/*
+ * Create the color index table.
+ */
+
+LOCAL(void)
+create_colorindex (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPROW indexptr;
+  int i,j,k, nci, blksize, val, pad;
+
+  /* For ordered dither, we pad the color index tables by MAXJSAMPLE in
+   * each direction (input index values can be -MAXJSAMPLE .. 2*MAXJSAMPLE).
+   * This is not necessary in the other dithering modes.  However, we
+   * flag whether it was done in case user changes dithering mode.
+   */
+  if (cinfo->dither_mode == JDITHER_ORDERED) {
+    pad = MAXJSAMPLE*2;
+    cquantize->is_padded = TRUE;
+  } else {
+    pad = 0;
+    cquantize->is_padded = FALSE;
+  }
+
+  cquantize->colorindex = (*cinfo->mem->alloc_sarray)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE,
+     (JDIMENSION) (MAXJSAMPLE+1 + pad),
+     (JDIMENSION) cinfo->out_color_components);
+
+  /* blksize is number of adjacent repeated entries for a component */
+  blksize = cquantize->sv_actual;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    /* fill in colorindex entries for i'th color component */
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    blksize = blksize / nci;
+
+    /* adjust colorindex pointers to provide padding at negative indexes. */
+    if (pad)
+      cquantize->colorindex[i] += MAXJSAMPLE;
+
+    /* in loop, val = index of current output value, */
+    /* and k = largest j that maps to current val */
+    indexptr = cquantize->colorindex[i];
+    val = 0;
+    k = largest_input_value(cinfo, i, 0, nci-1);
+    for (j = 0; j <= MAXJSAMPLE; j++) {
+      while (j > k)		/* advance val if past boundary */
+	k = largest_input_value(cinfo, i, ++val, nci-1);
+      /* premultiply so that no multiplication needed in main processing */
+      indexptr[j] = (JSAMPLE) (val * blksize);
+    }
+    /* Pad at both ends if necessary */
+    if (pad)
+      for (j = 1; j <= MAXJSAMPLE; j++) {
+	indexptr[-j] = indexptr[0];
+	indexptr[MAXJSAMPLE+j] = indexptr[MAXJSAMPLE];
+      }
+  }
+}
+
+
+/*
+ * Create an ordered-dither array for a component having ncolors
+ * distinct output values.
+ */
+
+LOCAL(ODITHER_MATRIX_PTR)
+make_odither_array (j_decompress_ptr cinfo, int ncolors)
+{
+  ODITHER_MATRIX_PTR odither;
+  int j,k;
+  long num,den;
+
+  odither = (ODITHER_MATRIX_PTR)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(ODITHER_MATRIX));
+  /* The inter-value distance for this color is MAXJSAMPLE/(ncolors-1).
+   * Hence the dither value for the matrix cell with fill order f
+   * (f=0..N-1) should be (N-1-2*f)/(2*N) * MAXJSAMPLE/(ncolors-1).
+   * On 16-bit-int machine, be careful to avoid overflow.
+   */
+  den = 2 * ODITHER_CELLS * ((long) (ncolors - 1));
+  for (j = 0; j < ODITHER_SIZE; j++) {
+    for (k = 0; k < ODITHER_SIZE; k++) {
+      num = ((long) (ODITHER_CELLS-1 - 2*((int)base_dither_matrix[j][k])))
+	    * MAXJSAMPLE;
+      /* Ensure round towards zero despite C's lack of consistency
+       * about rounding negative values in integer division...
+       */
+      odither[j][k] = (int) (num<0 ? -((-num)/den) : num/den);
+    }
+  }
+  return odither;
+}
+
+
+/*
+ * Create the ordered-dither tables.
+ * Components having the same number of representative colors may 
+ * share a dither table.
+ */
+
+LOCAL(void)
+create_odither_tables (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  ODITHER_MATRIX_PTR odither;
+  int i, j, nci;
+
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    nci = cquantize->Ncolors[i]; /* # of distinct values for this color */
+    odither = NULL;		/* search for matching prior component */
+    for (j = 0; j < i; j++) {
+      if (nci == cquantize->Ncolors[j]) {
+	odither = cquantize->odither[j];
+	break;
+      }
+    }
+    if (odither == NULL)	/* need a new table? */
+      odither = make_odither_array(cinfo, nci);
+    cquantize->odither[i] = odither;
+  }
+}
+
+
+/*
+ * Map some rows of pixels to the output colormapped representation.
+ */
+
+METHODDEF(void)
+color_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		JSAMPARRAY output_buf, int num_rows)
+/* General case, no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPARRAY colorindex = cquantize->colorindex;
+  int pixcode, ci;
+  JSAMPROW ptrin, ptrout;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  int nc = cinfo->out_color_components;
+
+  for (row = 0; row < num_rows; row++) {
+    ptrin = input_buf[row];
+    ptrout = output_buf[row];
+    for (col = width; col > 0; col--) {
+      pixcode = 0;
+      for (ci = 0; ci < nc; ci++) {
+	pixcode += GETJSAMPLE(colorindex[ci][GETJSAMPLE(*ptrin++)]);
+      }
+      *ptrout++ = (JSAMPLE) pixcode;
+    }
+  }
+}
+
+
+METHODDEF(void)
+color_quantize3 (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		 JSAMPARRAY output_buf, int num_rows)
+/* Fast path for out_color_components==3, no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  int pixcode;
+  JSAMPROW ptrin, ptrout;
+  JSAMPROW colorindex0 = cquantize->colorindex[0];
+  JSAMPROW colorindex1 = cquantize->colorindex[1];
+  JSAMPROW colorindex2 = cquantize->colorindex[2];
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    ptrin = input_buf[row];
+    ptrout = output_buf[row];
+    for (col = width; col > 0; col--) {
+      pixcode  = GETJSAMPLE(colorindex0[GETJSAMPLE(*ptrin++)]);
+      pixcode += GETJSAMPLE(colorindex1[GETJSAMPLE(*ptrin++)]);
+      pixcode += GETJSAMPLE(colorindex2[GETJSAMPLE(*ptrin++)]);
+      *ptrout++ = (JSAMPLE) pixcode;
+    }
+  }
+}
+
+
+METHODDEF(void)
+quantize_ord_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		     JSAMPARRAY output_buf, int num_rows)
+/* General case, with ordered dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPROW input_ptr;
+  JSAMPROW output_ptr;
+  JSAMPROW colorindex_ci;
+  int * dither;			/* points to active row of dither matrix */
+  int row_index, col_index;	/* current indexes into dither matrix */
+  int nc = cinfo->out_color_components;
+  int ci;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    /* Initialize output values to 0 so can process components separately */
+    jzero_far((void FAR *) output_buf[row],
+	      (size_t) (width * SIZEOF(JSAMPLE)));
+    row_index = cquantize->row_index;
+    for (ci = 0; ci < nc; ci++) {
+      input_ptr = input_buf[row] + ci;
+      output_ptr = output_buf[row];
+      colorindex_ci = cquantize->colorindex[ci];
+      dither = cquantize->odither[ci][row_index];
+      col_index = 0;
+
+      for (col = width; col > 0; col--) {
+	/* Form pixel value + dither, range-limit to 0..MAXJSAMPLE,
+	 * select output value, accumulate into output code for this pixel.
+	 * Range-limiting need not be done explicitly, as we have extended
+	 * the colorindex table to produce the right answers for out-of-range
+	 * inputs.  The maximum dither is +- MAXJSAMPLE; this sets the
+	 * required amount of padding.
+	 */
+	*output_ptr += colorindex_ci[GETJSAMPLE(*input_ptr)+dither[col_index]];
+	input_ptr += nc;
+	output_ptr++;
+	col_index = (col_index + 1) & ODITHER_MASK;
+      }
+    }
+    /* Advance row index for next row */
+    row_index = (row_index + 1) & ODITHER_MASK;
+    cquantize->row_index = row_index;
+  }
+}
+
+
+METHODDEF(void)
+quantize3_ord_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		      JSAMPARRAY output_buf, int num_rows)
+/* Fast path for out_color_components==3, with ordered dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  int pixcode;
+  JSAMPROW input_ptr;
+  JSAMPROW output_ptr;
+  JSAMPROW colorindex0 = cquantize->colorindex[0];
+  JSAMPROW colorindex1 = cquantize->colorindex[1];
+  JSAMPROW colorindex2 = cquantize->colorindex[2];
+  int * dither0;		/* points to active row of dither matrix */
+  int * dither1;
+  int * dither2;
+  int row_index, col_index;	/* current indexes into dither matrix */
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    row_index = cquantize->row_index;
+    input_ptr = input_buf[row];
+    output_ptr = output_buf[row];
+    dither0 = cquantize->odither[0][row_index];
+    dither1 = cquantize->odither[1][row_index];
+    dither2 = cquantize->odither[2][row_index];
+    col_index = 0;
+
+    for (col = width; col > 0; col--) {
+      pixcode  = GETJSAMPLE(colorindex0[GETJSAMPLE(*input_ptr++) +
+					dither0[col_index]]);
+      pixcode += GETJSAMPLE(colorindex1[GETJSAMPLE(*input_ptr++) +
+					dither1[col_index]]);
+      pixcode += GETJSAMPLE(colorindex2[GETJSAMPLE(*input_ptr++) +
+					dither2[col_index]]);
+      *output_ptr++ = (JSAMPLE) pixcode;
+      col_index = (col_index + 1) & ODITHER_MASK;
+    }
+    row_index = (row_index + 1) & ODITHER_MASK;
+    cquantize->row_index = row_index;
+  }
+}
+
+
+METHODDEF(void)
+quantize_fs_dither (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		    JSAMPARRAY output_buf, int num_rows)
+/* General case, with Floyd-Steinberg dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  LOCFSERROR cur;	/* current error or pixel value */
+  LOCFSERROR belowerr;		/* error for pixel below cur */
+  LOCFSERROR bpreverr;		/* error for below/prev col */
+  LOCFSERROR bnexterr;		/* error for below/next col */
+  LOCFSERROR delta;
+  FSERRPTR errorptr;	/* => fserrors[] at column before current */
+  JSAMPROW input_ptr;
+  JSAMPROW output_ptr;
+  JSAMPROW colorindex_ci;
+  JSAMPROW colormap_ci;
+  int pixcode;
+  int nc = cinfo->out_color_components;
+  int dir;			/* 1 for left-to-right, -1 for right-to-left */
+  int dirnc;			/* dir * nc */
+  int ci;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  JSAMPLE *range_limit = cinfo->sample_range_limit;
+  SHIFT_TEMPS
+
+  for (row = 0; row < num_rows; row++) {
+    /* Initialize output values to 0 so can process components separately */
+    jzero_far((void FAR *) output_buf[row],
+	      (size_t) (width * SIZEOF(JSAMPLE)));
+    for (ci = 0; ci < nc; ci++) {
+      input_ptr = input_buf[row] + ci;
+      output_ptr = output_buf[row];
+      if (cquantize->on_odd_row) {
+	/* work right to left in this row */
+	input_ptr += (width-1) * nc; /* so point to rightmost pixel */
+	output_ptr += width-1;
+	dir = -1;
+	dirnc = -nc;
+	errorptr = cquantize->fserrors[ci] + (width+1); /* => entry after last column */
+      } else {
+	/* work left to right in this row */
+	dir = 1;
+	dirnc = nc;
+	errorptr = cquantize->fserrors[ci]; /* => entry before first column */
+      }
+      colorindex_ci = cquantize->colorindex[ci];
+      colormap_ci = cquantize->sv_colormap[ci];
+      /* Preset error values: no error propagated to first pixel from left */
+      cur = 0;
+      /* and no error propagated to row below yet */
+      belowerr = bpreverr = 0;
+
+      for (col = width; col > 0; col--) {
+	/* cur holds the error propagated from the previous pixel on the
+	 * current line.  Add the error propagated from the previous line
+	 * to form the complete error correction term for this pixel, and
+	 * round the error term (which is expressed * 16) to an integer.
+	 * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
+	 * for either sign of the error value.
+	 * Note: errorptr points to *previous* column's array entry.
+	 */
+	cur = RIGHT_SHIFT(cur + errorptr[dir] + 8, 4);
+	/* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
+	 * The maximum error is +- MAXJSAMPLE; this sets the required size
+	 * of the range_limit array.
+	 */
+	cur += GETJSAMPLE(*input_ptr);
+	cur = GETJSAMPLE(range_limit[cur]);
+	/* Select output value, accumulate into output code for this pixel */
+	pixcode = GETJSAMPLE(colorindex_ci[cur]);
+	*output_ptr += (JSAMPLE) pixcode;
+	/* Compute actual representation error at this pixel */
+	/* Note: we can do this even though we don't have the final */
+	/* pixel code, because the colormap is orthogonal. */
+	cur -= GETJSAMPLE(colormap_ci[pixcode]);
+	/* Compute error fractions to be propagated to adjacent pixels.
+	 * Add these into the running sums, and simultaneously shift the
+	 * next-line error sums left by 1 column.
+	 */
+	bnexterr = cur;
+	delta = cur * 2;
+	cur += delta;		/* form error * 3 */
+	errorptr[0] = (FSERROR) (bpreverr + cur);
+	cur += delta;		/* form error * 5 */
+	bpreverr = belowerr + cur;
+	belowerr = bnexterr;
+	cur += delta;		/* form error * 7 */
+	/* At this point cur contains the 7/16 error value to be propagated
+	 * to the next pixel on the current line, and all the errors for the
+	 * next line have been shifted over. We are therefore ready to move on.
+	 */
+	input_ptr += dirnc;	/* advance input ptr to next column */
+	output_ptr += dir;	/* advance output ptr to next column */
+	errorptr += dir;	/* advance errorptr to current column */
+      }
+      /* Post-loop cleanup: we must unload the final error value into the
+       * final fserrors[] entry.  Note we need not unload belowerr because
+       * it is for the dummy column before or after the actual array.
+       */
+      errorptr[0] = (FSERROR) bpreverr; /* unload prev err into array */
+    }
+    cquantize->on_odd_row = (cquantize->on_odd_row ? FALSE : TRUE);
+  }
+}
+
+
+/*
+ * Allocate workspace for Floyd-Steinberg errors.
+ */
+
+LOCAL(void)
+alloc_fs_workspace (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  size_t arraysize;
+  int i;
+
+  arraysize = (size_t) ((cinfo->output_width + 2) * SIZEOF(FSERROR));
+  for (i = 0; i < cinfo->out_color_components; i++) {
+    cquantize->fserrors[i] = (FSERRPTR)
+      (*cinfo->mem->alloc_large)((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
+  }
+}
+
+
+/*
+ * Initialize for one-pass color quantization.
+ */
+
+METHODDEF(void)
+start_pass_1_quant (j_decompress_ptr cinfo, int )
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  size_t arraysize;
+  int i;
+
+  /* Install my colormap. */
+  cinfo->colormap = cquantize->sv_colormap;
+  cinfo->actual_number_of_colors = cquantize->sv_actual;
+
+  /* Initialize for desired dithering mode. */
+  switch (cinfo->dither_mode) {
+  case JDITHER_NONE:
+    if (cinfo->out_color_components == 3)
+      cquantize->pub.color_quantize = color_quantize3;
+    else
+      cquantize->pub.color_quantize = color_quantize;
+    break;
+  case JDITHER_ORDERED:
+    if (cinfo->out_color_components == 3)
+      cquantize->pub.color_quantize = quantize3_ord_dither;
+    else
+      cquantize->pub.color_quantize = quantize_ord_dither;
+    cquantize->row_index = 0;	/* initialize state for ordered dither */
+    /* If user changed to ordered dither from another mode,
+     * we must recreate the color index table with padding.
+     * This will cost extra space, but probably isn't very likely.
+     */
+    if (! cquantize->is_padded)
+      create_colorindex(cinfo);
+    /* Create ordered-dither tables if we didn't already. */
+    if (cquantize->odither[0] == NULL)
+      create_odither_tables(cinfo);
+    break;
+  case JDITHER_FS:
+    cquantize->pub.color_quantize = quantize_fs_dither;
+    cquantize->on_odd_row = FALSE; /* initialize state for F-S dither */
+    /* Allocate Floyd-Steinberg workspace if didn't already. */
+    if (cquantize->fserrors[0] == NULL)
+      alloc_fs_workspace(cinfo);
+    /* Initialize the propagated errors to zero. */
+    arraysize = (size_t) ((cinfo->output_width + 2) * SIZEOF(FSERROR));
+    for (i = 0; i < cinfo->out_color_components; i++)
+      jzero_far((void FAR *) cquantize->fserrors[i], arraysize);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+    break;
+  }
+}
+
+
+/*
+ * Finish up at the end of the pass.
+ */
+
+METHODDEF(void)
+finish_pass_1_quant (j_decompress_ptr )
+{
+  /* no work in 1-pass case */
+}
+
+
+/*
+ * Switch to a new external colormap between output passes.
+ * Shouldn't get to this module!
+ */
+
+METHODDEF(void)
+new_color_map_1_quant (j_decompress_ptr cinfo)
+{
+  ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+
+/*
+ * Module initialization routine for 1-pass color quantization.
+ */
+
+GLOBAL(void)
+jinit_1pass_quantizer (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize;
+
+  cquantize = (my_cquantize_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_cquantizer));
+  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
+  cquantize->pub.start_pass = start_pass_1_quant;
+  cquantize->pub.finish_pass = finish_pass_1_quant;
+  cquantize->pub.new_color_map = new_color_map_1_quant;
+  cquantize->fserrors[0] = NULL; /* Flag FS workspace not allocated */
+  cquantize->odither[0] = NULL;	/* Also flag odither arrays not allocated */
+
+  /* Make sure my internal arrays won't overflow */
+  if (cinfo->out_color_components > MAX_Q_COMPS)
+    ERREXIT1(cinfo, JERR_QUANT_COMPONENTS, MAX_Q_COMPS);
+  /* Make sure colormap indexes can be represented by JSAMPLEs */
+  if (cinfo->desired_number_of_colors > (MAXJSAMPLE+1))
+    ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXJSAMPLE+1);
+
+  /* Create the colormap and color index table. */
+  create_colormap(cinfo);
+  create_colorindex(cinfo);
+
+  /* Allocate Floyd-Steinberg workspace now if requested.
+   * We do this now since it is FAR storage and may affect the memory
+   * manager's space calculations.  If the user changes to FS dither
+   * mode in a later pass, we will allocate the space then, and will
+   * possibly overrun the max_memory_to_use setting.
+   */
+  if (cinfo->dither_mode == JDITHER_FS)
+    alloc_fs_workspace(cinfo);
+}
+
+#endif /* QUANT_1PASS_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jquant2.cpp b/ml/dlib/dlib/external/libjpeg/jquant2.cpp
new file mode 100644
index 000000000..0d7b5969a
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jquant2.cpp
@@ -0,0 +1,1310 @@
+/*
+ * jquant2.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains 2-pass color quantization (color mapping) routines.
+ * These routines provide selection of a custom color map for an image,
+ * followed by mapping of the image to that color map, with optional
+ * Floyd-Steinberg dithering.
+ * It is also possible to use just the second pass to map to an arbitrary
+ * externally-given color map.
+ *
+ * Note: ordered dithering is not supported, since there isn't any fast
+ * way to compute intercolor distances; it's unclear that ordered dither's
+ * fundamental assumptions even hold with an irregularly spaced color map.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+
+/*
+ * This module implements the well-known Heckbert paradigm for color
+ * quantization.  Most of the ideas used here can be traced back to
+ * Heckbert's seminal paper
+ *   Heckbert, Paul.  "Color Image Quantization for Frame Buffer Display",
+ *   Proc. SIGGRAPH '82, Computer Graphics v.16 #3 (July 1982), pp 297-304.
+ *
+ * In the first pass over the image, we accumulate a histogram showing the
+ * usage count of each possible color.  To keep the histogram to a reasonable
+ * size, we reduce the precision of the input; typical practice is to retain
+ * 5 or 6 bits per color, so that 8 or 4 different input values are counted
+ * in the same histogram cell.
+ *
+ * Next, the color-selection step begins with a box representing the whole
+ * color space, and repeatedly splits the "largest" remaining box until we
+ * have as many boxes as desired colors.  Then the mean color in each
+ * remaining box becomes one of the possible output colors.
+ * 
+ * The second pass over the image maps each input pixel to the closest output
+ * color (optionally after applying a Floyd-Steinberg dithering correction).
+ * This mapping is logically trivial, but making it go fast enough requires
+ * considerable care.
+ *
+ * Heckbert-style quantizers vary a good deal in their policies for choosing
+ * the "largest" box and deciding where to cut it.  The particular policies
+ * used here have proved out well in experimental comparisons, but better ones
+ * may yet be found.
+ *
+ * In earlier versions of the IJG code, this module quantized in YCbCr color
+ * space, processing the raw upsampled data without a color conversion step.
+ * This allowed the color conversion math to be done only once per colormap
+ * entry, not once per pixel.  However, that optimization precluded other
+ * useful optimizations (such as merging color conversion with upsampling)
+ * and it also interfered with desired capabilities such as quantizing to an
+ * externally-supplied colormap.  We have therefore abandoned that approach.
+ * The present code works in the post-conversion color space, typically RGB.
+ *
+ * To improve the visual quality of the results, we actually work in scaled
+ * RGB space, giving G distances more weight than R, and R in turn more than
+ * B.  To do everything in integer math, we must use integer scale factors.
+ * The 2/3/1 scale factors used here correspond loosely to the relative
+ * weights of the colors in the NTSC grayscale equation.
+ * If you want to use this code to quantize a non-RGB color space, you'll
+ * probably need to change these scale factors.
+ */
+
+#define R_SCALE 2		/* scale R distances by this much */
+#define G_SCALE 3		/* scale G distances by this much */
+#define B_SCALE 1		/* and B by this much */
+
+/* Relabel R/G/B as components 0/1/2, respecting the RGB ordering defined
+ * in jmorecfg.h.  As the code stands, it will do the right thing for R,G,B
+ * and B,G,R orders.  If you define some other weird order in jmorecfg.h,
+ * you'll get compile errors until you extend this logic.  In that case
+ * you'll probably want to tweak the histogram sizes too.
+ */
+
+#if RGB_RED == 0
+#define C0_SCALE R_SCALE
+#endif
+#if RGB_BLUE == 0
+#define C0_SCALE B_SCALE
+#endif
+#if RGB_GREEN == 1
+#define C1_SCALE G_SCALE
+#endif
+#if RGB_RED == 2
+#define C2_SCALE R_SCALE
+#endif
+#if RGB_BLUE == 2
+#define C2_SCALE B_SCALE
+#endif
+
+
+/*
+ * First we have the histogram data structure and routines for creating it.
+ *
+ * The number of bits of precision can be adjusted by changing these symbols.
+ * We recommend keeping 6 bits for G and 5 each for R and B.
+ * If you have plenty of memory and cycles, 6 bits all around gives marginally
+ * better results; if you are short of memory, 5 bits all around will save
+ * some space but degrade the results.
+ * To maintain a fully accurate histogram, we'd need to allocate a "long"
+ * (preferably unsigned long) for each cell.  In practice this is overkill;
+ * we can get by with 16 bits per cell.  Few of the cell counts will overflow,
+ * and clamping those that do overflow to the maximum value will give close-
+ * enough results.  This reduces the recommended histogram size from 256Kb
+ * to 128Kb, which is a useful savings on PC-class machines.
+ * (In the second pass the histogram space is re-used for pixel mapping data;
+ * in that capacity, each cell must be able to store zero to the number of
+ * desired colors.  16 bits/cell is plenty for that too.)
+ * Since the JPEG code is intended to run in small memory model on 80x86
+ * machines, we can't just allocate the histogram in one chunk.  Instead
+ * of a true 3-D array, we use a row of pointers to 2-D arrays.  Each
+ * pointer corresponds to a C0 value (typically 2^5 = 32 pointers) and
+ * each 2-D array has 2^6*2^5 = 2048 or 2^6*2^6 = 4096 entries.  Note that
+ * on 80x86 machines, the pointer row is in near memory but the actual
+ * arrays are in far memory (same arrangement as we use for image arrays).
+ */
+
+#define MAXNUMCOLORS  (MAXJSAMPLE+1) /* maximum size of colormap */
+
+/* These will do the right thing for either R,G,B or B,G,R color order,
+ * but you may not like the results for other color orders.
+ */
+#define HIST_C0_BITS  5		/* bits of precision in R/B histogram */
+#define HIST_C1_BITS  6		/* bits of precision in G histogram */
+#define HIST_C2_BITS  5		/* bits of precision in B/R histogram */
+
+/* Number of elements along histogram axes. */
+#define HIST_C0_ELEMS  (1<<HIST_C0_BITS)
+#define HIST_C1_ELEMS  (1<<HIST_C1_BITS)
+#define HIST_C2_ELEMS  (1<<HIST_C2_BITS)
+
+/* These are the amounts to shift an input value to get a histogram index. */
+#define C0_SHIFT  (BITS_IN_JSAMPLE-HIST_C0_BITS)
+#define C1_SHIFT  (BITS_IN_JSAMPLE-HIST_C1_BITS)
+#define C2_SHIFT  (BITS_IN_JSAMPLE-HIST_C2_BITS)
+
+
+typedef unsigned short histcell;	/* histogram cell; prefer an unsigned type */
+
+typedef histcell FAR * histptr;	/* for pointers to histogram cells */
+
+typedef histcell hist1d[HIST_C2_ELEMS]; /* typedefs for the array */
+typedef hist1d FAR * hist2d;	/* type for the 2nd-level pointers */
+typedef hist2d * hist3d;	/* type for top-level pointer */
+
+
+/* Declarations for Floyd-Steinberg dithering.
+ *
+ * Errors are accumulated into the array fserrors[], at a resolution of
+ * 1/16th of a pixel count.  The error at a given pixel is propagated
+ * to its not-yet-processed neighbors using the standard F-S fractions,
+ *		...	(here)	7/16
+ *		3/16	5/16	1/16
+ * We work left-to-right on even rows, right-to-left on odd rows.
+ *
+ * We can get away with a single array (holding one row's worth of errors)
+ * by using it to store the current row's errors at pixel columns not yet
+ * processed, but the next row's errors at columns already processed.  We
+ * need only a few extra variables to hold the errors immediately around the
+ * current column.  (If we are lucky, those variables are in registers, but
+ * even if not, they're probably cheaper to access than array elements are.)
+ *
+ * The fserrors[] array has (#columns + 2) entries; the extra entry at
+ * each end saves us from special-casing the first and last pixels.
+ * Each entry is three values long, one value for each color component.
+ *
+ * Note: on a wide image, we might not have enough room in a PC's near data
+ * segment to hold the error array; so it is allocated with alloc_large.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef short FSERROR;		/* 16 bits should be enough */
+typedef int LOCFSERROR;		/* use 'int' for calculation temps */
+#else
+typedef long FSERROR;		/* may need more than 16 bits */
+typedef long LOCFSERROR;	/* be sure calculation temps are big enough */
+#endif
+
+typedef FSERROR FAR *FSERRPTR;	/* pointer to error array (in FAR storage!) */
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_quantizer pub; /* public fields */
+
+  /* Space for the eventually created colormap is stashed here */
+  JSAMPARRAY sv_colormap;	/* colormap allocated at init time */
+  int desired;			/* desired # of colors = size of colormap */
+
+  /* Variables for accumulating image statistics */
+  hist3d histogram;		/* pointer to the histogram */
+
+  int needs_zeroed;		/* TRUE if next pass must zero histogram */
+
+  /* Variables for Floyd-Steinberg dithering */
+  FSERRPTR fserrors;		/* accumulated errors */
+  int on_odd_row;		/* flag to remember which row we are on */
+  int * error_limiter;		/* table for clamping the applied error */
+} my_cquantizer;
+
+typedef my_cquantizer * my_cquantize_ptr;
+
+
+/*
+ * Prescan some rows of pixels.
+ * In this module the prescan simply updates the histogram, which has been
+ * initialized to zeroes by start_pass.
+ * An output_buf parameter is required by the method signature, but no data
+ * is actually output (in fact the buffer controller is probably passing a
+ * NULL pointer).
+ */
+
+METHODDEF(void)
+prescan_quantize (j_decompress_ptr cinfo, JSAMPARRAY input_buf,
+		  JSAMPARRAY , int num_rows)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  JSAMPROW ptr;
+  histptr histp;
+  hist3d histogram = cquantize->histogram;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    ptr = input_buf[row];
+    for (col = width; col > 0; col--) {
+      /* get pixel value and index into the histogram */
+      histp = & histogram[GETJSAMPLE(ptr[0]) >> C0_SHIFT]
+			 [GETJSAMPLE(ptr[1]) >> C1_SHIFT]
+			 [GETJSAMPLE(ptr[2]) >> C2_SHIFT];
+      /* increment, check for overflow and undo increment if so. */
+      if (++(*histp) <= 0)
+	(*histp)--;
+      ptr += 3;
+    }
+  }
+}
+
+
+/*
+ * Next we have the really interesting routines: selection of a colormap
+ * given the completed histogram.
+ * These routines work with a list of "boxes", each representing a rectangular
+ * subset of the input color space (to histogram precision).
+ */
+
+typedef struct {
+  /* The bounds of the box (inclusive); expressed as histogram indexes */
+  int c0min, c0max;
+  int c1min, c1max;
+  int c2min, c2max;
+  /* The volume (actually 2-norm) of the box */
+  long volume;
+  /* The number of nonzero histogram cells within this box */
+  long colorcount;
+} box;
+
+typedef box * boxptr;
+
+
+LOCAL(boxptr)
+find_biggest_color_pop (boxptr boxlist, int numboxes)
+/* Find the splittable box with the largest color population */
+/* Returns NULL if no splittable boxes remain */
+{
+  boxptr boxp;
+  int i;
+  long maxc = 0;
+  boxptr which = NULL;
+  
+  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
+    if (boxp->colorcount > maxc && boxp->volume > 0) {
+      which = boxp;
+      maxc = boxp->colorcount;
+    }
+  }
+  return which;
+}
+
+
+LOCAL(boxptr)
+find_biggest_volume (boxptr boxlist, int numboxes)
+/* Find the splittable box with the largest (scaled) volume */
+/* Returns NULL if no splittable boxes remain */
+{
+  boxptr boxp;
+  int i;
+  long maxv = 0;
+  boxptr which = NULL;
+  
+  for (i = 0, boxp = boxlist; i < numboxes; i++, boxp++) {
+    if (boxp->volume > maxv) {
+      which = boxp;
+      maxv = boxp->volume;
+    }
+  }
+  return which;
+}
+
+
+LOCAL(void)
+update_box (j_decompress_ptr cinfo, boxptr boxp)
+/* Shrink the min/max bounds of a box to enclose only nonzero elements, */
+/* and recompute its volume and population */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  histptr histp;
+  int c0,c1,c2;
+  int c0min,c0max,c1min,c1max,c2min,c2max;
+  long dist0,dist1,dist2;
+  long ccount;
+  
+  c0min = boxp->c0min;  c0max = boxp->c0max;
+  c1min = boxp->c1min;  c1max = boxp->c1max;
+  c2min = boxp->c2min;  c2max = boxp->c2max;
+  
+  if (c0max > c0min)
+    for (c0 = c0min; c0 <= c0max; c0++)
+      for (c1 = c1min; c1 <= c1max; c1++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c0min = c0min = c0;
+	    goto have_c0min;
+	  }
+      }
+ have_c0min:
+  if (c0max > c0min)
+    for (c0 = c0max; c0 >= c0min; c0--)
+      for (c1 = c1min; c1 <= c1max; c1++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c0max = c0max = c0;
+	    goto have_c0max;
+	  }
+      }
+ have_c0max:
+  if (c1max > c1min)
+    for (c1 = c1min; c1 <= c1max; c1++)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c1min = c1min = c1;
+	    goto have_c1min;
+	  }
+      }
+ have_c1min:
+  if (c1max > c1min)
+    for (c1 = c1max; c1 >= c1min; c1--)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1][c2min];
+	for (c2 = c2min; c2 <= c2max; c2++)
+	  if (*histp++ != 0) {
+	    boxp->c1max = c1max = c1;
+	    goto have_c1max;
+	  }
+      }
+ have_c1max:
+  if (c2max > c2min)
+    for (c2 = c2min; c2 <= c2max; c2++)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1min][c2];
+	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
+	  if (*histp != 0) {
+	    boxp->c2min = c2min = c2;
+	    goto have_c2min;
+	  }
+      }
+ have_c2min:
+  if (c2max > c2min)
+    for (c2 = c2max; c2 >= c2min; c2--)
+      for (c0 = c0min; c0 <= c0max; c0++) {
+	histp = & histogram[c0][c1min][c2];
+	for (c1 = c1min; c1 <= c1max; c1++, histp += HIST_C2_ELEMS)
+	  if (*histp != 0) {
+	    boxp->c2max = c2max = c2;
+	    goto have_c2max;
+	  }
+      }
+ have_c2max:
+
+  /* Update box volume.
+   * We use 2-norm rather than real volume here; this biases the method
+   * against making long narrow boxes, and it has the side benefit that
+   * a box is splittable iff norm > 0.
+   * Since the differences are expressed in histogram-cell units,
+   * we have to shift back to JSAMPLE units to get consistent distances;
+   * after which, we scale according to the selected distance scale factors.
+   */
+  dist0 = ((c0max - c0min) << C0_SHIFT) * C0_SCALE;
+  dist1 = ((c1max - c1min) << C1_SHIFT) * C1_SCALE;
+  dist2 = ((c2max - c2min) << C2_SHIFT) * C2_SCALE;
+  boxp->volume = dist0*dist0 + dist1*dist1 + dist2*dist2;
+  
+  /* Now scan remaining volume of box and compute population */
+  ccount = 0;
+  for (c0 = c0min; c0 <= c0max; c0++)
+    for (c1 = c1min; c1 <= c1max; c1++) {
+      histp = & histogram[c0][c1][c2min];
+      for (c2 = c2min; c2 <= c2max; c2++, histp++)
+	if (*histp != 0) {
+	  ccount++;
+	}
+    }
+  boxp->colorcount = ccount;
+}
+
+
+LOCAL(int)
+median_cut (j_decompress_ptr cinfo, boxptr boxlist, int numboxes,
+	    int desired_colors)
+/* Repeatedly select and split the largest box until we have enough boxes */
+{
+  int n,lb;
+  int c0,c1,c2,cmax;
+  boxptr b1,b2;
+
+  while (numboxes < desired_colors) {
+    /* Select box to split.
+     * Current algorithm: by population for first half, then by volume.
+     */
+    if (numboxes*2 <= desired_colors) {
+      b1 = find_biggest_color_pop(boxlist, numboxes);
+    } else {
+      b1 = find_biggest_volume(boxlist, numboxes);
+    }
+    if (b1 == NULL)		/* no splittable boxes left! */
+      break;
+    b2 = &boxlist[numboxes];	/* where new box will go */
+    /* Copy the color bounds to the new box. */
+    b2->c0max = b1->c0max; b2->c1max = b1->c1max; b2->c2max = b1->c2max;
+    b2->c0min = b1->c0min; b2->c1min = b1->c1min; b2->c2min = b1->c2min;
+    /* Choose which axis to split the box on.
+     * Current algorithm: longest scaled axis.
+     * See notes in update_box about scaling distances.
+     */
+    c0 = ((b1->c0max - b1->c0min) << C0_SHIFT) * C0_SCALE;
+    c1 = ((b1->c1max - b1->c1min) << C1_SHIFT) * C1_SCALE;
+    c2 = ((b1->c2max - b1->c2min) << C2_SHIFT) * C2_SCALE;
+    /* We want to break any ties in favor of green, then red, blue last.
+     * This code does the right thing for R,G,B or B,G,R color orders only.
+     */
+#if RGB_RED == 0
+    cmax = c1; n = 1;
+    if (c0 > cmax) { cmax = c0; n = 0; }
+    if (c2 > cmax) { n = 2; }
+#else
+    cmax = c1; n = 1;
+    if (c2 > cmax) { cmax = c2; n = 2; }
+    if (c0 > cmax) { n = 0; }
+#endif
+    /* Choose split point along selected axis, and update box bounds.
+     * Current algorithm: split at halfway point.
+     * (Since the box has been shrunk to minimum volume,
+     * any split will produce two nonempty subboxes.)
+     * Note that lb value is max for lower box, so must be < old max.
+     */
+    switch (n) {
+    case 0:
+      lb = (b1->c0max + b1->c0min) / 2;
+      b1->c0max = lb;
+      b2->c0min = lb+1;
+      break;
+    case 1:
+      lb = (b1->c1max + b1->c1min) / 2;
+      b1->c1max = lb;
+      b2->c1min = lb+1;
+      break;
+    case 2:
+      lb = (b1->c2max + b1->c2min) / 2;
+      b1->c2max = lb;
+      b2->c2min = lb+1;
+      break;
+    }
+    /* Update stats for boxes */
+    update_box(cinfo, b1);
+    update_box(cinfo, b2);
+    numboxes++;
+  }
+  return numboxes;
+}
+
+
+LOCAL(void)
+compute_color (j_decompress_ptr cinfo, boxptr boxp, int icolor)
+/* Compute representative color for a box, put it in colormap[icolor] */
+{
+  /* Current algorithm: mean weighted by pixels (not colors) */
+  /* Note it is important to get the rounding correct! */
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  histptr histp;
+  int c0,c1,c2;
+  int c0min,c0max,c1min,c1max,c2min,c2max;
+  long count;
+  long total = 0;
+  long c0total = 0;
+  long c1total = 0;
+  long c2total = 0;
+  
+  c0min = boxp->c0min;  c0max = boxp->c0max;
+  c1min = boxp->c1min;  c1max = boxp->c1max;
+  c2min = boxp->c2min;  c2max = boxp->c2max;
+  
+  for (c0 = c0min; c0 <= c0max; c0++)
+    for (c1 = c1min; c1 <= c1max; c1++) {
+      histp = & histogram[c0][c1][c2min];
+      for (c2 = c2min; c2 <= c2max; c2++) {
+	if ((count = *histp++) != 0) {
+	  total += count;
+	  c0total += ((c0 << C0_SHIFT) + ((1<<C0_SHIFT)>>1)) * count;
+	  c1total += ((c1 << C1_SHIFT) + ((1<<C1_SHIFT)>>1)) * count;
+	  c2total += ((c2 << C2_SHIFT) + ((1<<C2_SHIFT)>>1)) * count;
+	}
+      }
+    }
+  
+  cinfo->colormap[0][icolor] = (JSAMPLE) ((c0total + (total>>1)) / total);
+  cinfo->colormap[1][icolor] = (JSAMPLE) ((c1total + (total>>1)) / total);
+  cinfo->colormap[2][icolor] = (JSAMPLE) ((c2total + (total>>1)) / total);
+}
+
+
+LOCAL(void)
+select_colors (j_decompress_ptr cinfo, int desired_colors)
+/* Master routine for color selection */
+{
+  boxptr boxlist;
+  int numboxes;
+  int i;
+
+  /* Allocate workspace for box list */
+  boxlist = (boxptr) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, desired_colors * SIZEOF(box));
+  /* Initialize one box containing whole space */
+  numboxes = 1;
+  boxlist[0].c0min = 0;
+  boxlist[0].c0max = MAXJSAMPLE >> C0_SHIFT;
+  boxlist[0].c1min = 0;
+  boxlist[0].c1max = MAXJSAMPLE >> C1_SHIFT;
+  boxlist[0].c2min = 0;
+  boxlist[0].c2max = MAXJSAMPLE >> C2_SHIFT;
+  /* Shrink it to actually-used volume and set its statistics */
+  update_box(cinfo, & boxlist[0]);
+  /* Perform median-cut to produce final box list */
+  numboxes = median_cut(cinfo, boxlist, numboxes, desired_colors);
+  /* Compute the representative color for each box, fill colormap */
+  for (i = 0; i < numboxes; i++)
+    compute_color(cinfo, & boxlist[i], i);
+  cinfo->actual_number_of_colors = numboxes;
+  TRACEMS1(cinfo, 1, JTRC_QUANT_SELECTED, numboxes);
+}
+
+
+/*
+ * These routines are concerned with the time-critical task of mapping input
+ * colors to the nearest color in the selected colormap.
+ *
+ * We re-use the histogram space as an "inverse color map", essentially a
+ * cache for the results of nearest-color searches.  All colors within a
+ * histogram cell will be mapped to the same colormap entry, namely the one
+ * closest to the cell's center.  This may not be quite the closest entry to
+ * the actual input color, but it's almost as good.  A zero in the cache
+ * indicates we haven't found the nearest color for that cell yet; the array
+ * is cleared to zeroes before starting the mapping pass.  When we find the
+ * nearest color for a cell, its colormap index plus one is recorded in the
+ * cache for future use.  The pass2 scanning routines call fill_inverse_cmap
+ * when they need to use an unfilled entry in the cache.
+ *
+ * Our method of efficiently finding nearest colors is based on the "locally
+ * sorted search" idea described by Heckbert and on the incremental distance
+ * calculation described by Spencer W. Thomas in chapter III.1 of Graphics
+ * Gems II (James Arvo, ed.  Academic Press, 1991).  Thomas points out that
+ * the distances from a given colormap entry to each cell of the histogram can
+ * be computed quickly using an incremental method: the differences between
+ * distances to adjacent cells themselves differ by a constant.  This allows a
+ * fairly fast implementation of the "brute force" approach of computing the
+ * distance from every colormap entry to every histogram cell.  Unfortunately,
+ * it needs a work array to hold the best-distance-so-far for each histogram
+ * cell (because the inner loop has to be over cells, not colormap entries).
+ * The work array elements have to be INT32s, so the work array would need
+ * 256Kb at our recommended precision.  This is not feasible in DOS machines.
+ *
+ * To get around these problems, we apply Thomas' method to compute the
+ * nearest colors for only the cells within a small subbox of the histogram.
+ * The work array need be only as big as the subbox, so the memory usage
+ * problem is solved.  Furthermore, we need not fill subboxes that are never
+ * referenced in pass2; many images use only part of the color gamut, so a
+ * fair amount of work is saved.  An additional advantage of this
+ * approach is that we can apply Heckbert's locality criterion to quickly
+ * eliminate colormap entries that are far away from the subbox; typically
+ * three-fourths of the colormap entries are rejected by Heckbert's criterion,
+ * and we need not compute their distances to individual cells in the subbox.
+ * The speed of this approach is heavily influenced by the subbox size: too
+ * small means too much overhead, too big loses because Heckbert's criterion
+ * can't eliminate as many colormap entries.  Empirically the best subbox
+ * size seems to be about 1/512th of the histogram (1/8th in each direction).
+ *
+ * Thomas' article also describes a refined method which is asymptotically
+ * faster than the brute-force method, but it is also far more complex and
+ * cannot efficiently be applied to small subboxes.  It is therefore not
+ * useful for programs intended to be portable to DOS machines.  On machines
+ * with plenty of memory, filling the whole histogram in one shot with Thomas'
+ * refined method might be faster than the present code --- but then again,
+ * it might not be any faster, and it's certainly more complicated.
+ */
+
+
+/* log2(histogram cells in update box) for each axis; this can be adjusted */
+#define BOX_C0_LOG  (HIST_C0_BITS-3)
+#define BOX_C1_LOG  (HIST_C1_BITS-3)
+#define BOX_C2_LOG  (HIST_C2_BITS-3)
+
+#define BOX_C0_ELEMS  (1<<BOX_C0_LOG) /* # of hist cells in update box */
+#define BOX_C1_ELEMS  (1<<BOX_C1_LOG)
+#define BOX_C2_ELEMS  (1<<BOX_C2_LOG)
+
+#define BOX_C0_SHIFT  (C0_SHIFT + BOX_C0_LOG)
+#define BOX_C1_SHIFT  (C1_SHIFT + BOX_C1_LOG)
+#define BOX_C2_SHIFT  (C2_SHIFT + BOX_C2_LOG)
+
+
+/*
+ * The next three routines implement inverse colormap filling.  They could
+ * all be folded into one big routine, but splitting them up this way saves
+ * some stack space (the mindist[] and bestdist[] arrays need not coexist)
+ * and may allow some compilers to produce better code by registerizing more
+ * inner-loop variables.
+ */
+
+LOCAL(int)
+find_nearby_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
+		    JSAMPLE colorlist[])
+/* Locate the colormap entries close enough to an update box to be candidates
+ * for the nearest entry to some cell(s) in the update box.  The update box
+ * is specified by the center coordinates of its first cell.  The number of
+ * candidate colormap entries is returned, and their colormap indexes are
+ * placed in colorlist[].
+ * This routine uses Heckbert's "locally sorted search" criterion to select
+ * the colors that need further consideration.
+ */
+{
+  int numcolors = cinfo->actual_number_of_colors;
+  int maxc0, maxc1, maxc2;
+  int centerc0, centerc1, centerc2;
+  int i, x, ncolors;
+  long minmaxdist, min_dist, max_dist, tdist;
+  long mindist[MAXNUMCOLORS];	/* min distance to colormap entry i */
+
+  /* Compute true coordinates of update box's upper corner and center.
+   * Actually we compute the coordinates of the center of the upper-corner
+   * histogram cell, which are the upper bounds of the volume we care about.
+   * Note that since ">>" rounds down, the "center" values may be closer to
+   * min than to max; hence comparisons to them must be "<=", not "<".
+   */
+  maxc0 = minc0 + ((1 << BOX_C0_SHIFT) - (1 << C0_SHIFT));
+  centerc0 = (minc0 + maxc0) >> 1;
+  maxc1 = minc1 + ((1 << BOX_C1_SHIFT) - (1 << C1_SHIFT));
+  centerc1 = (minc1 + maxc1) >> 1;
+  maxc2 = minc2 + ((1 << BOX_C2_SHIFT) - (1 << C2_SHIFT));
+  centerc2 = (minc2 + maxc2) >> 1;
+
+  /* For each color in colormap, find:
+   *  1. its minimum squared-distance to any point in the update box
+   *     (zero if color is within update box);
+   *  2. its maximum squared-distance to any point in the update box.
+   * Both of these can be found by considering only the corners of the box.
+   * We save the minimum distance for each color in mindist[];
+   * only the smallest maximum distance is of interest.
+   */
+  minmaxdist = 0x7FFFFFFFL;
+
+  for (i = 0; i < numcolors; i++) {
+    /* We compute the squared-c0-distance term, then add in the other two. */
+    x = GETJSAMPLE(cinfo->colormap[0][i]);
+    if (x < minc0) {
+      tdist = (x - minc0) * C0_SCALE;
+      min_dist = tdist*tdist;
+      tdist = (x - maxc0) * C0_SCALE;
+      max_dist = tdist*tdist;
+    } else if (x > maxc0) {
+      tdist = (x - maxc0) * C0_SCALE;
+      min_dist = tdist*tdist;
+      tdist = (x - minc0) * C0_SCALE;
+      max_dist = tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      min_dist = 0;
+      if (x <= centerc0) {
+	tdist = (x - maxc0) * C0_SCALE;
+	max_dist = tdist*tdist;
+      } else {
+	tdist = (x - minc0) * C0_SCALE;
+	max_dist = tdist*tdist;
+      }
+    }
+
+    x = GETJSAMPLE(cinfo->colormap[1][i]);
+    if (x < minc1) {
+      tdist = (x - minc1) * C1_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - maxc1) * C1_SCALE;
+      max_dist += tdist*tdist;
+    } else if (x > maxc1) {
+      tdist = (x - maxc1) * C1_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - minc1) * C1_SCALE;
+      max_dist += tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      if (x <= centerc1) {
+	tdist = (x - maxc1) * C1_SCALE;
+	max_dist += tdist*tdist;
+      } else {
+	tdist = (x - minc1) * C1_SCALE;
+	max_dist += tdist*tdist;
+      }
+    }
+
+    x = GETJSAMPLE(cinfo->colormap[2][i]);
+    if (x < minc2) {
+      tdist = (x - minc2) * C2_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - maxc2) * C2_SCALE;
+      max_dist += tdist*tdist;
+    } else if (x > maxc2) {
+      tdist = (x - maxc2) * C2_SCALE;
+      min_dist += tdist*tdist;
+      tdist = (x - minc2) * C2_SCALE;
+      max_dist += tdist*tdist;
+    } else {
+      /* within cell range so no contribution to min_dist */
+      if (x <= centerc2) {
+	tdist = (x - maxc2) * C2_SCALE;
+	max_dist += tdist*tdist;
+      } else {
+	tdist = (x - minc2) * C2_SCALE;
+	max_dist += tdist*tdist;
+      }
+    }
+
+    mindist[i] = min_dist;	/* save away the results */
+    if (max_dist < minmaxdist)
+      minmaxdist = max_dist;
+  }
+
+  /* Now we know that no cell in the update box is more than minmaxdist
+   * away from some colormap entry.  Therefore, only colors that are
+   * within minmaxdist of some part of the box need be considered.
+   */
+  ncolors = 0;
+  for (i = 0; i < numcolors; i++) {
+    if (mindist[i] <= minmaxdist)
+      colorlist[ncolors++] = (JSAMPLE) i;
+  }
+  return ncolors;
+}
+
+
+LOCAL(void)
+find_best_colors (j_decompress_ptr cinfo, int minc0, int minc1, int minc2,
+		  int numcolors, JSAMPLE colorlist[], JSAMPLE bestcolor[])
+/* Find the closest colormap entry for each cell in the update box,
+ * given the list of candidate colors prepared by find_nearby_colors.
+ * Return the indexes of the closest entries in the bestcolor[] array.
+ * This routine uses Thomas' incremental distance calculation method to
+ * find the distance from a colormap entry to successive cells in the box.
+ */
+{
+  int ic0, ic1, ic2;
+  int i, icolor;
+  long * bptr;	/* pointer into bestdist[] array */
+  JSAMPLE * cptr;		/* pointer into bestcolor[] array */
+  long dist0, dist1;		/* initial distance values */
+  long dist2;		/* current distance in inner loop */
+  long xx0, xx1;		/* distance increments */
+  long xx2;
+  long inc0, inc1, inc2;	/* initial values for increments */
+  /* This array holds the distance to the nearest-so-far color for each cell */
+  long bestdist[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
+
+  /* Initialize best-distance for each cell of the update box */
+  bptr = bestdist;
+  for (i = BOX_C0_ELEMS*BOX_C1_ELEMS*BOX_C2_ELEMS-1; i >= 0; i--)
+    *bptr++ = 0x7FFFFFFFL;
+  
+  /* For each color selected by find_nearby_colors,
+   * compute its distance to the center of each cell in the box.
+   * If that's less than best-so-far, update best distance and color number.
+   */
+  
+  /* Nominal steps between cell centers ("x" in Thomas article) */
+#define STEP_C0  ((1 << C0_SHIFT) * C0_SCALE)
+#define STEP_C1  ((1 << C1_SHIFT) * C1_SCALE)
+#define STEP_C2  ((1 << C2_SHIFT) * C2_SCALE)
+  
+  for (i = 0; i < numcolors; i++) {
+    icolor = GETJSAMPLE(colorlist[i]);
+    /* Compute (square of) distance from minc0/c1/c2 to this color */
+    inc0 = (minc0 - GETJSAMPLE(cinfo->colormap[0][icolor])) * C0_SCALE;
+    dist0 = inc0*inc0;
+    inc1 = (minc1 - GETJSAMPLE(cinfo->colormap[1][icolor])) * C1_SCALE;
+    dist0 += inc1*inc1;
+    inc2 = (minc2 - GETJSAMPLE(cinfo->colormap[2][icolor])) * C2_SCALE;
+    dist0 += inc2*inc2;
+    /* Form the initial difference increments */
+    inc0 = inc0 * (2 * STEP_C0) + STEP_C0 * STEP_C0;
+    inc1 = inc1 * (2 * STEP_C1) + STEP_C1 * STEP_C1;
+    inc2 = inc2 * (2 * STEP_C2) + STEP_C2 * STEP_C2;
+    /* Now loop over all cells in box, updating distance per Thomas method */
+    bptr = bestdist;
+    cptr = bestcolor;
+    xx0 = inc0;
+    for (ic0 = BOX_C0_ELEMS-1; ic0 >= 0; ic0--) {
+      dist1 = dist0;
+      xx1 = inc1;
+      for (ic1 = BOX_C1_ELEMS-1; ic1 >= 0; ic1--) {
+	dist2 = dist1;
+	xx2 = inc2;
+	for (ic2 = BOX_C2_ELEMS-1; ic2 >= 0; ic2--) {
+	  if (dist2 < *bptr) {
+	    *bptr = dist2;
+	    *cptr = (JSAMPLE) icolor;
+	  }
+	  dist2 += xx2;
+	  xx2 += 2 * STEP_C2 * STEP_C2;
+	  bptr++;
+	  cptr++;
+	}
+	dist1 += xx1;
+	xx1 += 2 * STEP_C1 * STEP_C1;
+      }
+      dist0 += xx0;
+      xx0 += 2 * STEP_C0 * STEP_C0;
+    }
+  }
+}
+
+
+LOCAL(void)
+fill_inverse_cmap (j_decompress_ptr cinfo, int c0, int c1, int c2)
+/* Fill the inverse-colormap entries in the update box that contains */
+/* histogram cell c0/c1/c2.  (Only that one cell MUST be filled, but */
+/* we can fill as many others as we wish.) */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  int minc0, minc1, minc2;	/* lower left corner of update box */
+  int ic0, ic1, ic2;
+  JSAMPLE * cptr;	/* pointer into bestcolor[] array */
+  histptr cachep;	/* pointer into main cache array */
+  /* This array lists the candidate colormap indexes. */
+  JSAMPLE colorlist[MAXNUMCOLORS];
+  int numcolors;		/* number of candidate colors */
+  /* This array holds the actually closest colormap index for each cell. */
+  JSAMPLE bestcolor[BOX_C0_ELEMS * BOX_C1_ELEMS * BOX_C2_ELEMS];
+
+  /* Convert cell coordinates to update box ID */
+  c0 >>= BOX_C0_LOG;
+  c1 >>= BOX_C1_LOG;
+  c2 >>= BOX_C2_LOG;
+
+  /* Compute true coordinates of update box's origin corner.
+   * Actually we compute the coordinates of the center of the corner
+   * histogram cell, which are the lower bounds of the volume we care about.
+   */
+  minc0 = (c0 << BOX_C0_SHIFT) + ((1 << C0_SHIFT) >> 1);
+  minc1 = (c1 << BOX_C1_SHIFT) + ((1 << C1_SHIFT) >> 1);
+  minc2 = (c2 << BOX_C2_SHIFT) + ((1 << C2_SHIFT) >> 1);
+  
+  /* Determine which colormap entries are close enough to be candidates
+   * for the nearest entry to some cell in the update box.
+   */
+  numcolors = find_nearby_colors(cinfo, minc0, minc1, minc2, colorlist);
+
+  /* Determine the actually nearest colors. */
+  find_best_colors(cinfo, minc0, minc1, minc2, numcolors, colorlist,
+		   bestcolor);
+
+  /* Save the best color numbers (plus 1) in the main cache array */
+  c0 <<= BOX_C0_LOG;		/* convert ID back to base cell indexes */
+  c1 <<= BOX_C1_LOG;
+  c2 <<= BOX_C2_LOG;
+  cptr = bestcolor;
+  for (ic0 = 0; ic0 < BOX_C0_ELEMS; ic0++) {
+    for (ic1 = 0; ic1 < BOX_C1_ELEMS; ic1++) {
+      cachep = & histogram[c0+ic0][c1+ic1][c2];
+      for (ic2 = 0; ic2 < BOX_C2_ELEMS; ic2++) {
+	*cachep++ = (histcell) (GETJSAMPLE(*cptr++) + 1);
+      }
+    }
+  }
+}
+
+
+/*
+ * Map some rows of pixels to the output colormapped representation.
+ */
+
+METHODDEF(void)
+pass2_no_dither (j_decompress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
+/* This version performs no dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  JSAMPROW inptr, outptr;
+  histptr cachep;
+  int c0, c1, c2;
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+
+  for (row = 0; row < num_rows; row++) {
+    inptr = input_buf[row];
+    outptr = output_buf[row];
+    for (col = width; col > 0; col--) {
+      /* get pixel value and index into the cache */
+      c0 = GETJSAMPLE(*inptr++) >> C0_SHIFT;
+      c1 = GETJSAMPLE(*inptr++) >> C1_SHIFT;
+      c2 = GETJSAMPLE(*inptr++) >> C2_SHIFT;
+      cachep = & histogram[c0][c1][c2];
+      /* If we have not seen this color before, find nearest colormap entry */
+      /* and update the cache */
+      if (*cachep == 0)
+	fill_inverse_cmap(cinfo, c0,c1,c2);
+      /* Now emit the colormap index for this cell */
+      *outptr++ = (JSAMPLE) (*cachep - 1);
+    }
+  }
+}
+
+
+METHODDEF(void)
+pass2_fs_dither (j_decompress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPARRAY output_buf, int num_rows)
+/* This version performs Floyd-Steinberg dithering */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  LOCFSERROR cur0, cur1, cur2;	/* current error or pixel value */
+  LOCFSERROR belowerr0, belowerr1, belowerr2; /* error for pixel below cur */
+  LOCFSERROR bpreverr0, bpreverr1, bpreverr2; /* error for below/prev col */
+  FSERRPTR errorptr;	/* => fserrors[] at column before current */
+  JSAMPROW inptr;		/* => current input pixel */
+  JSAMPROW outptr;		/* => current output pixel */
+  histptr cachep;
+  int dir;			/* +1 or -1 depending on direction */
+  int dir3;			/* 3*dir, for advancing inptr & errorptr */
+  int row;
+  JDIMENSION col;
+  JDIMENSION width = cinfo->output_width;
+  JSAMPLE *range_limit = cinfo->sample_range_limit;
+  int *error_limit = cquantize->error_limiter;
+  JSAMPROW colormap0 = cinfo->colormap[0];
+  JSAMPROW colormap1 = cinfo->colormap[1];
+  JSAMPROW colormap2 = cinfo->colormap[2];
+  SHIFT_TEMPS
+
+  for (row = 0; row < num_rows; row++) {
+    inptr = input_buf[row];
+    outptr = output_buf[row];
+    if (cquantize->on_odd_row) {
+      /* work right to left in this row */
+      inptr += (width-1) * 3;	/* so point to rightmost pixel */
+      outptr += width-1;
+      dir = -1;
+      dir3 = -3;
+      errorptr = cquantize->fserrors + (width+1)*3; /* => entry after last column */
+      cquantize->on_odd_row = FALSE; /* flip for next time */
+    } else {
+      /* work left to right in this row */
+      dir = 1;
+      dir3 = 3;
+      errorptr = cquantize->fserrors; /* => entry before first real column */
+      cquantize->on_odd_row = TRUE; /* flip for next time */
+    }
+    /* Preset error values: no error propagated to first pixel from left */
+    cur0 = cur1 = cur2 = 0;
+    /* and no error propagated to row below yet */
+    belowerr0 = belowerr1 = belowerr2 = 0;
+    bpreverr0 = bpreverr1 = bpreverr2 = 0;
+
+    for (col = width; col > 0; col--) {
+      /* curN holds the error propagated from the previous pixel on the
+       * current line.  Add the error propagated from the previous line
+       * to form the complete error correction term for this pixel, and
+       * round the error term (which is expressed * 16) to an integer.
+       * RIGHT_SHIFT rounds towards minus infinity, so adding 8 is correct
+       * for either sign of the error value.
+       * Note: errorptr points to *previous* column's array entry.
+       */
+      cur0 = RIGHT_SHIFT(cur0 + errorptr[dir3+0] + 8, 4);
+      cur1 = RIGHT_SHIFT(cur1 + errorptr[dir3+1] + 8, 4);
+      cur2 = RIGHT_SHIFT(cur2 + errorptr[dir3+2] + 8, 4);
+      /* Limit the error using transfer function set by init_error_limit.
+       * See comments with init_error_limit for rationale.
+       */
+      cur0 = error_limit[cur0];
+      cur1 = error_limit[cur1];
+      cur2 = error_limit[cur2];
+      /* Form pixel value + error, and range-limit to 0..MAXJSAMPLE.
+       * The maximum error is +- MAXJSAMPLE (or less with error limiting);
+       * this sets the required size of the range_limit array.
+       */
+      cur0 += GETJSAMPLE(inptr[0]);
+      cur1 += GETJSAMPLE(inptr[1]);
+      cur2 += GETJSAMPLE(inptr[2]);
+      cur0 = GETJSAMPLE(range_limit[cur0]);
+      cur1 = GETJSAMPLE(range_limit[cur1]);
+      cur2 = GETJSAMPLE(range_limit[cur2]);
+      /* Index into the cache with adjusted pixel value */
+      cachep = & histogram[cur0>>C0_SHIFT][cur1>>C1_SHIFT][cur2>>C2_SHIFT];
+      /* If we have not seen this color before, find nearest colormap */
+      /* entry and update the cache */
+      if (*cachep == 0)
+	fill_inverse_cmap(cinfo, cur0>>C0_SHIFT,cur1>>C1_SHIFT,cur2>>C2_SHIFT);
+      /* Now emit the colormap index for this cell */
+      { int pixcode = *cachep - 1;
+	*outptr = (JSAMPLE) pixcode;
+	/* Compute representation error for this pixel */
+	cur0 -= GETJSAMPLE(colormap0[pixcode]);
+	cur1 -= GETJSAMPLE(colormap1[pixcode]);
+	cur2 -= GETJSAMPLE(colormap2[pixcode]);
+      }
+      /* Compute error fractions to be propagated to adjacent pixels.
+       * Add these into the running sums, and simultaneously shift the
+       * next-line error sums left by 1 column.
+       */
+      { LOCFSERROR bnexterr, delta;
+
+	bnexterr = cur0;	/* Process component 0 */
+	delta = cur0 * 2;
+	cur0 += delta;		/* form error * 3 */
+	errorptr[0] = (FSERROR) (bpreverr0 + cur0);
+	cur0 += delta;		/* form error * 5 */
+	bpreverr0 = belowerr0 + cur0;
+	belowerr0 = bnexterr;
+	cur0 += delta;		/* form error * 7 */
+	bnexterr = cur1;	/* Process component 1 */
+	delta = cur1 * 2;
+	cur1 += delta;		/* form error * 3 */
+	errorptr[1] = (FSERROR) (bpreverr1 + cur1);
+	cur1 += delta;		/* form error * 5 */
+	bpreverr1 = belowerr1 + cur1;
+	belowerr1 = bnexterr;
+	cur1 += delta;		/* form error * 7 */
+	bnexterr = cur2;	/* Process component 2 */
+	delta = cur2 * 2;
+	cur2 += delta;		/* form error * 3 */
+	errorptr[2] = (FSERROR) (bpreverr2 + cur2);
+	cur2 += delta;		/* form error * 5 */
+	bpreverr2 = belowerr2 + cur2;
+	belowerr2 = bnexterr;
+	cur2 += delta;		/* form error * 7 */
+      }
+      /* At this point curN contains the 7/16 error value to be propagated
+       * to the next pixel on the current line, and all the errors for the
+       * next line have been shifted over.  We are therefore ready to move on.
+       */
+      inptr += dir3;		/* Advance pixel pointers to next column */
+      outptr += dir;
+      errorptr += dir3;		/* advance errorptr to current column */
+    }
+    /* Post-loop cleanup: we must unload the final error values into the
+     * final fserrors[] entry.  Note we need not unload belowerrN because
+     * it is for the dummy column before or after the actual array.
+     */
+    errorptr[0] = (FSERROR) bpreverr0; /* unload prev errs into array */
+    errorptr[1] = (FSERROR) bpreverr1;
+    errorptr[2] = (FSERROR) bpreverr2;
+  }
+}
+
+
+/*
+ * Initialize the error-limiting transfer function (lookup table).
+ * The raw F-S error computation can potentially compute error values of up to
+ * +- MAXJSAMPLE.  But we want the maximum correction applied to a pixel to be
+ * much less, otherwise obviously wrong pixels will be created.  (Typical
+ * effects include weird fringes at color-area boundaries, isolated bright
+ * pixels in a dark area, etc.)  The standard advice for avoiding this problem
+ * is to ensure that the "corners" of the color cube are allocated as output
+ * colors; then repeated errors in the same direction cannot cause cascading
+ * error buildup.  However, that only prevents the error from getting
+ * completely out of hand; Aaron Giles reports that error limiting improves
+ * the results even with corner colors allocated.
+ * A simple clamping of the error values to about +- MAXJSAMPLE/8 works pretty
+ * well, but the smoother transfer function used below is even better.  Thanks
+ * to Aaron Giles for this idea.
+ */
+
+LOCAL(void)
+init_error_limit (j_decompress_ptr cinfo)
+/* Allocate and fill in the error_limiter table */
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  int * table;
+  int in, out;
+
+  table = (int *) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, (MAXJSAMPLE*2+1) * SIZEOF(int));
+  table += MAXJSAMPLE;		/* so can index -MAXJSAMPLE .. +MAXJSAMPLE */
+  cquantize->error_limiter = table;
+
+#define STEPSIZE ((MAXJSAMPLE+1)/16)
+  /* Map errors 1:1 up to +- MAXJSAMPLE/16 */
+  out = 0;
+  for (in = 0; in < STEPSIZE; in++, out++) {
+    table[in] = out; table[-in] = -out;
+  }
+  /* Map errors 1:2 up to +- 3*MAXJSAMPLE/16 */
+  for (; in < STEPSIZE*3; in++, out += (in&1) ? 0 : 1) {
+    table[in] = out; table[-in] = -out;
+  }
+  /* Clamp the rest to final out value (which is (MAXJSAMPLE+1)/8) */
+  for (; in <= MAXJSAMPLE; in++) {
+    table[in] = out; table[-in] = -out;
+  }
+#undef STEPSIZE
+}
+
+
+/*
+ * Finish up at the end of each pass.
+ */
+
+METHODDEF(void)
+finish_pass1 (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+
+  /* Select the representative colors and fill in cinfo->colormap */
+  cinfo->colormap = cquantize->sv_colormap;
+  select_colors(cinfo, cquantize->desired);
+  /* Force next pass to zero the color index table */
+  cquantize->needs_zeroed = TRUE;
+}
+
+
+METHODDEF(void)
+finish_pass2 (j_decompress_ptr )
+{
+  /* no work */
+}
+
+
+/*
+ * Initialize for each processing pass.
+ */
+
+METHODDEF(void)
+start_pass_2_quant (j_decompress_ptr cinfo, int is_pre_scan)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+  hist3d histogram = cquantize->histogram;
+  int i;
+
+  /* Only F-S dithering or no dithering is supported. */
+  /* If user asks for ordered dither, give him F-S. */
+  if (cinfo->dither_mode != JDITHER_NONE)
+    cinfo->dither_mode = JDITHER_FS;
+
+  if (is_pre_scan) {
+    /* Set up method pointers */
+    cquantize->pub.color_quantize = prescan_quantize;
+    cquantize->pub.finish_pass = finish_pass1;
+    cquantize->needs_zeroed = TRUE; /* Always zero histogram */
+  } else {
+    /* Set up method pointers */
+    if (cinfo->dither_mode == JDITHER_FS)
+      cquantize->pub.color_quantize = pass2_fs_dither;
+    else
+      cquantize->pub.color_quantize = pass2_no_dither;
+    cquantize->pub.finish_pass = finish_pass2;
+
+    /* Make sure color count is acceptable */
+    i = cinfo->actual_number_of_colors;
+    if (i < 1)
+      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 1);
+    if (i > MAXNUMCOLORS)
+      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
+
+    if (cinfo->dither_mode == JDITHER_FS) {
+      size_t arraysize = (size_t) ((cinfo->output_width + 2) *
+				   (3 * SIZEOF(FSERROR)));
+      /* Allocate Floyd-Steinberg workspace if we didn't already. */
+      if (cquantize->fserrors == NULL)
+	cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, arraysize);
+      /* Initialize the propagated errors to zero. */
+      jzero_far((void FAR *) cquantize->fserrors, arraysize);
+      /* Make the error-limit table if we didn't already. */
+      if (cquantize->error_limiter == NULL)
+	init_error_limit(cinfo);
+      cquantize->on_odd_row = FALSE;
+    }
+
+  }
+  /* Zero the histogram or inverse color map, if necessary */
+  if (cquantize->needs_zeroed) {
+    for (i = 0; i < HIST_C0_ELEMS; i++) {
+      jzero_far((void FAR *) histogram[i],
+		HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
+    }
+    cquantize->needs_zeroed = FALSE;
+  }
+}
+
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+METHODDEF(void)
+new_color_map_2_quant (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize = (my_cquantize_ptr) cinfo->cquantize;
+
+  /* Reset the inverse color map */
+  cquantize->needs_zeroed = TRUE;
+}
+
+
+/*
+ * Module initialization routine for 2-pass color quantization.
+ */
+
+GLOBAL(void)
+jinit_2pass_quantizer (j_decompress_ptr cinfo)
+{
+  my_cquantize_ptr cquantize;
+  int i;
+
+  cquantize = (my_cquantize_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_cquantizer));
+  cinfo->cquantize = (struct jpeg_color_quantizer *) cquantize;
+  cquantize->pub.start_pass = start_pass_2_quant;
+  cquantize->pub.new_color_map = new_color_map_2_quant;
+  cquantize->fserrors = NULL;	/* flag optional arrays not allocated */
+  cquantize->error_limiter = NULL;
+
+  /* Make sure jdmaster didn't give me a case I can't handle */
+  if (cinfo->out_color_components != 3)
+    ERREXIT(cinfo, JERR_NOTIMPL);
+
+  /* Allocate the histogram/inverse colormap storage */
+  cquantize->histogram = (hist3d) (*cinfo->mem->alloc_small)
+    ((j_common_ptr) cinfo, JPOOL_IMAGE, HIST_C0_ELEMS * SIZEOF(hist2d));
+  for (i = 0; i < HIST_C0_ELEMS; i++) {
+    cquantize->histogram[i] = (hist2d) (*cinfo->mem->alloc_large)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       HIST_C1_ELEMS*HIST_C2_ELEMS * SIZEOF(histcell));
+  }
+  cquantize->needs_zeroed = TRUE; /* histogram is garbage now */
+
+  /* Allocate storage for the completed colormap, if required.
+   * We do this now since it is FAR storage and may affect
+   * the memory manager's space calculations.
+   */
+  if (cinfo->enable_2pass_quant) {
+    /* Make sure color count is acceptable */
+    int desired = cinfo->desired_number_of_colors;
+    /* Lower bound on # of colors ... somewhat arbitrary as long as > 0 */
+    if (desired < 8)
+      ERREXIT1(cinfo, JERR_QUANT_FEW_COLORS, 8);
+    /* Make sure colormap indexes can be represented by JSAMPLEs */
+    if (desired > MAXNUMCOLORS)
+      ERREXIT1(cinfo, JERR_QUANT_MANY_COLORS, MAXNUMCOLORS);
+    cquantize->sv_colormap = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo,JPOOL_IMAGE, (JDIMENSION) desired, (JDIMENSION) 3);
+    cquantize->desired = desired;
+  } else
+    cquantize->sv_colormap = NULL;
+
+  /* Only F-S dithering or no dithering is supported. */
+  /* If user asks for ordered dither, give him F-S. */
+  if (cinfo->dither_mode != JDITHER_NONE)
+    cinfo->dither_mode = JDITHER_FS;
+
+  /* Allocate Floyd-Steinberg workspace if necessary.
+   * This isn't really needed until pass 2, but again it is FAR storage.
+   * Although we will cope with a later change in dither_mode,
+   * we do not promise to honor max_memory_to_use if dither_mode changes.
+   */
+  if (cinfo->dither_mode == JDITHER_FS) {
+    cquantize->fserrors = (FSERRPTR) (*cinfo->mem->alloc_large)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (size_t) ((cinfo->output_width + 2) * (3 * SIZEOF(FSERROR))));
+    /* Might as well create the error-limiting table too. */
+    init_error_limit(cinfo);
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libjpeg/jutils.cpp b/ml/dlib/dlib/external/libjpeg/jutils.cpp
new file mode 100644
index 000000000..fd8906c83
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jutils.cpp
@@ -0,0 +1,179 @@
+/*
+ * jutils.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains tables and miscellaneous utility routines needed
+ * for both compression and decompression.
+ * Note we prefix all global names with "j" to minimize conflicts with
+ * a surrounding application.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * jpeg_zigzag_order[i] is the zigzag-order position of the i'th element
+ * of a DCT block read in natural order (left to right, top to bottom).
+ */
+
+#if 0				/* This table is not actually needed in v6a */
+
+const int jpeg_zigzag_order[DCTSIZE2] = {
+   0,  1,  5,  6, 14, 15, 27, 28,
+   2,  4,  7, 13, 16, 26, 29, 42,
+   3,  8, 12, 17, 25, 30, 41, 43,
+   9, 11, 18, 24, 31, 40, 44, 53,
+  10, 19, 23, 32, 39, 45, 52, 54,
+  20, 22, 33, 38, 46, 51, 55, 60,
+  21, 34, 37, 47, 50, 56, 59, 61,
+  35, 36, 48, 49, 57, 58, 62, 63
+};
+
+#endif
+
+/*
+ * jpeg_natural_order[i] is the natural-order position of the i'th element
+ * of zigzag order.
+ *
+ * When reading corrupted data, the Huffman decoders could attempt
+ * to reference an entry beyond the end of this array (if the decoded
+ * zero run length reaches past the end of the block).  To prevent
+ * wild stores without adding an inner-loop test, we put some extra
+ * "63"s after the real entries.  This will cause the extra coefficient
+ * to be stored in location 63 of the block, not somewhere random.
+ * The worst case would be a run-length of 15, which means we need 16
+ * fake entries.
+ */
+
+const int jpeg_natural_order[DCTSIZE2+16] = {
+  0,  1,  8, 16,  9,  2,  3, 10,
+ 17, 24, 32, 25, 18, 11,  4,  5,
+ 12, 19, 26, 33, 40, 48, 41, 34,
+ 27, 20, 13,  6,  7, 14, 21, 28,
+ 35, 42, 49, 56, 57, 50, 43, 36,
+ 29, 22, 15, 23, 30, 37, 44, 51,
+ 58, 59, 52, 45, 38, 31, 39, 46,
+ 53, 60, 61, 54, 47, 55, 62, 63,
+ 63, 63, 63, 63, 63, 63, 63, 63, /* extra entries for safety in decoder */
+ 63, 63, 63, 63, 63, 63, 63, 63
+};
+
+
+/*
+ * Arithmetic utilities
+ */
+
+GLOBAL(long)
+jdiv_round_up (long a, long b)
+/* Compute a/b rounded up to next integer, ie, ceil(a/b) */
+/* Assumes a >= 0, b > 0 */
+{
+  return (a + b - 1L) / b;
+}
+
+
+GLOBAL(long)
+jround_up (long a, long b)
+/* Compute a rounded up to next multiple of b, ie, ceil(a/b)*b */
+/* Assumes a >= 0, b > 0 */
+{
+  a += b - 1L;
+  return a - (a % b);
+}
+
+
+/* On normal machines we can apply MEMCOPY() and MEMZERO() to sample arrays
+ * and coefficient-block arrays.  This won't work on 80x86 because the arrays
+ * are FAR and we're assuming a small-pointer memory model.  However, some
+ * DOS compilers provide far-pointer versions of memcpy() and memset() even
+ * in the small-model libraries.  These will be used if USE_FMEM is defined.
+ * Otherwise, the routines below do it the hard way.  (The performance cost
+ * is not all that great, because these routines aren't very heavily used.)
+ */
+
+#ifndef NEED_FAR_POINTERS	/* normal case, same as regular macros */
+#define FMEMCOPY(dest,src,size)	MEMCOPY(dest,src,size)
+#define FMEMZERO(target,size)	MEMZERO(target,size)
+#else				/* 80x86 case, define if we can */
+#ifdef USE_FMEM
+#define FMEMCOPY(dest,src,size)	_fmemcpy((void FAR *)(dest), (const void FAR *)(src), (size_t)(size))
+#define FMEMZERO(target,size)	_fmemset((void FAR *)(target), 0, (size_t)(size))
+#endif
+#endif
+
+
+GLOBAL(void)
+jcopy_sample_rows (JSAMPARRAY input_array, int source_row,
+		   JSAMPARRAY output_array, int dest_row,
+		   int num_rows, JDIMENSION num_cols)
+/* Copy some rows of samples from one place to another.
+ * num_rows rows are copied from input_array[source_row++]
+ * to output_array[dest_row++]; these areas may overlap for duplication.
+ * The source and destination arrays must be at least as wide as num_cols.
+ */
+{
+  JSAMPROW inptr, outptr;
+#ifdef FMEMCOPY
+  size_t count = (size_t) (num_cols * SIZEOF(JSAMPLE));
+#else
+  JDIMENSION count;
+#endif
+  int row;
+
+  input_array += source_row;
+  output_array += dest_row;
+
+  for (row = num_rows; row > 0; row--) {
+    inptr = *input_array++;
+    outptr = *output_array++;
+#ifdef FMEMCOPY
+    FMEMCOPY(outptr, inptr, count);
+#else
+    for (count = num_cols; count > 0; count--)
+      *outptr++ = *inptr++;	/* needn't bother with GETJSAMPLE() here */
+#endif
+  }
+}
+
+
+GLOBAL(void)
+jcopy_block_row (JBLOCKROW input_row, JBLOCKROW output_row,
+		 JDIMENSION num_blocks)
+/* Copy a row of coefficient blocks from one place to another. */
+{
+#ifdef FMEMCOPY
+  FMEMCOPY(output_row, input_row, num_blocks * (DCTSIZE2 * SIZEOF(JCOEF)));
+#else
+  JCOEFPTR inptr, outptr;
+  long count;
+
+  inptr = (JCOEFPTR) input_row;
+  outptr = (JCOEFPTR) output_row;
+  for (count = (long) num_blocks * DCTSIZE2; count > 0; count--) {
+    *outptr++ = *inptr++;
+  }
+#endif
+}
+
+
+GLOBAL(void)
+jzero_far (void FAR * target, size_t bytestozero)
+/* Zero out a chunk of FAR memory. */
+/* This might be sample-array data, block-array data, or alloc_large data. */
+{
+#ifdef FMEMZERO
+  FMEMZERO(target, bytestozero);
+#else
+  char FAR * ptr = (char FAR *) target;
+  size_t count;
+
+  for (count = bytestozero; count > 0; count--) {
+    *ptr++ = 0;
+  }
+#endif
+}
diff --git a/ml/dlib/dlib/external/libjpeg/jversion.h b/ml/dlib/dlib/external/libjpeg/jversion.h
new file mode 100644
index 000000000..6472c58d3
--- /dev/null
+++ b/ml/dlib/dlib/external/libjpeg/jversion.h
@@ -0,0 +1,14 @@
+/*
+ * jversion.h
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains software version identification.
+ */
+
+
+#define JVERSION	"6b  27-Mar-1998"
+
+#define JCOPYRIGHT	"Copyright (C) 1998, Thomas G. Lane"
diff --git a/ml/dlib/dlib/external/libpng/LICENSE b/ml/dlib/dlib/external/libpng/LICENSE
new file mode 100644
index 000000000..b1b97ea57
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/LICENSE
@@ -0,0 +1,111 @@
+
+This copy of the libpng notices is provided for your convenience.  In case of
+any discrepancy between this copy and the notices in the file png.h that is
+included in the libpng distribution, the latter shall prevail.
+
+COPYRIGHT NOTICE, DISCLAIMER, and LICENSE:
+
+If you modify libpng you may insert additional notices immediately following
+this sentence.
+
+This code is released under the libpng license.
+
+libpng versions 1.2.6, August 15, 2004, through 1.6.7, November 14, 2013, are
+Copyright (c) 2004, 2006-2013 Glenn Randers-Pehrson, and are
+distributed according to the same disclaimer and license as libpng-1.2.5
+with the following individual added to the list of Contributing Authors
+
+   Cosmin Truta
+
+libpng versions 1.0.7, July 1, 2000, through 1.2.5 - October 3, 2002, are
+Copyright (c) 2000-2002 Glenn Randers-Pehrson, and are
+distributed according to the same disclaimer and license as libpng-1.0.6
+with the following individuals added to the list of Contributing Authors
+
+   Simon-Pierre Cadieux
+   Eric S. Raymond
+   Gilles Vollant
+
+and with the following additions to the disclaimer:
+
+   There is no warranty against interference with your enjoyment of the
+   library or against infringement.  There is no warranty that our
+   efforts or the library will fulfill any of your particular purposes
+   or needs.  This library is provided with all faults, and the entire
+   risk of satisfactory quality, performance, accuracy, and effort is with
+   the user.
+
+libpng versions 0.97, January 1998, through 1.0.6, March 20, 2000, are
+Copyright (c) 1998, 1999 Glenn Randers-Pehrson, and are
+distributed according to the same disclaimer and license as libpng-0.96,
+with the following individuals added to the list of Contributing Authors:
+
+   Tom Lane
+   Glenn Randers-Pehrson
+   Willem van Schaik
+
+libpng versions 0.89, June 1996, through 0.96, May 1997, are
+Copyright (c) 1996, 1997 Andreas Dilger
+Distributed according to the same disclaimer and license as libpng-0.88,
+with the following individuals added to the list of Contributing Authors:
+
+   John Bowler
+   Kevin Bracey
+   Sam Bushell
+   Magnus Holmgren
+   Greg Roelofs
+   Tom Tanner
+
+libpng versions 0.5, May 1995, through 0.88, January 1996, are
+Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.
+
+For the purposes of this copyright and license, "Contributing Authors"
+is defined as the following set of individuals:
+
+   Andreas Dilger
+   Dave Martindale
+   Guy Eric Schalnat
+   Paul Schmidt
+   Tim Wegner
+
+The PNG Reference Library is supplied "AS IS".  The Contributing Authors
+and Group 42, Inc. disclaim all warranties, expressed or implied,
+including, without limitation, the warranties of merchantability and of
+fitness for any purpose.  The Contributing Authors and Group 42, Inc.
+assume no liability for direct, indirect, incidental, special, exemplary,
+or consequential damages, which may result from the use of the PNG
+Reference Library, even if advised of the possibility of such damage.
+
+Permission is hereby granted to use, copy, modify, and distribute this
+source code, or portions hereof, for any purpose, without fee, subject
+to the following restrictions:
+
+1. The origin of this source code must not be misrepresented.
+
+2. Altered versions must be plainly marked as such and must not
+   be misrepresented as being the original source.
+
+3. This Copyright notice may not be removed or altered from any
+   source or altered source distribution.
+
+The Contributing Authors and Group 42, Inc. specifically permit, without
+fee, and encourage the use of this source code as a component to
+supporting the PNG file format in commercial products.  If you use this
+source code in a product, acknowledgment is not required but would be
+appreciated.
+
+
+A "png_get_copyright" function is available, for convenient use in "about"
+boxes and the like:
+
+   printf("%s",png_get_copyright(NULL));
+
+Also, the PNG logo (in PNG format, of course) is supplied in the
+files "pngbar.png" and "pngbar.jpg (88x31) and "pngnow.png" (98x31).
+
+Libpng is OSI Certified Open Source Software.  OSI Certified Open Source is a
+certification mark of the Open Source Initiative.
+
+Glenn Randers-Pehrson
+glennrp at users.sourceforge.net
+November 14, 2013
diff --git a/ml/dlib/dlib/external/libpng/README b/ml/dlib/dlib/external/libpng/README
new file mode 100644
index 000000000..80fc574ad
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/README
@@ -0,0 +1,202 @@
+README for libpng version 1.6.7 - November 14, 2013 (shared library 16.0)
+See the note about version numbers near the top of png.h
+
+See INSTALL for instructions on how to install libpng.
+
+Libpng comes in several distribution formats.  Get libpng-*.tar.gz or
+libpng-*.tar.xz or if you want UNIX-style line endings in the text files,
+or lpng*.7z or lpng*.zip if you want DOS-style line endings.
+
+Version 0.89 was the first official release of libpng.  Don't let the
+fact that it's the first release fool you.  The libpng library has been in
+extensive use and testing since mid-1995.  By late 1997 it had
+finally gotten to the stage where there hadn't been significant
+changes to the API in some time, and people have a bad feeling about
+libraries with versions < 1.0.  Version 1.0.0 was released in
+March 1998.
+
+****
+Note that some of the changes to the png_info structure render this
+version of the library binary incompatible with libpng-0.89 or
+earlier versions if you are using a shared library.  The type of the
+"filler" parameter for png_set_filler() has changed from png_byte to
+png_uint_32, which will affect shared-library applications that use
+this function.
+
+To avoid problems with changes to the internals of png_info_struct,
+new APIs have been made available in 0.95 to avoid direct application
+access to info_ptr.  These functions are the png_set_<chunk> and
+png_get_<chunk> functions.  These functions should be used when
+accessing/storing the info_struct data, rather than manipulating it
+directly, to avoid such problems in the future.
+
+It is important to note that the APIs do not make current programs
+that access the info struct directly incompatible with the new
+library.  However, it is strongly suggested that new programs use
+the new APIs (as shown in example.c and pngtest.c), and older programs
+be converted to the new format, to facilitate upgrades in the future.
+****
+
+Additions since 0.90 include the ability to compile libpng as a
+Windows DLL, and new APIs for accessing data in the info struct.
+Experimental functions include the ability to set weighting and cost
+factors for row filter selection, direct reads of integers from buffers
+on big-endian processors that support misaligned data access, faster
+methods of doing alpha composition, and more accurate 16->8 bit color
+conversion.
+
+The additions since 0.89 include the ability to read from a PNG stream
+which has had some (or all) of the signature bytes read by the calling
+application.  This also allows the reading of embedded PNG streams that
+do not have the PNG file signature.  As well, it is now possible to set
+the library action on the detection of chunk CRC errors.  It is possible
+to set different actions based on whether the CRC error occurred in a
+critical or an ancillary chunk.
+
+The changes made to the library, and bugs fixed are based on discussions
+on the PNG-implement mailing list and not on material submitted
+privately to Guy, Andreas, or Glenn.  They will forward any good
+suggestions to the list.
+
+For a detailed description on using libpng, read libpng-manual.txt.  For
+examples of libpng in a program, see example.c and pngtest.c.  For usage
+information and restrictions (what little they are) on libpng, see
+png.h.  For a description on using zlib (the compression library used by
+libpng) and zlib's restrictions, see zlib.h
+
+I have included a general makefile, as well as several machine and
+compiler specific ones, but you may have to modify one for your own needs.
+
+You should use zlib 1.0.4 or later to run this, but it MAY work with
+versions as old as zlib 0.95.  Even so, there are bugs in older zlib
+versions which can cause the output of invalid compression streams for
+some images.  You will definitely need zlib 1.0.4 or later if you are
+taking advantage of the MS-DOS "far" structure allocation for the small
+and medium memory models.  You should also note that zlib is a
+compression library that is useful for more things than just PNG files.
+You can use zlib as a drop-in replacement for fread() and fwrite() if
+you are so inclined.
+
+zlib should be available at the same place that libpng is, or at zlib.net.
+
+You may also want a copy of the PNG specification.  It is available
+as an RFC, a W3C Recommendation, and an ISO/IEC Standard.  You can find
+these at http://www.libpng.org/pub/png/documents/
+
+This code is currently being archived at libpng.sf.net in the
+[DOWNLOAD] area, and at ftp://ftp.simplesystems.org.  If you can't find it
+in any of those places, e-mail me, and I'll help you find it.
+
+If you have any code changes, requests, problems, etc., please e-mail
+them to me.  Also, I'd appreciate any make files or project files,
+and any modifications you needed to make to get libpng to compile,
+along with a #define variable to tell what compiler/system you are on.
+If you needed to add transformations to libpng, or wish libpng would
+provide the image in a different way, drop me a note (and code, if
+possible), so I can consider supporting the transformation.
+Finally, if you get any warning messages when compiling libpng
+(note: not zlib), and they are easy to fix, I'd appreciate the
+fix.  Please mention "libpng" somewhere in the subject line.  Thanks.
+
+This release was created and will be supported by myself (of course
+based in a large way on Guy's and Andreas' earlier work), and the PNG
+development group.
+
+Send comments/corrections/commendations to png-mng-implement at
+lists.sourceforge.net (subscription required; visit
+https://lists.sourceforge.net/lists/listinfo/png-mng-implement
+to subscribe) or to glennrp at users.sourceforge.net
+
+You can't reach Guy, the original libpng author, at the addresses
+given in previous versions of this document.  He and Andreas will
+read mail addressed to the png-implement list, however.
+
+Please do not send general questions about PNG.  Send them to
+png-mng-misc at lists.sf.net (subscription required; visit
+https://lists.sourceforge.net/lists/listinfo/png-mng-misc to
+subscribe).  If you have a question about something
+in the PNG specification that is related to using libpng, send it
+to me.  Send me any questions that start with "I was using libpng,
+and ...".  If in doubt, send questions to me.  I'll bounce them
+to others, if necessary.
+
+Please do not send suggestions on how to change PNG.  We have
+been discussing PNG for eighteen years now, and it is official and
+finished.  If you have suggestions for libpng, however, I'll
+gladly listen.  Even if your suggestion is not used immediately,
+it may be used later.
+
+Files in this distribution:
+
+      ANNOUNCE      =>  Announcement of this version, with recent changes
+      CHANGES       =>  Description of changes between libpng versions
+      KNOWNBUG      =>  List of known bugs and deficiencies
+      LICENSE       =>  License to use and redistribute libpng
+      README        =>  This file
+      TODO          =>  Things not implemented in the current library
+      Y2KINFO       =>  Statement of Y2K compliance
+      example.c     =>  Example code for using libpng functions
+      libpng.3      =>  manual page for libpng (includes libpng-manual.txt)
+      libpng-manual.txt  =>  Description of libpng and its functions
+      libpngpf.3    =>  manual page for libpng's private functions
+      png.5         =>  manual page for the PNG format
+      png.c         =>  Basic interface functions common to library
+      png.h         =>  Library function and interface declarations (public)
+      pngpriv.h     =>  Library function and interface declarations (private)
+      pngconf.h     =>  System specific library configuration (public)
+      pngstruct.h   =>  png_struct declaration (private)
+      pnginfo.h     =>  png_info struct declaration (private)
+      pngdebug.h    =>  debugging macros (private)
+      pngerror.c    =>  Error/warning message I/O functions
+      pngget.c      =>  Functions for retrieving info from struct
+      pngmem.c      =>  Memory handling functions
+      pngbar.png    =>  PNG logo, 88x31
+      pngnow.png    =>  PNG logo, 98x31
+      pngpread.c    =>  Progressive reading functions
+      pngread.c     =>  Read data/helper high-level functions
+      pngrio.c      =>  Lowest-level data read I/O functions
+      pngrtran.c    =>  Read data transformation functions
+      pngrutil.c    =>  Read data utility functions
+      pngset.c      =>  Functions for storing data into the info_struct
+      pngtest.c     =>  Library test program
+      pngtest.png   =>  Library test sample image
+      pngtrans.c    =>  Common data transformation functions
+      pngwio.c      =>  Lowest-level write I/O functions
+      pngwrite.c    =>  High-level write functions
+      pngwtran.c    =>  Write data transformations
+      pngwutil.c    =>  Write utility functions
+      arm           =>  Contains optimized code for the ARM platform
+      contrib       =>  Contributions
+       examples         =>  Example programs
+       gregbook         =>  source code for PNG reading and writing, from
+                            Greg Roelofs' "PNG: The Definitive Guide",
+                            O'Reilly, 1999
+       libtests         =>  Test programs
+       pngminim         =>  Minimal decoder, encoder, and progressive decoder
+                            programs demonstrating use of pngusr.dfa
+       pngminus         =>  Simple pnm2png and png2pnm programs
+       pngsuite         =>  Test images
+       tools            =>  Various tools
+       visupng      =>  Contains a MSVC workspace for VisualPng
+      projects      =>  Contains project files and workspaces for
+                        building a DLL
+       owatcom          =>  Contains a WATCOM project for building libpng
+       visualc71        =>  Contains a Microsoft Visual C++ (MSVC)
+                            workspace for building libpng and zlib
+       vstudio          =>  Contains a Microsoft Visual C++ (MSVC)
+                            workspace for building libpng and zlib
+      scripts       =>  Directory containing scripts for building libpng:
+                            (see scripts/README.txt for the list of scripts)
+
+Good luck, and happy coding.
+
+-Glenn Randers-Pehrson (current maintainer, since 1998)
+ Internet: glennrp at users.sourceforge.net
+
+-Andreas Eric Dilger (former maintainer, 1996-1997)
+ Internet: adilger at enel.ucalgary.ca
+ Web: http://www-mddsp.enel.ucalgary.ca/People/adilger/
+
+-Guy Eric Schalnat (original author and former maintainer, 1995-1996)
+ (formerly of Group 42, Inc)
+ Internet: gschal at infinet.com
diff --git a/ml/dlib/dlib/external/libpng/arm/arm_init.c b/ml/dlib/dlib/external/libpng/arm/arm_init.c
new file mode 100644
index 000000000..098771781
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/arm/arm_init.c
@@ -0,0 +1,232 @@
+
+/* arm_init.c - NEON optimised filter functions
+ *
+ * Copyright (c) 2013 Glenn Randers-Pehrson
+ * Written by Mans Rullgard, 2011.
+ * Last changed in libpng 1.6.6 [September 16, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+/* Below, after checking __linux__, various non-C90 POSIX 1003.1 functions are
+ * called.
+ */
+#define _POSIX_SOURCE 1
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_ARM_NEON_OPT > 0
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */
+#include <signal.h> /* for sig_atomic_t */
+
+#ifdef __ANDROID__
+/* Linux provides access to information about CPU capabilites via
+ * /proc/self/auxv, however Android blocks this while still claiming to be
+ * Linux.  The Andoid NDK, however, provides appropriate support.
+ *
+ * Documentation: http://www.kandroid.org/ndk/docs/CPU-ARM-NEON.html
+ */
+#include <cpu-features.h>
+
+static int
+png_have_neon(png_structp png_ptr)
+{
+   /* This is a whole lot easier than the mess below, however it is probably
+    * implemented as below, therefore it is better to cache the result (these
+    * function calls may be slow!)
+    */
+   PNG_UNUSED(png_ptr)
+   return android_getCpuFamily() == ANDROID_CPU_FAMILY_ARM &&
+      (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
+}
+#elif defined(__linux__)
+/* The generic __linux__ implementation requires reading /proc/self/auxv and
+ * looking at each element for one that records NEON capabilities.
+ */
+#include <unistd.h> /* for POSIX 1003.1 */
+#include <errno.h>  /* for EINTR */
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <elf.h>
+#include <asm/hwcap.h>
+
+/* A read call may be interrupted, in which case it returns -1 and sets errno to
+ * EINTR if nothing was done, otherwise (if something was done) a partial read
+ * may result.
+ */
+static size_t
+safe_read(png_structp png_ptr, int fd, void *buffer_in, size_t nbytes)
+{
+   size_t ntotal = 0;
+   char *buffer = png_voidcast(char*, buffer_in);
+
+   while (nbytes > 0)
+   {
+      unsigned int nread;
+      int iread;
+
+      /* Passing nread > INT_MAX to read is implementation defined in POSIX
+       * 1003.1, therefore despite the unsigned argument portable code must
+       * limit the value to INT_MAX!
+       */
+      if (nbytes > INT_MAX)
+         nread = INT_MAX;
+
+      else
+         nread = (unsigned int)/*SAFE*/nbytes;
+
+      iread = read(fd, buffer, nread);
+
+      if (iread == -1)
+      {
+         /* This is the devil in the details, a read can terminate early with 0
+          * bytes read because of EINTR, yet it still returns -1 otherwise end
+          * of file cannot be distinguished.
+          */
+         if (errno != EINTR)
+         {
+            png_warning(png_ptr, "/proc read failed");
+            return 0; /* I.e. a permanent failure */
+         }
+      }
+
+      else if (iread < 0)
+      {
+         /* Not a valid 'read' result: */
+         png_warning(png_ptr, "OS /proc read bug");
+         return 0;
+      }
+
+      else if (iread > 0)
+      {
+         /* Continue reading until a permanent failure, or EOF */
+         buffer += iread;
+         nbytes -= (unsigned int)/*SAFE*/iread;
+         ntotal += (unsigned int)/*SAFE*/iread;
+      }
+
+      else
+         return ntotal;
+   }
+
+   return ntotal; /* nbytes == 0 */
+}
+
+static int
+png_have_neon(png_structp png_ptr)
+{
+   int fd = open("/proc/self/auxv", O_RDONLY);
+   Elf32_auxv_t aux;
+
+   /* Failsafe: failure to open means no NEON */
+   if (fd == -1)
+   {
+      png_warning(png_ptr, "/proc/self/auxv open failed");
+      return 0;
+   }
+
+   while (safe_read(png_ptr, fd, &aux, sizeof aux) == sizeof aux)
+   {
+      if (aux.a_type == AT_HWCAP && (aux.a_un.a_val & HWCAP_NEON) != 0)
+      {
+         close(fd);
+         return 1;
+      }
+   }
+
+   close(fd);
+   return 0;
+}
+#else
+   /* We don't know how to do a run-time check on this system */
+#  error "no support for run-time ARM NEON checks"
+#endif /* OS checks */
+#endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
+
+#ifndef PNG_ALIGNED_MEMORY_SUPPORTED
+#  error "ALIGNED_MEMORY is required; set: -DPNG_ALIGNED_MEMORY_SUPPORTED"
+#endif
+
+void
+png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
+{
+   /* The switch statement is compiled in for ARM_NEON_API, the call to
+    * png_have_neon is compiled in for ARM_NEON_CHECK.  If both are defined
+    * the check is only performed if the API has not set the NEON option on
+    * or off explicitly.  In this case the check controls what happens.
+    *
+    * If the CHECK is not compiled in and the option is UNSET the behavior prior
+    * to 1.6.7 was to use the NEON code - this was a bug caused by having the
+    * wrong order of the 'ON' and 'default' cases.  UNSET now defaults to OFF,
+    * as documented in png.h
+    */
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+   switch ((pp->options >> PNG_ARM_NEON) & 3)
+   {
+      case PNG_OPTION_UNSET:
+         /* Allow the run-time check to execute if it has been enabled -
+          * thus both API and CHECK can be turned on.  If it isn't supported
+          * this case will fall through to the 'default' below, which just
+          * returns.
+          */
+#endif /* PNG_ARM_NEON_API_SUPPORTED */
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED
+         {
+            static volatile sig_atomic_t no_neon = -1; /* not checked */
+
+            if (no_neon < 0)
+               no_neon = !png_have_neon(pp);
+
+            if (no_neon)
+               return;
+         }
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+         break;
+#endif
+#endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
+
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+      default: /* OFF or INVALID */
+         return;
+
+      case PNG_OPTION_ON:
+         /* Option turned on */
+         break;
+   }
+#endif
+
+   /* IMPORTANT: any new external functions used here must be declared using
+    * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
+    * 'prefix' option to configure works:
+    *
+    *    ./configure --with-libpng-prefix=foobar_
+    *
+    * Verify you have got this right by running the above command, doing a build
+    * and examining pngprefix.h; it must contain a #define for every external
+    * function you add.  (Notice that this happens automatically for the
+    * initialization function.)
+    */
+   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
+
+   if (bpp == 3)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth3_neon;
+   }
+
+   else if (bpp == 4)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+          png_read_filter_row_paeth4_neon;
+   }
+}
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* PNG_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/arm/filter_neon.S b/ml/dlib/dlib/external/libpng/arm/filter_neon.S
new file mode 100644
index 000000000..3d1ccf505
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/arm/filter_neon.S
@@ -0,0 +1,245 @@
+
+/* filter_neon.S - NEON optimised filter functions
+ *
+ * Copyright (c) 2013 Glenn Randers-Pehrson
+ * Written by Mans Rullgard, 2011.
+ * Last changed in libpng 1.6.7 [November 14, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* This is required to get the symbol renames, which are #defines, and also
+ * includes the definition (or not) of PNG_ARM_NEON_OPT.
+ */
+#define PNG_VERSION_INFO_ONLY
+#include "../pngpriv.h"
+
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
+#endif
+
+/* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for
+ * ARM64).  The code in arm/filter_neon_intrinsics.c supports ARM64, however it
+ * only works if -mfpu=neon is specified on the GCC command line.  See pngpriv.h
+ * for the logic which sets PNG_USE_ARM_NEON_ASM:
+ */
+#if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_ARM_NEON_OPT > 0
+
+#ifdef __ELF__
+#   define ELF
+#else
+#   define ELF @
+#endif
+
+        .arch armv7-a
+        .fpu  neon
+
+.macro  func    name, export=0
+    .macro endfunc
+ELF     .size   \name, . - \name
+        .endfunc
+        .purgem endfunc
+    .endm
+        .text
+    .if \export
+        .global \name
+    .endif
+ELF     .type   \name, STT_FUNC
+        .func   \name
+\name:
+.endm
+
+func    png_read_filter_row_sub4_neon, export=1
+        ldr             r3,  [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+1:
+        vld4.32         {d4[],d5[],d6[],d7[]},    [r1,:128]
+        vadd.u8         d0,  d3,  d4
+        vadd.u8         d1,  d0,  d5
+        vadd.u8         d2,  d1,  d6
+        vadd.u8         d3,  d2,  d7
+        vst4.32         {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
+        subs            r3,  r3,  #16
+        bgt             1b
+
+        bx              lr
+endfunc
+
+func    png_read_filter_row_sub3_neon, export=1
+        ldr             r3,  [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+        mov             r0,  r1
+        mov             r2,  #3
+        mov             r12, #12
+        vld1.8          {q11},    [r0], r12
+1:
+        vext.8          d5,  d22, d23, #3
+        vadd.u8         d0,  d3,  d22
+        vext.8          d6,  d22, d23, #6
+        vadd.u8         d1,  d0,  d5
+        vext.8          d7,  d23, d23, #1
+        vld1.8          {q11},    [r0], r12
+        vst1.32         {d0[0]},  [r1,:32], r2
+        vadd.u8         d2,  d1,  d6
+        vst1.32         {d1[0]},  [r1], r2
+        vadd.u8         d3,  d2,  d7
+        vst1.32         {d2[0]},  [r1], r2
+        vst1.32         {d3[0]},  [r1], r2
+        subs            r3,  r3,  #12
+        bgt             1b
+
+        bx              lr
+endfunc
+
+func    png_read_filter_row_up_neon, export=1
+        ldr             r3,  [r0, #4]           @ rowbytes
+1:
+        vld1.8          {q0}, [r1,:128]
+        vld1.8          {q1}, [r2,:128]!
+        vadd.u8         q0,  q0,  q1
+        vst1.8          {q0}, [r1,:128]!
+        subs            r3,  r3,  #16
+        bgt             1b
+
+        bx              lr
+endfunc
+
+func    png_read_filter_row_avg4_neon, export=1
+        ldr             r12, [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+1:
+        vld4.32         {d4[],d5[],d6[],d7[]},    [r1,:128]
+        vld4.32         {d16[],d17[],d18[],d19[]},[r2,:128]!
+        vhadd.u8        d0,  d3,  d16
+        vadd.u8         d0,  d0,  d4
+        vhadd.u8        d1,  d0,  d17
+        vadd.u8         d1,  d1,  d5
+        vhadd.u8        d2,  d1,  d18
+        vadd.u8         d2,  d2,  d6
+        vhadd.u8        d3,  d2,  d19
+        vadd.u8         d3,  d3,  d7
+        vst4.32         {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
+        subs            r12, r12, #16
+        bgt             1b
+
+        bx              lr
+endfunc
+
+func    png_read_filter_row_avg3_neon, export=1
+        push            {r4,lr}
+        ldr             r12, [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+        mov             r0,  r1
+        mov             r4,  #3
+        mov             lr,  #12
+        vld1.8          {q11},    [r0], lr
+1:
+        vld1.8          {q10},    [r2], lr
+        vext.8          d5,  d22, d23, #3
+        vhadd.u8        d0,  d3,  d20
+        vext.8          d17, d20, d21, #3
+        vadd.u8         d0,  d0,  d22
+        vext.8          d6,  d22, d23, #6
+        vhadd.u8        d1,  d0,  d17
+        vext.8          d18, d20, d21, #6
+        vadd.u8         d1,  d1,  d5
+        vext.8          d7,  d23, d23, #1
+        vld1.8          {q11},    [r0], lr
+        vst1.32         {d0[0]},  [r1,:32], r4
+        vhadd.u8        d2,  d1,  d18
+        vst1.32         {d1[0]},  [r1], r4
+        vext.8          d19, d21, d21, #1
+        vadd.u8         d2,  d2,  d6
+        vhadd.u8        d3,  d2,  d19
+        vst1.32         {d2[0]},  [r1], r4
+        vadd.u8         d3,  d3,  d7
+        vst1.32         {d3[0]},  [r1], r4
+        subs            r12, r12, #12
+        bgt             1b
+
+        pop             {r4,pc}
+endfunc
+
+.macro  paeth           rx,  ra,  rb,  rc
+        vaddl.u8        q12, \ra, \rb           @ a + b
+        vaddl.u8        q15, \rc, \rc           @ 2*c
+        vabdl.u8        q13, \rb, \rc           @ pa
+        vabdl.u8        q14, \ra, \rc           @ pb
+        vabd.u16        q15, q12, q15           @ pc
+        vcle.u16        q12, q13, q14           @ pa <= pb
+        vcle.u16        q13, q13, q15           @ pa <= pc
+        vcle.u16        q14, q14, q15           @ pb <= pc
+        vand            q12, q12, q13           @ pa <= pb && pa <= pc
+        vmovn.u16       d28, q14
+        vmovn.u16       \rx, q12
+        vbsl            d28, \rb, \rc
+        vbsl            \rx, \ra, d28
+.endm
+
+func    png_read_filter_row_paeth4_neon, export=1
+        ldr             r12, [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+        vmov.i8         d20, #0
+1:
+        vld4.32         {d4[],d5[],d6[],d7[]},    [r1,:128]
+        vld4.32         {d16[],d17[],d18[],d19[]},[r2,:128]!
+        paeth           d0,  d3,  d16, d20
+        vadd.u8         d0,  d0,  d4
+        paeth           d1,  d0,  d17, d16
+        vadd.u8         d1,  d1,  d5
+        paeth           d2,  d1,  d18, d17
+        vadd.u8         d2,  d2,  d6
+        paeth           d3,  d2,  d19, d18
+        vmov            d20, d19
+        vadd.u8         d3,  d3,  d7
+        vst4.32         {d0[0],d1[0],d2[0],d3[0]},[r1,:128]!
+        subs            r12, r12, #16
+        bgt             1b
+
+        bx              lr
+endfunc
+
+func    png_read_filter_row_paeth3_neon, export=1
+        push            {r4,lr}
+        ldr             r12, [r0, #4]           @ rowbytes
+        vmov.i8         d3,  #0
+        vmov.i8         d4,  #0
+        mov             r0,  r1
+        mov             r4,  #3
+        mov             lr,  #12
+        vld1.8          {q11},    [r0], lr
+1:
+        vld1.8          {q10},    [r2], lr
+        paeth           d0,  d3,  d20, d4
+        vext.8          d5,  d22, d23, #3
+        vadd.u8         d0,  d0,  d22
+        vext.8          d17, d20, d21, #3
+        paeth           d1,  d0,  d17, d20
+        vst1.32         {d0[0]},  [r1,:32], r4
+        vext.8          d6,  d22, d23, #6
+        vadd.u8         d1,  d1,  d5
+        vext.8          d18, d20, d21, #6
+        paeth           d2,  d1,  d18, d17
+        vext.8          d7,  d23, d23, #1
+        vld1.8          {q11},    [r0], lr
+        vst1.32         {d1[0]},  [r1], r4
+        vadd.u8         d2,  d2,  d6
+        vext.8          d19, d21, d21, #1
+        paeth           d3,  d2,  d19, d18
+        vst1.32         {d2[0]},  [r1], r4
+        vmov            d4,  d19
+        vadd.u8         d3,  d3,  d7
+        vst1.32         {d3[0]},  [r1], r4
+        subs            r12, r12, #12
+        bgt             1b
+
+        pop             {r4,pc}
+endfunc
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* PNG_READ_SUPPORTED */
+#endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */
diff --git a/ml/dlib/dlib/external/libpng/arm/filter_neon_intrinsics.c b/ml/dlib/dlib/external/libpng/arm/filter_neon_intrinsics.c
new file mode 100644
index 000000000..e6a0217ab
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/arm/filter_neon_intrinsics.c
@@ -0,0 +1,372 @@
+
+/* filter_neon_intrinsics.c - NEON optimised filter functions
+ *
+ * Copyright (c) 2013 Glenn Randers-Pehrson
+ * Written by James Yu <james.yu at linaro.org>, October 2013.
+ * Based on filter_neon.S, written by Mans Rullgard, 2011.
+ *
+ * Last changed in libpng 1.6.7 [November 14, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+/* This code requires -mfpu=neon on the command line: */
+#if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code */
+
+#include <arm_neon.h>
+
+/* libpng row pointers are not necessarily aligned to any particular boundary,
+ * however this code will only work with appropriate alignment.  arm/arm_init.c
+ * checks for this (and will not compile unless it is done), this code uses
+ * variants of png_aligncast to avoid compiler warnings.
+ */
+#define png_ptr(type,pointer) png_aligncast(type *,pointer)
+#define png_ptrc(type,pointer) png_aligncastconst(const type *,pointer)
+
+/* The following relies on a variable 'temp_pointer' being declared with type
+ * 'type'.  This is written this way just to hide the GCC strict aliasing
+ * warning; note that the code is safe because there never is an alias between
+ * the input and output pointers.
+ */
+#define png_ldr(type,pointer)\
+   (temp_pointer = png_ptr(type,pointer), *temp_pointer)
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_ARM_NEON_OPT > 0
+
+void
+png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint8x16_t qrp, qpp;
+
+      qrp = vld1q_u8(rp);
+      qpp = vld1q_u8(pp);
+      qrp = vaddq_u8(qrp, qpp);
+      vst1q_u8(rp, qrp);
+   }
+}
+
+void
+png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp = vld1q_u8(rp);
+   uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   uint8x8x2_t vrp = *vrpt;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop;)
+   {
+      uint8x8_t vtmp1, vtmp2;
+      uint32x2_t *temp_pointer;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
+      vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vdest.val[2] = vadd_u8(vdest.val[1], vtmp2);
+      vdest.val[3] = vadd_u8(vdest.val[2], vtmp1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+
+   PNG_UNUSED(prev_row)
+}
+
+void
+png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16)
+   {
+      uint32x2x4_t vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      uint8x8x4_t *vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      uint8x8x4_t vrp = *vrpt;
+      uint32x2x4_t *temp_pointer;
+
+      vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]);
+      vdest.val[2] = vadd_u8(vdest.val[1], vrp.val[2]);
+      vdest.val[3] = vadd_u8(vdest.val[2], vrp.val[3]);
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+
+   PNG_UNUSED(prev_row)
+}
+
+void
+png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t,&vtmp);
+   vrp = *vrpt;
+
+   for (; rp < rp_stop; pp += 12)
+   {
+      uint8x8_t vtmp1, vtmp2, vtmp3;
+
+      uint8x8x2_t *vppt;
+      uint8x8x2_t vpp;
+
+      uint32x2_t *temp_pointer;
+
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t,&vtmp);
+      vpp = *vppt;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 6);
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t,&vtmp);
+      vrp = *vrpt;
+
+      vdest.val[2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[2] = vadd_u8(vdest.val[2], vtmp3);
+
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vdest.val[3] = vhadd_u8(vdest.val[2], vtmp2);
+      vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+}
+
+void
+png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+
+      vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t,pp));
+      vppt = png_ptr(uint8x8x4_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[2] = vhadd_u8(vdest.val[1], vpp.val[2]);
+      vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
+      vdest.val[3] = vhadd_u8(vdest.val[2], vpp.val[3]);
+      vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
+
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+}
+
+static uint8x8_t
+paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c)
+{
+   uint8x8_t d, e;
+   uint16x8_t p1, pa, pb, pc;
+
+   p1 = vaddl_u8(a, b); /* a + b */
+   pc = vaddl_u8(c, c); /* c * 2 */
+   pa = vabdl_u8(b, c); /* pa */
+   pb = vabdl_u8(a, c); /* pb */
+   pc = vabdq_u16(p1, pc); /* pc */
+
+   p1 = vcleq_u16(pa, pb); /* pa <= pb */
+   pa = vcleq_u16(pa, pc); /* pa <= pc */
+   pb = vcleq_u16(pb, pc); /* pb <= pc */
+
+   p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */
+
+   d = vmovn_u16(pb);
+   e = vmovn_u16(p1);
+
+   d = vbsl_u8(d, b, c);
+   e = vbsl_u8(e, a, d);
+
+   return e;
+}
+
+void
+png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t,&vtmp);
+   vrp = *vrpt;
+
+   for (; rp < rp_stop; pp += 12)
+   {
+      uint8x8x2_t *vppt;
+      uint8x8x2_t vpp;
+      uint8x8_t vtmp1, vtmp2, vtmp3;
+      uint32x2_t *temp_pointer;
+
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6);
+      vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[2] = vadd_u8(vdest.val[2], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t,&vtmp);
+      vrp = *vrpt;
+
+      vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3);
+      vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
+
+      vlast = vtmp2;
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+}
+
+void
+png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+
+      vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t,pp));
+      vppt = png_ptr(uint8x8x4_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[2] = paeth(vdest.val[1], vpp.val[2], vpp.val[1]);
+      vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
+      vdest.val[3] = paeth(vdest.val[2], vpp.val[3], vpp.val[2]);
+      vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
+
+      vlast = vpp.val[3];
+
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+}
+
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* PNG_READ_SUPPORTED */
+#endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */
diff --git a/ml/dlib/dlib/external/libpng/png.c b/ml/dlib/dlib/external/libpng/png.c
new file mode 100644
index 000000000..efcc6eead
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/png.c
@@ -0,0 +1,4299 @@
+
+/* png.c - location for general purpose libpng functions
+ *
+ * Last changed in libpng 1.6.2 [April 25, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+
+/* Generate a compiler error if there is an old png.h in the search path. */
+typedef png_libpng_version_1_6_7 Your_png_h_is_not_version_1_6_7;
+
+/* Tells libpng that we have already handled the first "num_bytes" bytes
+ * of the PNG file signature.  If the PNG data is embedded into another
+ * stream we can set num_bytes = 8 so that libpng will not attempt to read
+ * or write any of the magic bytes before it starts on the IHDR.
+ */
+
+#ifdef PNG_READ_SUPPORTED
+void PNGAPI
+png_set_sig_bytes(png_structrp png_ptr, int num_bytes)
+{
+   png_debug(1, "in png_set_sig_bytes");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (num_bytes > 8)
+      png_error(png_ptr, "Too many bytes for PNG signature");
+
+   png_ptr->sig_bytes = (png_byte)(num_bytes < 0 ? 0 : num_bytes);
+}
+
+/* Checks whether the supplied bytes match the PNG signature.  We allow
+ * checking less than the full 8-byte signature so that those apps that
+ * already read the first few bytes of a file to determine the file type
+ * can simply check the remaining bytes for extra assurance.  Returns
+ * an integer less than, equal to, or greater than zero if sig is found,
+ * respectively, to be less than, to match, or be greater than the correct
+ * PNG signature (this is the same behavior as strcmp, memcmp, etc).
+ */
+int PNGAPI
+png_sig_cmp(png_const_bytep sig, png_size_t start, png_size_t num_to_check)
+{
+   png_byte png_signature[8] = {137, 80, 78, 71, 13, 10, 26, 10};
+
+   if (num_to_check > 8)
+      num_to_check = 8;
+
+   else if (num_to_check < 1)
+      return (-1);
+
+   if (start > 7)
+      return (-1);
+
+   if (start + num_to_check > 8)
+      num_to_check = 8 - start;
+
+   return ((int)(memcmp(&sig[start], &png_signature[start], num_to_check)));
+}
+
+#endif /* PNG_READ_SUPPORTED */
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+/* Function to allocate memory for zlib */
+PNG_FUNCTION(voidpf /* PRIVATE */,
+png_zalloc,(voidpf png_ptr, uInt items, uInt size),PNG_ALLOCATED)
+{
+   png_alloc_size_t num_bytes = size;
+
+   if (png_ptr == NULL)
+      return NULL;
+
+   if (items >= (~(png_alloc_size_t)0)/size)
+   {
+      png_warning (png_voidcast(png_structrp, png_ptr),
+         "Potential overflow in png_zalloc()");
+      return NULL;
+   }
+
+   num_bytes *= items;
+   return png_malloc_warn(png_voidcast(png_structrp, png_ptr), num_bytes);
+}
+
+/* Function to free memory for zlib */
+void /* PRIVATE */
+png_zfree(voidpf png_ptr, voidpf ptr)
+{
+   png_free(png_voidcast(png_const_structrp,png_ptr), ptr);
+}
+
+/* Reset the CRC variable to 32 bits of 1's.  Care must be taken
+ * in case CRC is > 32 bits to leave the top bits 0.
+ */
+void /* PRIVATE */
+png_reset_crc(png_structrp png_ptr)
+{
+   /* The cast is safe because the crc is a 32 bit value. */
+   png_ptr->crc = (png_uint_32)crc32(0, Z_NULL, 0);
+}
+
+/* Calculate the CRC over a section of data.  We can only pass as
+ * much data to this routine as the largest single buffer size.  We
+ * also check that this data will actually be used before going to the
+ * trouble of calculating it.
+ */
+void /* PRIVATE */
+png_calculate_crc(png_structrp png_ptr, png_const_bytep ptr, png_size_t length)
+{
+   int need_crc = 1;
+
+   if (PNG_CHUNK_ANCILLARY(png_ptr->chunk_name))
+   {
+      if ((png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_MASK) ==
+          (PNG_FLAG_CRC_ANCILLARY_USE | PNG_FLAG_CRC_ANCILLARY_NOWARN))
+         need_crc = 0;
+   }
+
+   else /* critical */
+   {
+      if (png_ptr->flags & PNG_FLAG_CRC_CRITICAL_IGNORE)
+         need_crc = 0;
+   }
+
+   /* 'uLong' is defined in zlib.h as unsigned long; this means that on some
+    * systems it is a 64 bit value.  crc32, however, returns 32 bits so the
+    * following cast is safe.  'uInt' may be no more than 16 bits, so it is
+    * necessary to perform a loop here.
+    */
+   if (need_crc && length > 0)
+   {
+      uLong crc = png_ptr->crc; /* Should never issue a warning */
+
+      do
+      {
+         uInt safe_length = (uInt)length;
+         if (safe_length == 0)
+            safe_length = (uInt)-1; /* evil, but safe */
+
+         crc = crc32(crc, ptr, safe_length);
+
+         /* The following should never issue compiler warnings; if they do the
+          * target system has characteristics that will probably violate other
+          * assumptions within the libpng code.
+          */
+         ptr += safe_length;
+         length -= safe_length;
+      }
+      while (length > 0);
+
+      /* And the following is always safe because the crc is only 32 bits. */
+      png_ptr->crc = (png_uint_32)crc;
+   }
+}
+
+/* Check a user supplied version number, called from both read and write
+ * functions that create a png_struct.
+ */
+int
+png_user_version_check(png_structrp png_ptr, png_const_charp user_png_ver)
+{
+   if (user_png_ver)
+   {
+      int i = 0;
+
+      do
+      {
+         if (user_png_ver[i] != png_libpng_ver[i])
+            png_ptr->flags |= PNG_FLAG_LIBRARY_MISMATCH;
+      } while (png_libpng_ver[i++]);
+   }
+
+   else
+      png_ptr->flags |= PNG_FLAG_LIBRARY_MISMATCH;
+
+   if (png_ptr->flags & PNG_FLAG_LIBRARY_MISMATCH)
+   {
+     /* Libpng 0.90 and later are binary incompatible with libpng 0.89, so
+      * we must recompile any applications that use any older library version.
+      * For versions after libpng 1.0, we will be compatible, so we need
+      * only check the first and third digits (note that when we reach version
+      * 1.10 we will need to check the fourth symbol, namely user_png_ver[3]).
+      */
+      if (user_png_ver == NULL || user_png_ver[0] != png_libpng_ver[0] ||
+          (user_png_ver[0] == '1' && (user_png_ver[2] != png_libpng_ver[2] ||
+          user_png_ver[3] != png_libpng_ver[3])) ||
+          (user_png_ver[0] == '0' && user_png_ver[2] < '9'))
+      {
+#ifdef PNG_WARNINGS_SUPPORTED
+         size_t pos = 0;
+         char m[128];
+
+         pos = png_safecat(m, (sizeof m), pos,
+             "Application built with libpng-");
+         pos = png_safecat(m, (sizeof m), pos, user_png_ver);
+         pos = png_safecat(m, (sizeof m), pos, " but running with ");
+         pos = png_safecat(m, (sizeof m), pos, png_libpng_ver);
+
+         png_warning(png_ptr, m);
+#endif
+
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+         png_ptr->flags = 0;
+#endif
+
+         return 0;
+      }
+   }
+
+   /* Success return. */
+   return 1;
+}
+
+/* Generic function to create a png_struct for either read or write - this
+ * contains the common initialization.
+ */
+PNG_FUNCTION(png_structp /* PRIVATE */,
+png_create_png_struct,(png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn, png_voidp mem_ptr,
+    png_malloc_ptr malloc_fn, png_free_ptr free_fn),PNG_ALLOCATED)
+{
+   png_struct create_struct;
+#  ifdef PNG_SETJMP_SUPPORTED
+      jmp_buf create_jmp_buf;
+#  endif
+
+   /* This temporary stack-allocated structure is used to provide a place to
+    * build enough context to allow the user provided memory allocator (if any)
+    * to be called.
+    */
+   memset(&create_struct, 0, (sizeof create_struct));
+
+   /* Added at libpng-1.2.6 */
+#  ifdef PNG_USER_LIMITS_SUPPORTED
+      create_struct.user_width_max = PNG_USER_WIDTH_MAX;
+      create_struct.user_height_max = PNG_USER_HEIGHT_MAX;
+
+#     ifdef PNG_USER_CHUNK_CACHE_MAX
+         /* Added at libpng-1.2.43 and 1.4.0 */
+         create_struct.user_chunk_cache_max = PNG_USER_CHUNK_CACHE_MAX;
+#     endif
+
+#     ifdef PNG_USER_CHUNK_MALLOC_MAX
+         /* Added at libpng-1.2.43 and 1.4.1, required only for read but exists
+          * in png_struct regardless.
+          */
+         create_struct.user_chunk_malloc_max = PNG_USER_CHUNK_MALLOC_MAX;
+#     endif
+#  endif
+
+   /* The following two API calls simply set fields in png_struct, so it is safe
+    * to do them now even though error handling is not yet set up.
+    */
+#  ifdef PNG_USER_MEM_SUPPORTED
+      png_set_mem_fn(&create_struct, mem_ptr, malloc_fn, free_fn);
+#  endif
+
+   /* (*error_fn) can return control to the caller after the error_ptr is set,
+    * this will result in a memory leak unless the error_fn does something
+    * extremely sophisticated.  The design lacks merit but is implicit in the
+    * API.
+    */
+   png_set_error_fn(&create_struct, error_ptr, error_fn, warn_fn);
+
+#  ifdef PNG_SETJMP_SUPPORTED
+      if (!setjmp(create_jmp_buf))
+      {
+         /* Temporarily fake out the longjmp information until we have
+          * successfully completed this function.  This only works if we have
+          * setjmp() support compiled in, but it is safe - this stuff should
+          * never happen.
+          */
+         create_struct.jmp_buf_ptr = &create_jmp_buf;
+         create_struct.jmp_buf_size = 0; /*stack allocation*/
+         create_struct.longjmp_fn = longjmp;
+#  else
+      {
+#  endif
+         /* Call the general version checker (shared with read and write code):
+          */
+         if (png_user_version_check(&create_struct, user_png_ver))
+         {
+            png_structrp png_ptr = png_voidcast(png_structrp,
+               png_malloc_warn(&create_struct, (sizeof *png_ptr)));
+
+            if (png_ptr != NULL)
+            {
+               /* png_ptr->zstream holds a back-pointer to the png_struct, so
+                * this can only be done now:
+                */
+               create_struct.zstream.zalloc = png_zalloc;
+               create_struct.zstream.zfree = png_zfree;
+               create_struct.zstream.opaque = png_ptr;
+
+#              ifdef PNG_SETJMP_SUPPORTED
+                  /* Eliminate the local error handling: */
+                  create_struct.jmp_buf_ptr = NULL;
+                  create_struct.jmp_buf_size = 0;
+                  create_struct.longjmp_fn = 0;
+#              endif
+
+               *png_ptr = create_struct;
+
+               /* This is the successful return point */
+               return png_ptr;
+            }
+         }
+      }
+
+   /* A longjmp because of a bug in the application storage allocator or a
+    * simple failure to allocate the png_struct.
+    */
+   return NULL;
+}
+
+/* Allocate the memory for an info_struct for the application. */
+PNG_FUNCTION(png_infop,PNGAPI
+png_create_info_struct,(png_const_structrp png_ptr),PNG_ALLOCATED)
+{
+   png_inforp info_ptr;
+
+   png_debug(1, "in png_create_info_struct");
+
+   if (png_ptr == NULL)
+      return NULL;
+
+   /* Use the internal API that does not (or at least should not) error out, so
+    * that this call always returns ok.  The application typically sets up the
+    * error handling *after* creating the info_struct because this is the way it
+    * has always been done in 'example.c'.
+    */
+   info_ptr = png_voidcast(png_inforp, png_malloc_base(png_ptr,
+      (sizeof *info_ptr)));
+
+   if (info_ptr != NULL)
+      memset(info_ptr, 0, (sizeof *info_ptr));
+
+   return info_ptr;
+}
+
+/* This function frees the memory associated with a single info struct.
+ * Normally, one would use either png_destroy_read_struct() or
+ * png_destroy_write_struct() to free an info struct, but this may be
+ * useful for some applications.  From libpng 1.6.0 this function is also used
+ * internally to implement the png_info release part of the 'struct' destroy
+ * APIs.  This ensures that all possible approaches free the same data (all of
+ * it).
+ */
+void PNGAPI
+png_destroy_info_struct(png_const_structrp png_ptr, png_infopp info_ptr_ptr)
+{
+   png_inforp info_ptr = NULL;
+
+   png_debug(1, "in png_destroy_info_struct");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (info_ptr_ptr != NULL)
+      info_ptr = *info_ptr_ptr;
+
+   if (info_ptr != NULL)
+   {
+      /* Do this first in case of an error below; if the app implements its own
+       * memory management this can lead to png_free calling png_error, which
+       * will abort this routine and return control to the app error handler.
+       * An infinite loop may result if it then tries to free the same info
+       * ptr.
+       */
+      *info_ptr_ptr = NULL;
+
+      png_free_data(png_ptr, info_ptr, PNG_FREE_ALL, -1);
+      memset(info_ptr, 0, (sizeof *info_ptr));
+      png_free(png_ptr, info_ptr);
+   }
+}
+
+/* Initialize the info structure.  This is now an internal function (0.89)
+ * and applications using it are urged to use png_create_info_struct()
+ * instead.  Use deprecated in 1.6.0, internal use removed (used internally it
+ * is just a memset).
+ *
+ * NOTE: it is almost inconceivable that this API is used because it bypasses
+ * the user-memory mechanism and the user error handling/warning mechanisms in
+ * those cases where it does anything other than a memset.
+ */
+PNG_FUNCTION(void,PNGAPI
+png_info_init_3,(png_infopp ptr_ptr, png_size_t png_info_struct_size),
+   PNG_DEPRECATED)
+{
+   png_inforp info_ptr = *ptr_ptr;
+
+   png_debug(1, "in png_info_init_3");
+
+   if (info_ptr == NULL)
+      return;
+
+   if ((sizeof (png_info)) > png_info_struct_size)
+   {
+      *ptr_ptr = NULL;
+      /* The following line is why this API should not be used: */
+      free(info_ptr);
+      info_ptr = png_voidcast(png_inforp, png_malloc_base(NULL,
+         (sizeof *info_ptr)));
+      *ptr_ptr = info_ptr;
+   }
+
+   /* Set everything to 0 */
+   memset(info_ptr, 0, (sizeof *info_ptr));
+}
+
+/* The following API is not called internally */
+void PNGAPI
+png_data_freer(png_const_structrp png_ptr, png_inforp info_ptr,
+   int freer, png_uint_32 mask)
+{
+   png_debug(1, "in png_data_freer");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (freer == PNG_DESTROY_WILL_FREE_DATA)
+      info_ptr->free_me |= mask;
+
+   else if (freer == PNG_USER_WILL_FREE_DATA)
+      info_ptr->free_me &= ~mask;
+
+   else
+      png_error(png_ptr, "Unknown freer parameter in png_data_freer");
+}
+
+void PNGAPI
+png_free_data(png_const_structrp png_ptr, png_inforp info_ptr, png_uint_32 mask,
+   int num)
+{
+   png_debug(1, "in png_free_data");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+#ifdef PNG_TEXT_SUPPORTED
+   /* Free text item num or (if num == -1) all text items */
+   if ((mask & PNG_FREE_TEXT) & info_ptr->free_me)
+   {
+      if (num != -1)
+      {
+         if (info_ptr->text && info_ptr->text[num].key)
+         {
+            png_free(png_ptr, info_ptr->text[num].key);
+            info_ptr->text[num].key = NULL;
+         }
+      }
+
+      else
+      {
+         int i;
+         for (i = 0; i < info_ptr->num_text; i++)
+             png_free_data(png_ptr, info_ptr, PNG_FREE_TEXT, i);
+         png_free(png_ptr, info_ptr->text);
+         info_ptr->text = NULL;
+         info_ptr->num_text=0;
+      }
+   }
+#endif
+
+#ifdef PNG_tRNS_SUPPORTED
+   /* Free any tRNS entry */
+   if ((mask & PNG_FREE_TRNS) & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->trans_alpha);
+      info_ptr->trans_alpha = NULL;
+      info_ptr->valid &= ~PNG_INFO_tRNS;
+   }
+#endif
+
+#ifdef PNG_sCAL_SUPPORTED
+   /* Free any sCAL entry */
+   if ((mask & PNG_FREE_SCAL) & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->scal_s_width);
+      png_free(png_ptr, info_ptr->scal_s_height);
+      info_ptr->scal_s_width = NULL;
+      info_ptr->scal_s_height = NULL;
+      info_ptr->valid &= ~PNG_INFO_sCAL;
+   }
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+   /* Free any pCAL entry */
+   if ((mask & PNG_FREE_PCAL) & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->pcal_purpose);
+      png_free(png_ptr, info_ptr->pcal_units);
+      info_ptr->pcal_purpose = NULL;
+      info_ptr->pcal_units = NULL;
+      if (info_ptr->pcal_params != NULL)
+         {
+            unsigned int i;
+            for (i = 0; i < info_ptr->pcal_nparams; i++)
+            {
+               png_free(png_ptr, info_ptr->pcal_params[i]);
+               info_ptr->pcal_params[i] = NULL;
+            }
+            png_free(png_ptr, info_ptr->pcal_params);
+            info_ptr->pcal_params = NULL;
+         }
+      info_ptr->valid &= ~PNG_INFO_pCAL;
+   }
+#endif
+
+#ifdef PNG_iCCP_SUPPORTED
+   /* Free any profile entry */
+   if ((mask & PNG_FREE_ICCP) & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->iccp_name);
+      png_free(png_ptr, info_ptr->iccp_profile);
+      info_ptr->iccp_name = NULL;
+      info_ptr->iccp_profile = NULL;
+      info_ptr->valid &= ~PNG_INFO_iCCP;
+   }
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+   /* Free a given sPLT entry, or (if num == -1) all sPLT entries */
+   if ((mask & PNG_FREE_SPLT) & info_ptr->free_me)
+   {
+      if (num != -1)
+      {
+         if (info_ptr->splt_palettes)
+         {
+            png_free(png_ptr, info_ptr->splt_palettes[num].name);
+            png_free(png_ptr, info_ptr->splt_palettes[num].entries);
+            info_ptr->splt_palettes[num].name = NULL;
+            info_ptr->splt_palettes[num].entries = NULL;
+         }
+      }
+
+      else
+      {
+         if (info_ptr->splt_palettes_num)
+         {
+            int i;
+            for (i = 0; i < info_ptr->splt_palettes_num; i++)
+               png_free_data(png_ptr, info_ptr, PNG_FREE_SPLT, (int)i);
+
+            png_free(png_ptr, info_ptr->splt_palettes);
+            info_ptr->splt_palettes = NULL;
+            info_ptr->splt_palettes_num = 0;
+         }
+         info_ptr->valid &= ~PNG_INFO_sPLT;
+      }
+   }
+#endif
+
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+   if ((mask & PNG_FREE_UNKN) & info_ptr->free_me)
+   {
+      if (num != -1)
+      {
+          if (info_ptr->unknown_chunks)
+          {
+             png_free(png_ptr, info_ptr->unknown_chunks[num].data);
+             info_ptr->unknown_chunks[num].data = NULL;
+          }
+      }
+
+      else
+      {
+         int i;
+
+         if (info_ptr->unknown_chunks_num)
+         {
+            for (i = 0; i < info_ptr->unknown_chunks_num; i++)
+               png_free_data(png_ptr, info_ptr, PNG_FREE_UNKN, (int)i);
+
+            png_free(png_ptr, info_ptr->unknown_chunks);
+            info_ptr->unknown_chunks = NULL;
+            info_ptr->unknown_chunks_num = 0;
+         }
+      }
+   }
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+   /* Free any hIST entry */
+   if ((mask & PNG_FREE_HIST)  & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->hist);
+      info_ptr->hist = NULL;
+      info_ptr->valid &= ~PNG_INFO_hIST;
+   }
+#endif
+
+   /* Free any PLTE entry that was internally allocated */
+   if ((mask & PNG_FREE_PLTE) & info_ptr->free_me)
+   {
+      png_free(png_ptr, info_ptr->palette);
+      info_ptr->palette = NULL;
+      info_ptr->valid &= ~PNG_INFO_PLTE;
+      info_ptr->num_palette = 0;
+   }
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+   /* Free any image bits attached to the info structure */
+   if ((mask & PNG_FREE_ROWS) & info_ptr->free_me)
+   {
+      if (info_ptr->row_pointers)
+      {
+         png_uint_32 row;
+         for (row = 0; row < info_ptr->height; row++)
+         {
+            png_free(png_ptr, info_ptr->row_pointers[row]);
+            info_ptr->row_pointers[row] = NULL;
+         }
+         png_free(png_ptr, info_ptr->row_pointers);
+         info_ptr->row_pointers = NULL;
+      }
+      info_ptr->valid &= ~PNG_INFO_IDAT;
+   }
+#endif
+
+   if (num != -1)
+      mask &= ~PNG_FREE_MUL;
+
+   info_ptr->free_me &= ~mask;
+}
+#endif /* defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED) */
+
+/* This function returns a pointer to the io_ptr associated with the user
+ * functions.  The application should free any memory associated with this
+ * pointer before png_write_destroy() or png_read_destroy() are called.
+ */
+png_voidp PNGAPI
+png_get_io_ptr(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return (NULL);
+
+   return (png_ptr->io_ptr);
+}
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+#  ifdef PNG_STDIO_SUPPORTED
+/* Initialize the default input/output functions for the PNG file.  If you
+ * use your own read or write routines, you can call either png_set_read_fn()
+ * or png_set_write_fn() instead of png_init_io().  If you have defined
+ * PNG_NO_STDIO or otherwise disabled PNG_STDIO_SUPPORTED, you must use a
+ * function of your own because "FILE *" isn't necessarily available.
+ */
+void PNGAPI
+png_init_io(png_structrp png_ptr, png_FILE_p fp)
+{
+   png_debug(1, "in png_init_io");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->io_ptr = (png_voidp)fp;
+}
+#  endif
+
+#ifdef PNG_SAVE_INT_32_SUPPORTED
+/* The png_save_int_32 function assumes integers are stored in two's
+ * complement format.  If this isn't the case, then this routine needs to
+ * be modified to write data in two's complement format.  Note that,
+ * the following works correctly even if png_int_32 has more than 32 bits
+ * (compare the more complex code required on read for sign extension.)
+ */
+void PNGAPI
+png_save_int_32(png_bytep buf, png_int_32 i)
+{
+   buf[0] = (png_byte)((i >> 24) & 0xff);
+   buf[1] = (png_byte)((i >> 16) & 0xff);
+   buf[2] = (png_byte)((i >> 8) & 0xff);
+   buf[3] = (png_byte)(i & 0xff);
+}
+#endif
+
+#  ifdef PNG_TIME_RFC1123_SUPPORTED
+/* Convert the supplied time into an RFC 1123 string suitable for use in
+ * a "Creation Time" or other text-based time string.
+ */
+int PNGAPI
+png_convert_to_rfc1123_buffer(char out[29], png_const_timep ptime)
+{
+   static PNG_CONST char short_months[12][4] =
+        {"Jan", "Feb", "Mar", "Apr", "May", "Jun",
+         "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"};
+
+   if (out == NULL)
+      return 0;
+
+   if (ptime->year > 9999 /* RFC1123 limitation */ ||
+       ptime->month == 0    ||  ptime->month > 12  ||
+       ptime->day   == 0    ||  ptime->day   > 31  ||
+       ptime->hour  > 23    ||  ptime->minute > 59 ||
+       ptime->second > 60)
+      return 0;
+
+   {
+      size_t pos = 0;
+      char number_buf[5]; /* enough for a four-digit year */
+
+#     define APPEND_STRING(string) pos = png_safecat(out, 29, pos, (string))
+#     define APPEND_NUMBER(format, value)\
+         APPEND_STRING(PNG_FORMAT_NUMBER(number_buf, format, (value)))
+#     define APPEND(ch) if (pos < 28) out[pos++] = (ch)
+
+      APPEND_NUMBER(PNG_NUMBER_FORMAT_u, (unsigned)ptime->day);
+      APPEND(' ');
+      APPEND_STRING(short_months[(ptime->month - 1)]);
+      APPEND(' ');
+      APPEND_NUMBER(PNG_NUMBER_FORMAT_u, ptime->year);
+      APPEND(' ');
+      APPEND_NUMBER(PNG_NUMBER_FORMAT_02u, (unsigned)ptime->hour);
+      APPEND(':');
+      APPEND_NUMBER(PNG_NUMBER_FORMAT_02u, (unsigned)ptime->minute);
+      APPEND(':');
+      APPEND_NUMBER(PNG_NUMBER_FORMAT_02u, (unsigned)ptime->second);
+      APPEND_STRING(" +0000"); /* This reliably terminates the buffer */
+
+#     undef APPEND
+#     undef APPEND_NUMBER
+#     undef APPEND_STRING
+   }
+
+   return 1;
+}
+
+#     if PNG_LIBPNG_VER < 10700
+/* To do: remove the following from libpng-1.7 */
+/* Original API that uses a private buffer in png_struct.
+ * Deprecated because it causes png_struct to carry a spurious temporary
+ * buffer (png_struct::time_buffer), better to have the caller pass this in.
+ */
+png_const_charp PNGAPI
+png_convert_to_rfc1123(png_structrp png_ptr, png_const_timep ptime)
+{
+   if (png_ptr != NULL)
+   {
+      /* The only failure above if png_ptr != NULL is from an invalid ptime */
+      if (!png_convert_to_rfc1123_buffer(png_ptr->time_buffer, ptime))
+         png_warning(png_ptr, "Ignoring invalid time value");
+
+      else
+         return png_ptr->time_buffer;
+   }
+
+   return NULL;
+}
+#     endif
+#  endif /* PNG_TIME_RFC1123_SUPPORTED */
+
+#endif /* defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED) */
+
+png_const_charp PNGAPI
+png_get_copyright(png_const_structrp png_ptr)
+{
+   PNG_UNUSED(png_ptr)  /* Silence compiler warning about unused png_ptr */
+#ifdef PNG_STRING_COPYRIGHT
+   return PNG_STRING_COPYRIGHT
+#else
+#  ifdef __STDC__
+   return PNG_STRING_NEWLINE \
+     "libpng version 1.6.7 - November 14, 2013" PNG_STRING_NEWLINE \
+     "Copyright (c) 1998-2013 Glenn Randers-Pehrson" PNG_STRING_NEWLINE \
+     "Copyright (c) 1996-1997 Andreas Dilger" PNG_STRING_NEWLINE \
+     "Copyright (c) 1995-1996 Guy Eric Schalnat, Group 42, Inc." \
+     PNG_STRING_NEWLINE;
+#  else
+      return "libpng version 1.6.7 - November 14, 2013\
+      Copyright (c) 1998-2013 Glenn Randers-Pehrson\
+      Copyright (c) 1996-1997 Andreas Dilger\
+      Copyright (c) 1995-1996 Guy Eric Schalnat, Group 42, Inc.";
+#  endif
+#endif
+}
+
+/* The following return the library version as a short string in the
+ * format 1.0.0 through 99.99.99zz.  To get the version of *.h files
+ * used with your application, print out PNG_LIBPNG_VER_STRING, which
+ * is defined in png.h.
+ * Note: now there is no difference between png_get_libpng_ver() and
+ * png_get_header_ver().  Due to the version_nn_nn_nn typedef guard,
+ * it is guaranteed that png.c uses the correct version of png.h.
+ */
+png_const_charp PNGAPI
+png_get_libpng_ver(png_const_structrp png_ptr)
+{
+   /* Version of *.c files used when building libpng */
+   return png_get_header_ver(png_ptr);
+}
+
+png_const_charp PNGAPI
+png_get_header_ver(png_const_structrp png_ptr)
+{
+   /* Version of *.h files used when building libpng */
+   PNG_UNUSED(png_ptr)  /* Silence compiler warning about unused png_ptr */
+   return PNG_LIBPNG_VER_STRING;
+}
+
+png_const_charp PNGAPI
+png_get_header_version(png_const_structrp png_ptr)
+{
+   /* Returns longer string containing both version and date */
+   PNG_UNUSED(png_ptr)  /* Silence compiler warning about unused png_ptr */
+#ifdef __STDC__
+   return PNG_HEADER_VERSION_STRING
+#  ifndef PNG_READ_SUPPORTED
+   "     (NO READ SUPPORT)"
+#  endif
+   PNG_STRING_NEWLINE;
+#else
+   return PNG_HEADER_VERSION_STRING;
+#endif
+}
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+int PNGAPI
+png_handle_as_unknown(png_const_structrp png_ptr, png_const_bytep chunk_name)
+{
+   /* Check chunk_name and return "keep" value if it's on the list, else 0 */
+   png_const_bytep p, p_end;
+
+   if (png_ptr == NULL || chunk_name == NULL || png_ptr->num_chunk_list == 0)
+      return PNG_HANDLE_CHUNK_AS_DEFAULT;
+
+   p_end = png_ptr->chunk_list;
+   p = p_end + png_ptr->num_chunk_list*5; /* beyond end */
+
+   /* The code is the fifth byte after each four byte string.  Historically this
+    * code was always searched from the end of the list, this is no longer
+    * necessary because the 'set' routine handles duplicate entries correcty.
+    */
+   do /* num_chunk_list > 0, so at least one */
+   {
+      p -= 5;
+
+      if (!memcmp(chunk_name, p, 4))
+         return p[4];
+   }
+   while (p > p_end);
+
+   /* This means that known chunks should be processed and unknown chunks should
+    * be handled according to the value of png_ptr->unknown_default; this can be
+    * confusing because, as a result, there are two levels of defaulting for
+    * unknown chunks.
+    */
+   return PNG_HANDLE_CHUNK_AS_DEFAULT;
+}
+
+#if defined(PNG_READ_UNKNOWN_CHUNKS_SUPPORTED) ||\
+   defined(PNG_HANDLE_AS_UNKNOWN_SUPPORTED)
+int /* PRIVATE */
+png_chunk_unknown_handling(png_const_structrp png_ptr, png_uint_32 chunk_name)
+{
+   png_byte chunk_string[5];
+
+   PNG_CSTRING_FROM_CHUNK(chunk_string, chunk_name);
+   return png_handle_as_unknown(png_ptr, chunk_string);
+}
+#endif /* READ_UNKNOWN_CHUNKS || HANDLE_AS_UNKNOWN */
+#endif /* SET_UNKNOWN_CHUNKS */
+
+#ifdef PNG_READ_SUPPORTED
+/* This function, added to libpng-1.0.6g, is untested. */
+int PNGAPI
+png_reset_zstream(png_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return Z_STREAM_ERROR;
+
+   /* WARNING: this resets the window bits to the maximum! */
+   return (inflateReset(&png_ptr->zstream));
+}
+#endif /* PNG_READ_SUPPORTED */
+
+/* This function was added to libpng-1.0.7 */
+png_uint_32 PNGAPI
+png_access_version_number(void)
+{
+   /* Version of *.c files used when building libpng */
+   return((png_uint_32)PNG_LIBPNG_VER);
+}
+
+
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+/* Ensure that png_ptr->zstream.msg holds some appropriate error message string.
+ * If it doesn't 'ret' is used to set it to something appropriate, even in cases
+ * like Z_OK or Z_STREAM_END where the error code is apparently a success code.
+ */
+void /* PRIVATE */
+png_zstream_error(png_structrp png_ptr, int ret)
+{
+   /* Translate 'ret' into an appropriate error string, priority is given to the
+    * one in zstream if set.  This always returns a string, even in cases like
+    * Z_OK or Z_STREAM_END where the error code is a success code.
+    */
+   if (png_ptr->zstream.msg == NULL) switch (ret)
+   {
+      default:
+      case Z_OK:
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("unexpected zlib return code");
+         break;
+
+      case Z_STREAM_END:
+         /* Normal exit */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("unexpected end of LZ stream");
+         break;
+
+      case Z_NEED_DICT:
+         /* This means the deflate stream did not have a dictionary; this
+          * indicates a bogus PNG.
+          */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("missing LZ dictionary");
+         break;
+
+      case Z_ERRNO:
+         /* gz APIs only: should not happen */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("zlib IO error");
+         break;
+
+      case Z_STREAM_ERROR:
+         /* internal libpng error */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("bad parameters to zlib");
+         break;
+
+      case Z_DATA_ERROR:
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("damaged LZ stream");
+         break;
+
+      case Z_MEM_ERROR:
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("insufficient memory");
+         break;
+
+      case Z_BUF_ERROR:
+         /* End of input or output; not a problem if the caller is doing
+          * incremental read or write.
+          */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("truncated");
+         break;
+
+      case Z_VERSION_ERROR:
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("unsupported zlib version");
+         break;
+
+      case PNG_UNEXPECTED_ZLIB_RETURN:
+         /* Compile errors here mean that zlib now uses the value co-opted in
+          * pngpriv.h for PNG_UNEXPECTED_ZLIB_RETURN; update the switch above
+          * and change pngpriv.h.  Note that this message is "... return",
+          * whereas the default/Z_OK one is "... return code".
+          */
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("unexpected zlib return");
+         break;
+   }
+}
+
+/* png_convert_size: a PNGAPI but no longer in png.h, so deleted
+ * at libpng 1.5.5!
+ */
+
+/* Added at libpng version 1.2.34 and 1.4.0 (moved from pngset.c) */
+#ifdef PNG_GAMMA_SUPPORTED /* always set if COLORSPACE */
+static int
+png_colorspace_check_gamma(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_fixed_point gAMA, int from)
+   /* This is called to check a new gamma value against an existing one.  The
+    * routine returns false if the new gamma value should not be written.
+    *
+    * 'from' says where the new gamma value comes from:
+    *
+    *    0: the new gamma value is the libpng estimate for an ICC profile
+    *    1: the new gamma value comes from a gAMA chunk
+    *    2: the new gamma value comes from an sRGB chunk
+    */
+{
+   png_fixed_point gtest;
+
+   if ((colorspace->flags & PNG_COLORSPACE_HAVE_GAMMA) != 0 &&
+      (!png_muldiv(&gtest, colorspace->gamma, PNG_FP_1, gAMA) ||
+      png_gamma_significant(gtest)))
+   {
+      /* Either this is an sRGB image, in which case the calculated gamma
+       * approximation should match, or this is an image with a profile and the
+       * value libpng calculates for the gamma of the profile does not match the
+       * value recorded in the file.  The former, sRGB, case is an error, the
+       * latter is just a warning.
+       */
+      if ((colorspace->flags & PNG_COLORSPACE_FROM_sRGB) != 0 || from == 2)
+      {
+         png_chunk_report(png_ptr, "gamma value does not match sRGB",
+            PNG_CHUNK_ERROR);
+         /* Do not overwrite an sRGB value */
+         return from == 2;
+      }
+
+      else /* sRGB tag not involved */
+      {
+         png_chunk_report(png_ptr, "gamma value does not match libpng estimate",
+            PNG_CHUNK_WARNING);
+         return from == 1;
+      }
+   }
+
+   return 1;
+}
+
+void /* PRIVATE */
+png_colorspace_set_gamma(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_fixed_point gAMA)
+{
+   /* Changed in libpng-1.5.4 to limit the values to ensure overflow can't
+    * occur.  Since the fixed point representation is assymetrical it is
+    * possible for 1/gamma to overflow the limit of 21474 and this means the
+    * gamma value must be at least 5/100000 and hence at most 20000.0.  For
+    * safety the limits here are a little narrower.  The values are 0.00016 to
+    * 6250.0, which are truly ridiculous gamma values (and will produce
+    * displays that are all black or all white.)
+    *
+    * In 1.6.0 this test replaces the ones in pngrutil.c, in the gAMA chunk
+    * handling code, which only required the value to be >0.
+    */
+   png_const_charp errmsg;
+
+   if (gAMA < 16 || gAMA > 625000000)
+      errmsg = "gamma value out of range";
+
+#  ifdef PNG_READ_gAMA_SUPPORTED
+      /* Allow the application to set the gamma value more than once */
+      else if ((png_ptr->mode & PNG_IS_READ_STRUCT) != 0 &&
+         (colorspace->flags & PNG_COLORSPACE_FROM_gAMA) != 0)
+         errmsg = "duplicate";
+#  endif
+
+   /* Do nothing if the colorspace is already invalid */
+   else if (colorspace->flags & PNG_COLORSPACE_INVALID)
+      return;
+
+   else
+   {
+      if (png_colorspace_check_gamma(png_ptr, colorspace, gAMA, 1/*from gAMA*/))
+      {
+         /* Store this gamma value. */
+         colorspace->gamma = gAMA;
+         colorspace->flags |=
+            (PNG_COLORSPACE_HAVE_GAMMA | PNG_COLORSPACE_FROM_gAMA);
+      }
+
+      /* At present if the check_gamma test fails the gamma of the colorspace is
+       * not updated however the colorspace is not invalidated.  This
+       * corresponds to the case where the existing gamma comes from an sRGB
+       * chunk or profile.  An error message has already been output.
+       */
+      return;
+   }
+
+   /* Error exit - errmsg has been set. */
+   colorspace->flags |= PNG_COLORSPACE_INVALID;
+   png_chunk_report(png_ptr, errmsg, PNG_CHUNK_WRITE_ERROR);
+}
+
+void /* PRIVATE */
+png_colorspace_sync_info(png_const_structrp png_ptr, png_inforp info_ptr)
+{
+   if (info_ptr->colorspace.flags & PNG_COLORSPACE_INVALID)
+   {
+      /* Everything is invalid */
+      info_ptr->valid &= ~(PNG_INFO_gAMA|PNG_INFO_cHRM|PNG_INFO_sRGB|
+         PNG_INFO_iCCP);
+
+#     ifdef PNG_COLORSPACE_SUPPORTED
+         /* Clean up the iCCP profile now if it won't be used. */
+         png_free_data(png_ptr, info_ptr, PNG_FREE_ICCP, -1/*not used*/);
+#     else
+         PNG_UNUSED(png_ptr)
+#     endif
+   }
+
+   else
+   {
+#     ifdef PNG_COLORSPACE_SUPPORTED
+         /* Leave the INFO_iCCP flag set if the pngset.c code has already set
+          * it; this allows a PNG to contain a profile which matches sRGB and
+          * yet still have that profile retrievable by the application.
+          */
+         if (info_ptr->colorspace.flags & PNG_COLORSPACE_MATCHES_sRGB)
+            info_ptr->valid |= PNG_INFO_sRGB;
+
+         else
+            info_ptr->valid &= ~PNG_INFO_sRGB;
+
+         if (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS)
+            info_ptr->valid |= PNG_INFO_cHRM;
+
+         else
+            info_ptr->valid &= ~PNG_INFO_cHRM;
+#     endif
+
+      if (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_GAMMA)
+         info_ptr->valid |= PNG_INFO_gAMA;
+
+      else
+         info_ptr->valid &= ~PNG_INFO_gAMA;
+   }
+}
+
+#ifdef PNG_READ_SUPPORTED
+void /* PRIVATE */
+png_colorspace_sync(png_const_structrp png_ptr, png_inforp info_ptr)
+{
+   if (info_ptr == NULL) /* reduce code size; check here not in the caller */
+      return;
+
+   info_ptr->colorspace = png_ptr->colorspace;
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+#endif
+#endif
+
+#ifdef PNG_COLORSPACE_SUPPORTED
+/* Added at libpng-1.5.5 to support read and write of true CIEXYZ values for
+ * cHRM, as opposed to using chromaticities.  These internal APIs return
+ * non-zero on a parameter error.  The X, Y and Z values are required to be
+ * positive and less than 1.0.
+ */
+static int
+png_xy_from_XYZ(png_xy *xy, const png_XYZ *XYZ)
+{
+   png_int_32 d, dwhite, whiteX, whiteY;
+
+   d = XYZ->red_X + XYZ->red_Y + XYZ->red_Z;
+   if (!png_muldiv(&xy->redx, XYZ->red_X, PNG_FP_1, d)) return 1;
+   if (!png_muldiv(&xy->redy, XYZ->red_Y, PNG_FP_1, d)) return 1;
+   dwhite = d;
+   whiteX = XYZ->red_X;
+   whiteY = XYZ->red_Y;
+
+   d = XYZ->green_X + XYZ->green_Y + XYZ->green_Z;
+   if (!png_muldiv(&xy->greenx, XYZ->green_X, PNG_FP_1, d)) return 1;
+   if (!png_muldiv(&xy->greeny, XYZ->green_Y, PNG_FP_1, d)) return 1;
+   dwhite += d;
+   whiteX += XYZ->green_X;
+   whiteY += XYZ->green_Y;
+
+   d = XYZ->blue_X + XYZ->blue_Y + XYZ->blue_Z;
+   if (!png_muldiv(&xy->bluex, XYZ->blue_X, PNG_FP_1, d)) return 1;
+   if (!png_muldiv(&xy->bluey, XYZ->blue_Y, PNG_FP_1, d)) return 1;
+   dwhite += d;
+   whiteX += XYZ->blue_X;
+   whiteY += XYZ->blue_Y;
+
+   /* The reference white is simply the sum of the end-point (X,Y,Z) vectors,
+    * thus:
+    */
+   if (!png_muldiv(&xy->whitex, whiteX, PNG_FP_1, dwhite)) return 1;
+   if (!png_muldiv(&xy->whitey, whiteY, PNG_FP_1, dwhite)) return 1;
+
+   return 0;
+}
+
+static int
+png_XYZ_from_xy(png_XYZ *XYZ, const png_xy *xy)
+{
+   png_fixed_point red_inverse, green_inverse, blue_scale;
+   png_fixed_point left, right, denominator;
+
+   /* Check xy and, implicitly, z.  Note that wide gamut color spaces typically
+    * have end points with 0 tristimulus values (these are impossible end
+    * points, but they are used to cover the possible colors.)
+    */
+   if (xy->redx < 0 || xy->redx > PNG_FP_1) return 1;
+   if (xy->redy < 0 || xy->redy > PNG_FP_1-xy->redx) return 1;
+   if (xy->greenx < 0 || xy->greenx > PNG_FP_1) return 1;
+   if (xy->greeny < 0 || xy->greeny > PNG_FP_1-xy->greenx) return 1;
+   if (xy->bluex < 0 || xy->bluex > PNG_FP_1) return 1;
+   if (xy->bluey < 0 || xy->bluey > PNG_FP_1-xy->bluex) return 1;
+   if (xy->whitex < 0 || xy->whitex > PNG_FP_1) return 1;
+   if (xy->whitey < 0 || xy->whitey > PNG_FP_1-xy->whitex) return 1;
+
+   /* The reverse calculation is more difficult because the original tristimulus
+    * value had 9 independent values (red,green,blue)x(X,Y,Z) however only 8
+    * derived values were recorded in the cHRM chunk;
+    * (red,green,blue,white)x(x,y).  This loses one degree of freedom and
+    * therefore an arbitrary ninth value has to be introduced to undo the
+    * original transformations.
+    *
+    * Think of the original end-points as points in (X,Y,Z) space.  The
+    * chromaticity values (c) have the property:
+    *
+    *           C
+    *   c = ---------
+    *       X + Y + Z
+    *
+    * For each c (x,y,z) from the corresponding original C (X,Y,Z).  Thus the
+    * three chromaticity values (x,y,z) for each end-point obey the
+    * relationship:
+    *
+    *   x + y + z = 1
+    *
+    * This describes the plane in (X,Y,Z) space that intersects each axis at the
+    * value 1.0; call this the chromaticity plane.  Thus the chromaticity
+    * calculation has scaled each end-point so that it is on the x+y+z=1 plane
+    * and chromaticity is the intersection of the vector from the origin to the
+    * (X,Y,Z) value with the chromaticity plane.
+    *
+    * To fully invert the chromaticity calculation we would need the three
+    * end-point scale factors, (red-scale, green-scale, blue-scale), but these
+    * were not recorded.  Instead we calculated the reference white (X,Y,Z) and
+    * recorded the chromaticity of this.  The reference white (X,Y,Z) would have
+    * given all three of the scale factors since:
+    *
+    *    color-C = color-c * color-scale
+    *    white-C = red-C + green-C + blue-C
+    *            = red-c*red-scale + green-c*green-scale + blue-c*blue-scale
+    *
+    * But cHRM records only white-x and white-y, so we have lost the white scale
+    * factor:
+    *
+    *    white-C = white-c*white-scale
+    *
+    * To handle this the inverse transformation makes an arbitrary assumption
+    * about white-scale:
+    *
+    *    Assume: white-Y = 1.0
+    *    Hence:  white-scale = 1/white-y
+    *    Or:     red-Y + green-Y + blue-Y = 1.0
+    *
+    * Notice the last statement of the assumption gives an equation in three of
+    * the nine values we want to calculate.  8 more equations come from the
+    * above routine as summarised at the top above (the chromaticity
+    * calculation):
+    *
+    *    Given: color-x = color-X / (color-X + color-Y + color-Z)
+    *    Hence: (color-x - 1)*color-X + color.x*color-Y + color.x*color-Z = 0
+    *
+    * This is 9 simultaneous equations in the 9 variables "color-C" and can be
+    * solved by Cramer's rule.  Cramer's rule requires calculating 10 9x9 matrix
+    * determinants, however this is not as bad as it seems because only 28 of
+    * the total of 90 terms in the various matrices are non-zero.  Nevertheless
+    * Cramer's rule is notoriously numerically unstable because the determinant
+    * calculation involves the difference of large, but similar, numbers.  It is
+    * difficult to be sure that the calculation is stable for real world values
+    * and it is certain that it becomes unstable where the end points are close
+    * together.
+    *
+    * So this code uses the perhaps slightly less optimal but more
+    * understandable and totally obvious approach of calculating color-scale.
+    *
+    * This algorithm depends on the precision in white-scale and that is
+    * (1/white-y), so we can immediately see that as white-y approaches 0 the
+    * accuracy inherent in the cHRM chunk drops off substantially.
+    *
+    * libpng arithmetic: a simple invertion of the above equations
+    * ------------------------------------------------------------
+    *
+    *    white_scale = 1/white-y
+    *    white-X = white-x * white-scale
+    *    white-Y = 1.0
+    *    white-Z = (1 - white-x - white-y) * white_scale
+    *
+    *    white-C = red-C + green-C + blue-C
+    *            = red-c*red-scale + green-c*green-scale + blue-c*blue-scale
+    *
+    * This gives us three equations in (red-scale,green-scale,blue-scale) where
+    * all the coefficients are now known:
+    *
+    *    red-x*red-scale + green-x*green-scale + blue-x*blue-scale
+    *       = white-x/white-y
+    *    red-y*red-scale + green-y*green-scale + blue-y*blue-scale = 1
+    *    red-z*red-scale + green-z*green-scale + blue-z*blue-scale
+    *       = (1 - white-x - white-y)/white-y
+    *
+    * In the last equation color-z is (1 - color-x - color-y) so we can add all
+    * three equations together to get an alternative third:
+    *
+    *    red-scale + green-scale + blue-scale = 1/white-y = white-scale
+    *
+    * So now we have a Cramer's rule solution where the determinants are just
+    * 3x3 - far more tractible.  Unfortunately 3x3 determinants still involve
+    * multiplication of three coefficients so we can't guarantee to avoid
+    * overflow in the libpng fixed point representation.  Using Cramer's rule in
+    * floating point is probably a good choice here, but it's not an option for
+    * fixed point.  Instead proceed to simplify the first two equations by
+    * eliminating what is likely to be the largest value, blue-scale:
+    *
+    *    blue-scale = white-scale - red-scale - green-scale
+    *
+    * Hence:
+    *
+    *    (red-x - blue-x)*red-scale + (green-x - blue-x)*green-scale =
+    *                (white-x - blue-x)*white-scale
+    *
+    *    (red-y - blue-y)*red-scale + (green-y - blue-y)*green-scale =
+    *                1 - blue-y*white-scale
+    *
+    * And now we can trivially solve for (red-scale,green-scale):
+    *
+    *    green-scale =
+    *                (white-x - blue-x)*white-scale - (red-x - blue-x)*red-scale
+    *                -----------------------------------------------------------
+    *                                  green-x - blue-x
+    *
+    *    red-scale =
+    *                1 - blue-y*white-scale - (green-y - blue-y) * green-scale
+    *                ---------------------------------------------------------
+    *                                  red-y - blue-y
+    *
+    * Hence:
+    *
+    *    red-scale =
+    *          ( (green-x - blue-x) * (white-y - blue-y) -
+    *            (green-y - blue-y) * (white-x - blue-x) ) / white-y
+    * -------------------------------------------------------------------------
+    *  (green-x - blue-x)*(red-y - blue-y)-(green-y - blue-y)*(red-x - blue-x)
+    *
+    *    green-scale =
+    *          ( (red-y - blue-y) * (white-x - blue-x) -
+    *            (red-x - blue-x) * (white-y - blue-y) ) / white-y
+    * -------------------------------------------------------------------------
+    *  (green-x - blue-x)*(red-y - blue-y)-(green-y - blue-y)*(red-x - blue-x)
+    *
+    * Accuracy:
+    * The input values have 5 decimal digits of accuracy.  The values are all in
+    * the range 0 < value < 1, so simple products are in the same range but may
+    * need up to 10 decimal digits to preserve the original precision and avoid
+    * underflow.  Because we are using a 32-bit signed representation we cannot
+    * match this; the best is a little over 9 decimal digits, less than 10.
+    *
+    * The approach used here is to preserve the maximum precision within the
+    * signed representation.  Because the red-scale calculation above uses the
+    * difference between two products of values that must be in the range -1..+1
+    * it is sufficient to divide the product by 7; ceil(100,000/32767*2).  The
+    * factor is irrelevant in the calculation because it is applied to both
+    * numerator and denominator.
+    *
+    * Note that the values of the differences of the products of the
+    * chromaticities in the above equations tend to be small, for example for
+    * the sRGB chromaticities they are:
+    *
+    * red numerator:    -0.04751
+    * green numerator:  -0.08788
+    * denominator:      -0.2241 (without white-y multiplication)
+    *
+    *  The resultant Y coefficients from the chromaticities of some widely used
+    *  color space definitions are (to 15 decimal places):
+    *
+    *  sRGB
+    *    0.212639005871510 0.715168678767756 0.072192315360734
+    *  Kodak ProPhoto
+    *    0.288071128229293 0.711843217810102 0.000085653960605
+    *  Adobe RGB
+    *    0.297344975250536 0.627363566255466 0.075291458493998
+    *  Adobe Wide Gamut RGB
+    *    0.258728243040113 0.724682314948566 0.016589442011321
+    */
+   /* By the argument, above overflow should be impossible here. The return
+    * value of 2 indicates an internal error to the caller.
+    */
+   if (!png_muldiv(&left, xy->greenx-xy->bluex, xy->redy - xy->bluey, 7))
+      return 2;
+   if (!png_muldiv(&right, xy->greeny-xy->bluey, xy->redx - xy->bluex, 7))
+      return 2;
+   denominator = left - right;
+
+   /* Now find the red numerator. */
+   if (!png_muldiv(&left, xy->greenx-xy->bluex, xy->whitey-xy->bluey, 7))
+      return 2;
+   if (!png_muldiv(&right, xy->greeny-xy->bluey, xy->whitex-xy->bluex, 7))
+      return 2;
+
+   /* Overflow is possible here and it indicates an extreme set of PNG cHRM
+    * chunk values.  This calculation actually returns the reciprocal of the
+    * scale value because this allows us to delay the multiplication of white-y
+    * into the denominator, which tends to produce a small number.
+    */
+   if (!png_muldiv(&red_inverse, xy->whitey, denominator, left-right) ||
+       red_inverse <= xy->whitey /* r+g+b scales = white scale */)
+      return 1;
+
+   /* Similarly for green_inverse: */
+   if (!png_muldiv(&left, xy->redy-xy->bluey, xy->whitex-xy->bluex, 7))
+      return 2;
+   if (!png_muldiv(&right, xy->redx-xy->bluex, xy->whitey-xy->bluey, 7))
+      return 2;
+   if (!png_muldiv(&green_inverse, xy->whitey, denominator, left-right) ||
+       green_inverse <= xy->whitey)
+      return 1;
+
+   /* And the blue scale, the checks above guarantee this can't overflow but it
+    * can still produce 0 for extreme cHRM values.
+    */
+   blue_scale = png_reciprocal(xy->whitey) - png_reciprocal(red_inverse) -
+      png_reciprocal(green_inverse);
+   if (blue_scale <= 0) return 1;
+
+
+   /* And fill in the png_XYZ: */
+   if (!png_muldiv(&XYZ->red_X, xy->redx, PNG_FP_1, red_inverse)) return 1;
+   if (!png_muldiv(&XYZ->red_Y, xy->redy, PNG_FP_1, red_inverse)) return 1;
+   if (!png_muldiv(&XYZ->red_Z, PNG_FP_1 - xy->redx - xy->redy, PNG_FP_1,
+      red_inverse))
+      return 1;
+
+   if (!png_muldiv(&XYZ->green_X, xy->greenx, PNG_FP_1, green_inverse))
+      return 1;
+   if (!png_muldiv(&XYZ->green_Y, xy->greeny, PNG_FP_1, green_inverse))
+      return 1;
+   if (!png_muldiv(&XYZ->green_Z, PNG_FP_1 - xy->greenx - xy->greeny, PNG_FP_1,
+      green_inverse))
+      return 1;
+
+   if (!png_muldiv(&XYZ->blue_X, xy->bluex, blue_scale, PNG_FP_1)) return 1;
+   if (!png_muldiv(&XYZ->blue_Y, xy->bluey, blue_scale, PNG_FP_1)) return 1;
+   if (!png_muldiv(&XYZ->blue_Z, PNG_FP_1 - xy->bluex - xy->bluey, blue_scale,
+      PNG_FP_1))
+      return 1;
+
+   return 0; /*success*/
+}
+
+static int
+png_XYZ_normalize(png_XYZ *XYZ)
+{
+   png_int_32 Y;
+
+   if (XYZ->red_Y < 0 || XYZ->green_Y < 0 || XYZ->blue_Y < 0 ||
+      XYZ->red_X < 0 || XYZ->green_X < 0 || XYZ->blue_X < 0 ||
+      XYZ->red_Z < 0 || XYZ->green_Z < 0 || XYZ->blue_Z < 0)
+      return 1;
+
+   /* Normalize by scaling so the sum of the end-point Y values is PNG_FP_1.
+    * IMPLEMENTATION NOTE: ANSI requires signed overflow not to occur, therefore
+    * relying on addition of two positive values producing a negative one is not
+    * safe.
+    */
+   Y = XYZ->red_Y;
+   if (0x7fffffff - Y < XYZ->green_X) return 1;
+   Y += XYZ->green_Y;
+   if (0x7fffffff - Y < XYZ->blue_X) return 1;
+   Y += XYZ->blue_Y;
+
+   if (Y != PNG_FP_1)
+   {
+      if (!png_muldiv(&XYZ->red_X, XYZ->red_X, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->red_Y, XYZ->red_Y, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->red_Z, XYZ->red_Z, PNG_FP_1, Y)) return 1;
+
+      if (!png_muldiv(&XYZ->green_X, XYZ->green_X, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->green_Y, XYZ->green_Y, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->green_Z, XYZ->green_Z, PNG_FP_1, Y)) return 1;
+
+      if (!png_muldiv(&XYZ->blue_X, XYZ->blue_X, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->blue_Y, XYZ->blue_Y, PNG_FP_1, Y)) return 1;
+      if (!png_muldiv(&XYZ->blue_Z, XYZ->blue_Z, PNG_FP_1, Y)) return 1;
+   }
+
+   return 0;
+}
+
+static int
+png_colorspace_endpoints_match(const png_xy *xy1, const png_xy *xy2, int delta)
+{
+   /* Allow an error of +/-0.01 (absolute value) on each chromaticity */
+   return !(PNG_OUT_OF_RANGE(xy1->whitex, xy2->whitex,delta) ||
+      PNG_OUT_OF_RANGE(xy1->whitey, xy2->whitey,delta) ||
+      PNG_OUT_OF_RANGE(xy1->redx,   xy2->redx,  delta) ||
+      PNG_OUT_OF_RANGE(xy1->redy,   xy2->redy,  delta) ||
+      PNG_OUT_OF_RANGE(xy1->greenx, xy2->greenx,delta) ||
+      PNG_OUT_OF_RANGE(xy1->greeny, xy2->greeny,delta) ||
+      PNG_OUT_OF_RANGE(xy1->bluex,  xy2->bluex, delta) ||
+      PNG_OUT_OF_RANGE(xy1->bluey,  xy2->bluey, delta));
+}
+
+/* Added in libpng-1.6.0, a different check for the validity of a set of cHRM
+ * chunk chromaticities.  Earlier checks used to simply look for the overflow
+ * condition (where the determinant of the matrix to solve for XYZ ends up zero
+ * because the chromaticity values are not all distinct.)  Despite this it is
+ * theoretically possible to produce chromaticities that are apparently valid
+ * but that rapidly degrade to invalid, potentially crashing, sets because of
+ * arithmetic inaccuracies when calculations are performed on them.  The new
+ * check is to round-trip xy -> XYZ -> xy and then check that the result is
+ * within a small percentage of the original.
+ */
+static int
+png_colorspace_check_xy(png_XYZ *XYZ, const png_xy *xy)
+{
+   int result;
+   png_xy xy_test;
+
+   /* As a side-effect this routine also returns the XYZ endpoints. */
+   result = png_XYZ_from_xy(XYZ, xy);
+   if (result) return result;
+
+   result = png_xy_from_XYZ(&xy_test, XYZ);
+   if (result) return result;
+
+   if (png_colorspace_endpoints_match(xy, &xy_test,
+      5/*actually, the math is pretty accurate*/))
+      return 0;
+
+   /* Too much slip */
+   return 1;
+}
+
+/* This is the check going the other way.  The XYZ is modified to normalize it
+ * (another side-effect) and the xy chromaticities are returned.
+ */
+static int
+png_colorspace_check_XYZ(png_xy *xy, png_XYZ *XYZ)
+{
+   int result;
+   png_XYZ XYZtemp;
+
+   result = png_XYZ_normalize(XYZ);
+   if (result) return result;
+
+   result = png_xy_from_XYZ(xy, XYZ);
+   if (result) return result;
+
+   XYZtemp = *XYZ;
+   return png_colorspace_check_xy(&XYZtemp, xy);
+}
+
+/* Used to check for an endpoint match against sRGB */
+static const png_xy sRGB_xy = /* From ITU-R BT.709-3 */
+{
+   /* color      x       y */
+   /* red   */ 64000, 33000,
+   /* green */ 30000, 60000,
+   /* blue  */ 15000,  6000,
+   /* white */ 31270, 32900
+};
+
+static int
+png_colorspace_set_xy_and_XYZ(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, const png_xy *xy, const png_XYZ *XYZ,
+   int preferred)
+{
+   if (colorspace->flags & PNG_COLORSPACE_INVALID)
+      return 0;
+
+   /* The consistency check is performed on the chromaticities; this factors out
+    * variations because of the normalization (or not) of the end point Y
+    * values.
+    */
+   if (preferred < 2 && (colorspace->flags & PNG_COLORSPACE_HAVE_ENDPOINTS))
+   {
+      /* The end points must be reasonably close to any we already have.  The
+       * following allows an error of up to +/-.001
+       */
+      if (!png_colorspace_endpoints_match(xy, &colorspace->end_points_xy, 100))
+      {
+         colorspace->flags |= PNG_COLORSPACE_INVALID;
+         png_benign_error(png_ptr, "inconsistent chromaticities");
+         return 0; /* failed */
+      }
+
+      /* Only overwrite with preferred values */
+      if (!preferred)
+         return 1; /* ok, but no change */
+   }
+
+   colorspace->end_points_xy = *xy;
+   colorspace->end_points_XYZ = *XYZ;
+   colorspace->flags |= PNG_COLORSPACE_HAVE_ENDPOINTS;
+
+   /* The end points are normally quoted to two decimal digits, so allow +/-0.01
+    * on this test.
+    */
+   if (png_colorspace_endpoints_match(xy, &sRGB_xy, 1000))
+      colorspace->flags |= PNG_COLORSPACE_ENDPOINTS_MATCH_sRGB;
+
+   else
+      colorspace->flags &= PNG_COLORSPACE_CANCEL(
+         PNG_COLORSPACE_ENDPOINTS_MATCH_sRGB);
+
+   return 2; /* ok and changed */
+}
+
+int /* PRIVATE */
+png_colorspace_set_chromaticities(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, const png_xy *xy, int preferred)
+{
+   /* We must check the end points to ensure they are reasonable - in the past
+    * color management systems have crashed as a result of getting bogus
+    * colorant values, while this isn't the fault of libpng it is the
+    * responsibility of libpng because PNG carries the bomb and libpng is in a
+    * position to protect against it.
+    */
+   png_XYZ XYZ;
+
+   switch (png_colorspace_check_xy(&XYZ, xy))
+   {
+      case 0: /* success */
+         return png_colorspace_set_xy_and_XYZ(png_ptr, colorspace, xy, &XYZ,
+            preferred);
+
+      case 1:
+         /* We can't invert the chromaticities so we can't produce value XYZ
+          * values.  Likely as not a color management system will fail too.
+          */
+         colorspace->flags |= PNG_COLORSPACE_INVALID;
+         png_benign_error(png_ptr, "invalid chromaticities");
+         break;
+
+      default:
+         /* libpng is broken; this should be a warning but if it happens we
+          * want error reports so for the moment it is an error.
+          */
+         colorspace->flags |= PNG_COLORSPACE_INVALID;
+         png_error(png_ptr, "internal error checking chromaticities");
+         break;
+   }
+
+   return 0; /* failed */
+}
+
+int /* PRIVATE */
+png_colorspace_set_endpoints(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, const png_XYZ *XYZ_in, int preferred)
+{
+   png_XYZ XYZ = *XYZ_in;
+   png_xy xy;
+
+   switch (png_colorspace_check_XYZ(&xy, &XYZ))
+   {
+      case 0:
+         return png_colorspace_set_xy_and_XYZ(png_ptr, colorspace, &xy, &XYZ,
+            preferred);
+
+      case 1:
+         /* End points are invalid. */
+         colorspace->flags |= PNG_COLORSPACE_INVALID;
+         png_benign_error(png_ptr, "invalid end points");
+         break;
+
+      default:
+         colorspace->flags |= PNG_COLORSPACE_INVALID;
+         png_error(png_ptr, "internal error checking chromaticities");
+         break;
+   }
+
+   return 0; /* failed */
+}
+
+#if defined(PNG_sRGB_SUPPORTED) || defined(PNG_iCCP_SUPPORTED)
+/* Error message generation */
+static char
+png_icc_tag_char(png_uint_32 byte)
+{
+   byte &= 0xff;
+   if (byte >= 32 && byte <= 126)
+      return (char)byte;
+   else
+      return '?';
+}
+
+static void
+png_icc_tag_name(char *name, png_uint_32 tag)
+{
+   name[0] = '\'';
+   name[1] = png_icc_tag_char(tag >> 24);
+   name[2] = png_icc_tag_char(tag >> 16);
+   name[3] = png_icc_tag_char(tag >>  8);
+   name[4] = png_icc_tag_char(tag      );
+   name[5] = '\'';
+}
+
+static int
+is_ICC_signature_char(png_alloc_size_t it)
+{
+   return it == 32 || (it >= 48 && it <= 57) || (it >= 65 && it <= 90) ||
+      (it >= 97 && it <= 122);
+}
+
+static int is_ICC_signature(png_alloc_size_t it)
+{
+   return is_ICC_signature_char(it >> 24) /* checks all the top bits */ &&
+      is_ICC_signature_char((it >> 16) & 0xff) &&
+      is_ICC_signature_char((it >> 8) & 0xff) &&
+      is_ICC_signature_char(it & 0xff);
+}
+
+static int
+png_icc_profile_error(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_charp name, png_alloc_size_t value, png_const_charp reason)
+{
+   size_t pos;
+   char message[196]; /* see below for calculation */
+
+   if (colorspace != NULL)
+      colorspace->flags |= PNG_COLORSPACE_INVALID;
+
+   pos = png_safecat(message, (sizeof message), 0, "profile '"); /* 9 chars */
+   pos = png_safecat(message, pos+79, pos, name); /* Truncate to 79 chars */
+   pos = png_safecat(message, (sizeof message), pos, "': "); /* +2 = 90 */
+   if (is_ICC_signature(value))
+   {
+      /* So 'value' is at most 4 bytes and the following cast is safe */
+      png_icc_tag_name(message+pos, (png_uint_32)value);
+      pos += 6; /* total +8; less than the else clause */
+      message[pos++] = ':';
+      message[pos++] = ' ';
+   }
+#  ifdef PNG_WARNINGS_SUPPORTED
+   else
+      {
+         char number[PNG_NUMBER_BUFFER_SIZE]; /* +24 = 114*/
+
+         pos = png_safecat(message, (sizeof message), pos,
+            png_format_number(number, number+(sizeof number),
+               PNG_NUMBER_FORMAT_x, value));
+         pos = png_safecat(message, (sizeof message), pos, "h: "); /*+2 = 116*/
+      }
+#  endif
+   /* The 'reason' is an arbitrary message, allow +79 maximum 195 */
+   pos = png_safecat(message, (sizeof message), pos, reason);
+
+   /* This is recoverable, but make it unconditionally an app_error on write to
+    * avoid writing invalid ICC profiles into PNG files.  (I.e.  we handle them
+    * on read, with a warning, but on write unless the app turns off
+    * application errors the PNG won't be written.)
+    */
+   png_chunk_report(png_ptr, message,
+      (colorspace != NULL) ? PNG_CHUNK_ERROR : PNG_CHUNK_WRITE_ERROR);
+
+   return 0;
+}
+#endif /* sRGB || iCCP */
+
+#ifdef PNG_sRGB_SUPPORTED
+int /* PRIVATE */
+png_colorspace_set_sRGB(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   int intent)
+{
+   /* sRGB sets known gamma, end points and (from the chunk) intent. */
+   /* IMPORTANT: these are not necessarily the values found in an ICC profile
+    * because ICC profiles store values adapted to a D50 environment; it is
+    * expected that the ICC profile mediaWhitePointTag will be D50, see the
+    * checks and code elsewhere to understand this better.
+    *
+    * These XYZ values, which are accurate to 5dp, produce rgb to gray
+    * coefficients of (6968,23435,2366), which are reduced (because they add up
+    * to 32769 not 32768) to (6968,23434,2366).  These are the values that
+    * libpng has traditionally used (and are the best values given the 15bit
+    * algorithm used by the rgb to gray code.)
+    */
+   static const png_XYZ sRGB_XYZ = /* D65 XYZ (*not* the D50 adapted values!) */
+   {
+      /* color      X      Y      Z */
+      /* red   */ 41239, 21264,  1933,
+      /* green */ 35758, 71517, 11919,
+      /* blue  */ 18048,  7219, 95053
+   };
+
+   /* Do nothing if the colorspace is already invalidated. */
+   if (colorspace->flags & PNG_COLORSPACE_INVALID)
+      return 0;
+
+   /* Check the intent, then check for existing settings.  It is valid for the
+    * PNG file to have cHRM or gAMA chunks along with sRGB, but the values must
+    * be consistent with the correct values.  If, however, this function is
+    * called below because an iCCP chunk matches sRGB then it is quite
+    * conceivable that an older app recorded incorrect gAMA and cHRM because of
+    * an incorrect calculation based on the values in the profile - this does
+    * *not* invalidate the profile (though it still produces an error, which can
+    * be ignored.)
+    */
+   if (intent < 0 || intent >= PNG_sRGB_INTENT_LAST)
+      return png_icc_profile_error(png_ptr, colorspace, "sRGB",
+         (unsigned)intent, "invalid sRGB rendering intent");
+
+   if ((colorspace->flags & PNG_COLORSPACE_HAVE_INTENT) != 0 &&
+      colorspace->rendering_intent != intent)
+      return png_icc_profile_error(png_ptr, colorspace, "sRGB",
+         (unsigned)intent, "inconsistent rendering intents");
+
+   if ((colorspace->flags & PNG_COLORSPACE_FROM_sRGB) != 0)
+   {
+      png_benign_error(png_ptr, "duplicate sRGB information ignored");
+      return 0;
+   }
+
+   /* If the standard sRGB cHRM chunk does not match the one from the PNG file
+    * warn but overwrite the value with the correct one.
+    */
+   if ((colorspace->flags & PNG_COLORSPACE_HAVE_ENDPOINTS) != 0 &&
+      !png_colorspace_endpoints_match(&sRGB_xy, &colorspace->end_points_xy,
+         100))
+      png_chunk_report(png_ptr, "cHRM chunk does not match sRGB",
+         PNG_CHUNK_ERROR);
+
+   /* This check is just done for the error reporting - the routine always
+    * returns true when the 'from' argument corresponds to sRGB (2).
+    */
+   (void)png_colorspace_check_gamma(png_ptr, colorspace, PNG_GAMMA_sRGB_INVERSE,
+      2/*from sRGB*/);
+
+   /* intent: bugs in GCC force 'int' to be used as the parameter type. */
+   colorspace->rendering_intent = (png_uint_16)intent;
+   colorspace->flags |= PNG_COLORSPACE_HAVE_INTENT;
+
+   /* endpoints */
+   colorspace->end_points_xy = sRGB_xy;
+   colorspace->end_points_XYZ = sRGB_XYZ;
+   colorspace->flags |=
+      (PNG_COLORSPACE_HAVE_ENDPOINTS|PNG_COLORSPACE_ENDPOINTS_MATCH_sRGB);
+
+   /* gamma */
+   colorspace->gamma = PNG_GAMMA_sRGB_INVERSE;
+   colorspace->flags |= PNG_COLORSPACE_HAVE_GAMMA;
+
+   /* Finally record that we have an sRGB profile */
+   colorspace->flags |=
+      (PNG_COLORSPACE_MATCHES_sRGB|PNG_COLORSPACE_FROM_sRGB);
+
+   return 1; /* set */
+}
+#endif /* sRGB */
+
+#ifdef PNG_iCCP_SUPPORTED
+/* Encoded value of D50 as an ICC XYZNumber.  From the ICC 2010 spec the value
+ * is XYZ(0.9642,1.0,0.8249), which scales to:
+ *
+ *    (63189.8112, 65536, 54060.6464)
+ */
+static const png_byte D50_nCIEXYZ[12] =
+   { 0x00, 0x00, 0xf6, 0xd6, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0xd3, 0x2d };
+
+int /* PRIVATE */
+png_icc_check_length(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_charp name, png_uint_32 profile_length)
+{
+   if (profile_length < 132)
+      return png_icc_profile_error(png_ptr, colorspace, name, profile_length,
+         "too short");
+
+   if (profile_length & 3)
+      return png_icc_profile_error(png_ptr, colorspace, name, profile_length,
+         "invalid length");
+
+   return 1;
+}
+
+int /* PRIVATE */
+png_icc_check_header(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_charp name, png_uint_32 profile_length,
+   png_const_bytep profile/* first 132 bytes only */, int color_type)
+{
+   png_uint_32 temp;
+
+   /* Length check; this cannot be ignored in this code because profile_length
+    * is used later to check the tag table, so even if the profile seems over
+    * long profile_length from the caller must be correct.  The caller can fix
+    * this up on read or write by just passing in the profile header length.
+    */
+   temp = png_get_uint_32(profile);
+   if (temp != profile_length)
+      return png_icc_profile_error(png_ptr, colorspace, name, temp,
+         "length does not match profile");
+
+   temp = png_get_uint_32(profile+128); /* tag count: 12 bytes/tag */
+   if (temp > 357913930 || /* (2^32-4-132)/12: maximum possible tag count */
+      profile_length < 132+12*temp) /* truncated tag table */
+      return png_icc_profile_error(png_ptr, colorspace, name, temp,
+         "tag count too large");
+
+   /* The 'intent' must be valid or we can't store it, ICC limits the intent to
+    * 16 bits.
+    */
+   temp = png_get_uint_32(profile+64);
+   if (temp >= 0xffff) /* The ICC limit */
+      return png_icc_profile_error(png_ptr, colorspace, name, temp,
+         "invalid rendering intent");
+
+   /* This is just a warning because the profile may be valid in future
+    * versions.
+    */
+   if (temp >= PNG_sRGB_INTENT_LAST)
+      (void)png_icc_profile_error(png_ptr, NULL, name, temp,
+         "intent outside defined range");
+
+   /* At this point the tag table can't be checked because it hasn't necessarily
+    * been loaded; however, various header fields can be checked.  These checks
+    * are for values permitted by the PNG spec in an ICC profile; the PNG spec
+    * restricts the profiles that can be passed in an iCCP chunk (they must be
+    * appropriate to processing PNG data!)
+    */
+
+   /* Data checks (could be skipped).  These checks must be independent of the
+    * version number; however, the version number doesn't accomodate changes in
+    * the header fields (just the known tags and the interpretation of the
+    * data.)
+    */
+   temp = png_get_uint_32(profile+36); /* signature 'ascp' */
+   if (temp != 0x61637370)
+      return png_icc_profile_error(png_ptr, colorspace, name, temp,
+         "invalid signature");
+
+   /* Currently the PCS illuminant/adopted white point (the computational
+    * white point) are required to be D50,
+    * however the profile contains a record of the illuminant so perhaps ICC
+    * expects to be able to change this in the future (despite the rationale in
+    * the introduction for using a fixed PCS adopted white.)  Consequently the
+    * following is just a warning.
+    */
+   if (memcmp(profile+68, D50_nCIEXYZ, 12) != 0)
+      (void)png_icc_profile_error(png_ptr, NULL, name, 0/*no tag value*/,
+         "PCS illuminant is not D50");
+
+   /* The PNG spec requires this:
+    * "If the iCCP chunk is present, the image samples conform to the colour
+    * space represented by the embedded ICC profile as defined by the
+    * International Color Consortium [ICC]. The colour space of the ICC profile
+    * shall be an RGB colour space for colour images (PNG colour types 2, 3, and
+    * 6), or a greyscale colour space for greyscale images (PNG colour types 0
+    * and 4)."
+    *
+    * This checking code ensures the embedded profile (on either read or write)
+    * conforms to the specification requirements.  Notice that an ICC 'gray'
+    * color-space profile contains the information to transform the monochrome
+    * data to XYZ or L*a*b (according to which PCS the profile uses) and this
+    * should be used in preference to the standard libpng K channel replication
+    * into R, G and B channels.
+    *
+    * Previously it was suggested that an RGB profile on grayscale data could be
+    * handled.  However it it is clear that using an RGB profile in this context
+    * must be an error - there is no specification of what it means.  Thus it is
+    * almost certainly more correct to ignore the profile.
+    */
+   temp = png_get_uint_32(profile+16); /* data colour space field */
+   switch (temp)
+   {
+      case 0x52474220: /* 'RGB ' */
+         if (!(color_type & PNG_COLOR_MASK_COLOR))
+            return png_icc_profile_error(png_ptr, colorspace, name, temp,
+               "RGB color space not permitted on grayscale PNG");
+         break;
+
+      case 0x47524159: /* 'GRAY' */
+         if (color_type & PNG_COLOR_MASK_COLOR)
+            return png_icc_profile_error(png_ptr, colorspace, name, temp,
+               "Gray color space not permitted on RGB PNG");
+         break;
+
+      default:
+         return png_icc_profile_error(png_ptr, colorspace, name, temp,
+            "invalid ICC profile color space");
+   }
+
+   /* It is up to the application to check that the profile class matches the
+    * application requirements; the spec provides no guidance, but it's pretty
+    * weird if the profile is not scanner ('scnr'), monitor ('mntr'), printer
+    * ('prtr') or 'spac' (for generic color spaces).  Issue a warning in these
+    * cases.  Issue an error for device link or abstract profiles - these don't
+    * contain the records necessary to transform the color-space to anything
+    * other than the target device (and not even that for an abstract profile).
+    * Profiles of these classes may not be embedded in images.
+    */
+   temp = png_get_uint_32(profile+12); /* profile/device class */
+   switch (temp)
+   {
+      case 0x73636E72: /* 'scnr' */
+      case 0x6D6E7472: /* 'mntr' */
+      case 0x70727472: /* 'prtr' */
+      case 0x73706163: /* 'spac' */
+         /* All supported */
+         break;
+
+      case 0x61627374: /* 'abst' */
+         /* May not be embedded in an image */
+         return png_icc_profile_error(png_ptr, colorspace, name, temp,
+            "invalid embedded Abstract ICC profile");
+
+      case 0x6C696E6B: /* 'link' */
+         /* DeviceLink profiles cannnot be interpreted in a non-device specific
+          * fashion, if an app uses the AToB0Tag in the profile the results are
+          * undefined unless the result is sent to the intended device,
+          * therefore a DeviceLink profile should not be found embedded in a
+          * PNG.
+          */
+         return png_icc_profile_error(png_ptr, colorspace, name, temp,
+            "unexpected DeviceLink ICC profile class");
+
+      case 0x6E6D636C: /* 'nmcl' */
+         /* A NamedColor profile is also device specific, however it doesn't
+          * contain an AToB0 tag that is open to misintrepretation.  Almost
+          * certainly it will fail the tests below.
+          */
+         (void)png_icc_profile_error(png_ptr, NULL, name, temp,
+            "unexpected NamedColor ICC profile class");
+         break;
+
+      default:
+         /* To allow for future enhancements to the profile accept unrecognized
+          * profile classes with a warning, these then hit the test below on the
+          * tag content to ensure they are backward compatible with one of the
+          * understood profiles.
+          */
+         (void)png_icc_profile_error(png_ptr, NULL, name, temp,
+            "unrecognized ICC profile class");
+         break;
+   }
+
+   /* For any profile other than a device link one the PCS must be encoded
+    * either in XYZ or Lab.
+    */
+   temp = png_get_uint_32(profile+20);
+   switch (temp)
+   {
+      case 0x58595A20: /* 'XYZ ' */
+      case 0x4C616220: /* 'Lab ' */
+         break;
+
+      default:
+         return png_icc_profile_error(png_ptr, colorspace, name, temp,
+            "unexpected ICC PCS encoding");
+   }
+
+   return 1;
+}
+
+int /* PRIVATE */
+png_icc_check_tag_table(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_charp name, png_uint_32 profile_length,
+   png_const_bytep profile /* header plus whole tag table */)
+{
+   png_uint_32 tag_count = png_get_uint_32(profile+128);
+   png_uint_32 itag;
+   png_const_bytep tag = profile+132; /* The first tag */
+
+   /* First scan all the tags in the table and add bits to the icc_info value
+    * (temporarily in 'tags').
+    */
+   for (itag=0; itag < tag_count; ++itag, tag += 12)
+   {
+      png_uint_32 tag_id = png_get_uint_32(tag+0);
+      png_uint_32 tag_start = png_get_uint_32(tag+4); /* must be aligned */
+      png_uint_32 tag_length = png_get_uint_32(tag+8);/* not padded */
+
+      /* The ICC specification does not exclude zero length tags, therefore the
+       * start might actually be anywhere if there is no data, but this would be
+       * a clear abuse of the intent of the standard so the start is checked for
+       * being in range.  All defined tag types have an 8 byte header - a 4 byte
+       * type signature then 0.
+       */
+      if ((tag_start & 3) != 0)
+      {
+         /* CNHP730S.icc shipped with Microsoft Windows 64 violates this, it is
+          * only a warning here because libpng does not care about the
+          * alignment.
+          */
+         (void)png_icc_profile_error(png_ptr, NULL, name, tag_id,
+            "ICC profile tag start not a multiple of 4");
+      }
+
+      /* This is a hard error; potentially it can cause read outside the
+       * profile.
+       */
+      if (tag_start > profile_length || tag_length > profile_length - tag_start)
+         return png_icc_profile_error(png_ptr, colorspace, name, tag_id,
+            "ICC profile tag outside profile");
+   }
+
+   return 1; /* success, maybe with warnings */
+}
+
+#ifdef PNG_sRGB_SUPPORTED
+/* Information about the known ICC sRGB profiles */
+static const struct
+{
+   png_uint_32 adler, crc, length;
+   png_uint_32 md5[4];
+   png_byte    have_md5;
+   png_byte    is_broken;
+   png_uint_16 intent;
+
+#  define PNG_MD5(a,b,c,d) { a, b, c, d }, (a!=0)||(b!=0)||(c!=0)||(d!=0)
+#  define PNG_ICC_CHECKSUM(adler, crc, md5, intent, broke, date, length, fname)\
+      { adler, crc, length, md5, broke, intent },
+
+} png_sRGB_checks[] =
+{
+   /* This data comes from contrib/tools/checksum-icc run on downloads of
+    * all four ICC sRGB profiles from www.color.org.
+    */
+   /* adler32, crc32, MD5[4], intent, date, length, file-name */
+   PNG_ICC_CHECKSUM(0x0a3fd9f6, 0x3b8772b9,
+      PNG_MD5(0x29f83dde, 0xaff255ae, 0x7842fae4, 0xca83390d), 0, 0,
+      "2009/03/27 21:36:31", 3048, "sRGB_IEC61966-2-1_black_scaled.icc")
+
+   /* ICC sRGB v2 perceptual no black-compensation: */
+   PNG_ICC_CHECKSUM(0x4909e5e1, 0x427ebb21,
+      PNG_MD5(0xc95bd637, 0xe95d8a3b, 0x0df38f99, 0xc1320389), 1, 0,
+      "2009/03/27 21:37:45", 3052, "sRGB_IEC61966-2-1_no_black_scaling.icc")
+
+   PNG_ICC_CHECKSUM(0xfd2144a1, 0x306fd8ae,
+      PNG_MD5(0xfc663378, 0x37e2886b, 0xfd72e983, 0x8228f1b8), 0, 0,
+      "2009/08/10 17:28:01", 60988, "sRGB_v4_ICC_preference_displayclass.icc")
+
+   /* ICC sRGB v4 perceptual */
+   PNG_ICC_CHECKSUM(0x209c35d2, 0xbbef7812,
+      PNG_MD5(0x34562abf, 0x994ccd06, 0x6d2c5721, 0xd0d68c5d), 0, 0,
+      "2007/07/25 00:05:37", 60960, "sRGB_v4_ICC_preference.icc")
+
+   /* The following profiles have no known MD5 checksum. If there is a match
+    * on the (empty) MD5 the other fields are used to attempt a match and
+    * a warning is produced.  The first two of these profiles have a 'cprt' tag
+    * which suggests that they were also made by Hewlett Packard.
+    */
+   PNG_ICC_CHECKSUM(0xa054d762, 0x5d5129ce,
+      PNG_MD5(0x00000000, 0x00000000, 0x00000000, 0x00000000), 1, 0,
+      "2004/07/21 18:57:42", 3024, "sRGB_IEC61966-2-1_noBPC.icc")
+
+   /* This is a 'mntr' (display) profile with a mediaWhitePointTag that does not
+    * match the D50 PCS illuminant in the header (it is in fact the D65 values,
+    * so the white point is recorded as the un-adapted value.)  The profiles
+    * below only differ in one byte - the intent - and are basically the same as
+    * the previous profile except for the mediaWhitePointTag error and a missing
+    * chromaticAdaptationTag.
+    */
+   PNG_ICC_CHECKSUM(0xf784f3fb, 0x182ea552,
+      PNG_MD5(0x00000000, 0x00000000, 0x00000000, 0x00000000), 0, 1/*broken*/,
+      "1998/02/09 06:49:00", 3144, "HP-Microsoft sRGB v2 perceptual")
+
+   PNG_ICC_CHECKSUM(0x0398f3fc, 0xf29e526d,
+      PNG_MD5(0x00000000, 0x00000000, 0x00000000, 0x00000000), 1, 1/*broken*/,
+      "1998/02/09 06:49:00", 3144, "HP-Microsoft sRGB v2 media-relative")
+};
+
+static int
+png_compare_ICC_profile_with_sRGB(png_const_structrp png_ptr,
+   png_const_bytep profile, uLong adler)
+{
+   /* The quick check is to verify just the MD5 signature and trust the
+    * rest of the data.  Because the profile has already been verified for
+    * correctness this is safe.  png_colorspace_set_sRGB will check the 'intent'
+    * field too, so if the profile has been edited with an intent not defined
+    * by sRGB (but maybe defined by a later ICC specification) the read of
+    * the profile will fail at that point.
+    */
+   png_uint_32 length = 0;
+   png_uint_32 intent = 0x10000; /* invalid */
+#if PNG_sRGB_PROFILE_CHECKS > 1
+   uLong crc = 0; /* the value for 0 length data */
+#endif
+   unsigned int i;
+
+   for (i=0; i < (sizeof png_sRGB_checks) / (sizeof png_sRGB_checks[0]); ++i)
+   {
+      if (png_get_uint_32(profile+84) == png_sRGB_checks[i].md5[0] &&
+         png_get_uint_32(profile+88) == png_sRGB_checks[i].md5[1] &&
+         png_get_uint_32(profile+92) == png_sRGB_checks[i].md5[2] &&
+         png_get_uint_32(profile+96) == png_sRGB_checks[i].md5[3])
+      {
+         /* This may be one of the old HP profiles without an MD5, in that
+          * case we can only use the length and Adler32 (note that these
+          * are not used by default if there is an MD5!)
+          */
+#        if PNG_sRGB_PROFILE_CHECKS == 0
+            if (png_sRGB_checks[i].have_md5)
+               return 1+png_sRGB_checks[i].is_broken;
+#        endif
+
+         /* Profile is unsigned or more checks have been configured in. */
+         if (length == 0)
+         {
+            length = png_get_uint_32(profile);
+            intent = png_get_uint_32(profile+64);
+         }
+
+         /* Length *and* intent must match */
+         if (length == png_sRGB_checks[i].length &&
+            intent == png_sRGB_checks[i].intent)
+         {
+            /* Now calculate the adler32 if not done already. */
+            if (adler == 0)
+            {
+               adler = adler32(0, NULL, 0);
+               adler = adler32(adler, profile, length);
+            }
+
+            if (adler == png_sRGB_checks[i].adler)
+            {
+               /* These basic checks suggest that the data has not been
+                * modified, but if the check level is more than 1 perform
+                * our own crc32 checksum on the data.
+                */
+#              if PNG_sRGB_PROFILE_CHECKS > 1
+                  if (crc == 0)
+                  {
+                     crc = crc32(0, NULL, 0);
+                     crc = crc32(crc, profile, length);
+                  }
+
+                  /* So this check must pass for the 'return' below to happen.
+                   */
+                  if (crc == png_sRGB_checks[i].crc)
+#              endif
+               {
+                  if (png_sRGB_checks[i].is_broken)
+                  {
+                     /* These profiles are known to have bad data that may cause
+                      * problems if they are used, therefore attempt to
+                      * discourage their use, skip the 'have_md5' warning below,
+                      * which is made irrelevant by this error.
+                      */
+                     png_chunk_report(png_ptr, "known incorrect sRGB profile",
+                        PNG_CHUNK_ERROR);
+                  }
+
+                  /* Warn that this being done; this isn't even an error since
+                   * the profile is perfectly valid, but it would be nice if
+                   * people used the up-to-date ones.
+                   */
+                  else if (!png_sRGB_checks[i].have_md5)
+                  {
+                     png_chunk_report(png_ptr,
+                        "out-of-date sRGB profile with no signature",
+                        PNG_CHUNK_WARNING);
+                  }
+
+                  return 1+png_sRGB_checks[i].is_broken;
+               }
+            }
+         }
+
+#        if PNG_sRGB_PROFILE_CHECKS > 0
+            /* The signature matched, but the profile had been changed in some
+             * way.  This is an apparent violation of the ICC terms of use and,
+             * anyway, probably indicates a data error or uninformed hacking.
+             */
+            if (png_sRGB_checks[i].have_md5)
+               png_benign_error(png_ptr,
+                  "copyright violation: edited ICC profile ignored");
+#        endif
+      }
+   }
+
+   return 0; /* no match */
+}
+#endif
+
+#ifdef PNG_sRGB_SUPPORTED
+void /* PRIVATE */
+png_icc_set_sRGB(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_const_bytep profile, uLong adler)
+{
+   /* Is this profile one of the known ICC sRGB profiles?  If it is, just set
+    * the sRGB information.
+    */
+   if (png_compare_ICC_profile_with_sRGB(png_ptr, profile, adler))
+      (void)png_colorspace_set_sRGB(png_ptr, colorspace,
+         (int)/*already checked*/png_get_uint_32(profile+64));
+}
+#endif /* PNG_READ_sRGB_SUPPORTED */
+
+int /* PRIVATE */
+png_colorspace_set_ICC(png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_charp name, png_uint_32 profile_length, png_const_bytep profile,
+   int color_type)
+{
+   if (colorspace->flags & PNG_COLORSPACE_INVALID)
+      return 0;
+
+   if (png_icc_check_length(png_ptr, colorspace, name, profile_length) &&
+      png_icc_check_header(png_ptr, colorspace, name, profile_length, profile,
+         color_type) &&
+      png_icc_check_tag_table(png_ptr, colorspace, name, profile_length,
+         profile))
+   {
+#     ifdef PNG_sRGB_SUPPORTED
+         /* If no sRGB support, don't try storing sRGB information */
+         png_icc_set_sRGB(png_ptr, colorspace, profile, 0);
+#     endif
+      return 1;
+   }
+
+   /* Failure case */
+   return 0;
+}
+#endif /* iCCP */
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+void /* PRIVATE */
+png_colorspace_set_rgb_coefficients(png_structrp png_ptr)
+{
+   /* Set the rgb_to_gray coefficients from the colorspace. */
+   if (!png_ptr->rgb_to_gray_coefficients_set &&
+      (png_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS) != 0)
+   {
+      /* png_set_background has not been called, get the coefficients from the Y
+       * values of the colorspace colorants.
+       */
+      png_fixed_point r = png_ptr->colorspace.end_points_XYZ.red_Y;
+      png_fixed_point g = png_ptr->colorspace.end_points_XYZ.green_Y;
+      png_fixed_point b = png_ptr->colorspace.end_points_XYZ.blue_Y;
+      png_fixed_point total = r+g+b;
+
+      if (total > 0 &&
+         r >= 0 && png_muldiv(&r, r, 32768, total) && r >= 0 && r <= 32768 &&
+         g >= 0 && png_muldiv(&g, g, 32768, total) && g >= 0 && g <= 32768 &&
+         b >= 0 && png_muldiv(&b, b, 32768, total) && b >= 0 && b <= 32768 &&
+         r+g+b <= 32769)
+      {
+         /* We allow 0 coefficients here.  r+g+b may be 32769 if two or
+          * all of the coefficients were rounded up.  Handle this by
+          * reducing the *largest* coefficient by 1; this matches the
+          * approach used for the default coefficients in pngrtran.c
+          */
+         int add = 0;
+
+         if (r+g+b > 32768)
+            add = -1;
+         else if (r+g+b < 32768)
+            add = 1;
+
+         if (add != 0)
+         {
+            if (g >= r && g >= b)
+               g += add;
+            else if (r >= g && r >= b)
+               r += add;
+            else
+               b += add;
+         }
+
+         /* Check for an internal error. */
+         if (r+g+b != 32768)
+            png_error(png_ptr,
+               "internal error handling cHRM coefficients");
+
+         else
+         {
+            png_ptr->rgb_to_gray_red_coeff   = (png_uint_16)r;
+            png_ptr->rgb_to_gray_green_coeff = (png_uint_16)g;
+         }
+      }
+
+      /* This is a png_error at present even though it could be ignored -
+       * it should never happen, but it is important that if it does, the
+       * bug is fixed.
+       */
+      else
+         png_error(png_ptr, "internal error handling cHRM->XYZ");
+   }
+}
+#endif
+
+#endif /* COLORSPACE */
+
+void /* PRIVATE */
+png_check_IHDR(png_const_structrp png_ptr,
+   png_uint_32 width, png_uint_32 height, int bit_depth,
+   int color_type, int interlace_type, int compression_type,
+   int filter_type)
+{
+   int error = 0;
+
+   /* Check for width and height valid values */
+   if (width == 0)
+   {
+      png_warning(png_ptr, "Image width is zero in IHDR");
+      error = 1;
+   }
+
+   if (height == 0)
+   {
+      png_warning(png_ptr, "Image height is zero in IHDR");
+      error = 1;
+   }
+
+#  ifdef PNG_SET_USER_LIMITS_SUPPORTED
+   if (width > png_ptr->user_width_max)
+
+#  else
+   if (width > PNG_USER_WIDTH_MAX)
+#  endif
+   {
+      png_warning(png_ptr, "Image width exceeds user limit in IHDR");
+      error = 1;
+   }
+
+#  ifdef PNG_SET_USER_LIMITS_SUPPORTED
+   if (height > png_ptr->user_height_max)
+#  else
+   if (height > PNG_USER_HEIGHT_MAX)
+#  endif
+   {
+      png_warning(png_ptr, "Image height exceeds user limit in IHDR");
+      error = 1;
+   }
+
+   if (width > PNG_UINT_31_MAX)
+   {
+      png_warning(png_ptr, "Invalid image width in IHDR");
+      error = 1;
+   }
+
+   if (height > PNG_UINT_31_MAX)
+   {
+      png_warning(png_ptr, "Invalid image height in IHDR");
+      error = 1;
+   }
+
+   if (width > (PNG_UINT_32_MAX
+                 >> 3)      /* 8-byte RGBA pixels */
+                 - 48       /* bigrowbuf hack */
+                 - 1        /* filter byte */
+                 - 7*8      /* rounding of width to multiple of 8 pixels */
+                 - 8)       /* extra max_pixel_depth pad */
+      png_warning(png_ptr, "Width is too large for libpng to process pixels");
+
+   /* Check other values */
+   if (bit_depth != 1 && bit_depth != 2 && bit_depth != 4 &&
+       bit_depth != 8 && bit_depth != 16)
+   {
+      png_warning(png_ptr, "Invalid bit depth in IHDR");
+      error = 1;
+   }
+
+   if (color_type < 0 || color_type == 1 ||
+       color_type == 5 || color_type > 6)
+   {
+      png_warning(png_ptr, "Invalid color type in IHDR");
+      error = 1;
+   }
+
+   if (((color_type == PNG_COLOR_TYPE_PALETTE) && bit_depth > 8) ||
+       ((color_type == PNG_COLOR_TYPE_RGB ||
+         color_type == PNG_COLOR_TYPE_GRAY_ALPHA ||
+         color_type == PNG_COLOR_TYPE_RGB_ALPHA) && bit_depth < 8))
+   {
+      png_warning(png_ptr, "Invalid color type/bit depth combination in IHDR");
+      error = 1;
+   }
+
+   if (interlace_type >= PNG_INTERLACE_LAST)
+   {
+      png_warning(png_ptr, "Unknown interlace method in IHDR");
+      error = 1;
+   }
+
+   if (compression_type != PNG_COMPRESSION_TYPE_BASE)
+   {
+      png_warning(png_ptr, "Unknown compression method in IHDR");
+      error = 1;
+   }
+
+#  ifdef PNG_MNG_FEATURES_SUPPORTED
+   /* Accept filter_method 64 (intrapixel differencing) only if
+    * 1. Libpng was compiled with PNG_MNG_FEATURES_SUPPORTED and
+    * 2. Libpng did not read a PNG signature (this filter_method is only
+    *    used in PNG datastreams that are embedded in MNG datastreams) and
+    * 3. The application called png_permit_mng_features with a mask that
+    *    included PNG_FLAG_MNG_FILTER_64 and
+    * 4. The filter_method is 64 and
+    * 5. The color_type is RGB or RGBA
+    */
+   if ((png_ptr->mode & PNG_HAVE_PNG_SIGNATURE) &&
+       png_ptr->mng_features_permitted)
+      png_warning(png_ptr, "MNG features are not allowed in a PNG datastream");
+
+   if (filter_type != PNG_FILTER_TYPE_BASE)
+   {
+      if (!((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) &&
+          (filter_type == PNG_INTRAPIXEL_DIFFERENCING) &&
+          ((png_ptr->mode & PNG_HAVE_PNG_SIGNATURE) == 0) &&
+          (color_type == PNG_COLOR_TYPE_RGB ||
+          color_type == PNG_COLOR_TYPE_RGB_ALPHA)))
+      {
+         png_warning(png_ptr, "Unknown filter method in IHDR");
+         error = 1;
+      }
+
+      if (png_ptr->mode & PNG_HAVE_PNG_SIGNATURE)
+      {
+         png_warning(png_ptr, "Invalid filter method in IHDR");
+         error = 1;
+      }
+   }
+
+#  else
+   if (filter_type != PNG_FILTER_TYPE_BASE)
+   {
+      png_warning(png_ptr, "Unknown filter method in IHDR");
+      error = 1;
+   }
+#  endif
+
+   if (error == 1)
+      png_error(png_ptr, "Invalid IHDR data");
+}
+
+#if defined(PNG_sCAL_SUPPORTED) || defined(PNG_pCAL_SUPPORTED)
+/* ASCII to fp functions */
+/* Check an ASCII formated floating point value, see the more detailed
+ * comments in pngpriv.h
+ */
+/* The following is used internally to preserve the sticky flags */
+#define png_fp_add(state, flags) ((state) |= (flags))
+#define png_fp_set(state, value) ((state) = (value) | ((state) & PNG_FP_STICKY))
+
+int /* PRIVATE */
+png_check_fp_number(png_const_charp string, png_size_t size, int *statep,
+   png_size_tp whereami)
+{
+   int state = *statep;
+   png_size_t i = *whereami;
+
+   while (i < size)
+   {
+      int type;
+      /* First find the type of the next character */
+      switch (string[i])
+      {
+      case 43:  type = PNG_FP_SAW_SIGN;                   break;
+      case 45:  type = PNG_FP_SAW_SIGN + PNG_FP_NEGATIVE; break;
+      case 46:  type = PNG_FP_SAW_DOT;                    break;
+      case 48:  type = PNG_FP_SAW_DIGIT;                  break;
+      case 49: case 50: case 51: case 52:
+      case 53: case 54: case 55: case 56:
+      case 57:  type = PNG_FP_SAW_DIGIT + PNG_FP_NONZERO; break;
+      case 69:
+      case 101: type = PNG_FP_SAW_E;                      break;
+      default:  goto PNG_FP_End;
+      }
+
+      /* Now deal with this type according to the current
+       * state, the type is arranged to not overlap the
+       * bits of the PNG_FP_STATE.
+       */
+      switch ((state & PNG_FP_STATE) + (type & PNG_FP_SAW_ANY))
+      {
+      case PNG_FP_INTEGER + PNG_FP_SAW_SIGN:
+         if (state & PNG_FP_SAW_ANY)
+            goto PNG_FP_End; /* not a part of the number */
+
+         png_fp_add(state, type);
+         break;
+
+      case PNG_FP_INTEGER + PNG_FP_SAW_DOT:
+         /* Ok as trailer, ok as lead of fraction. */
+         if (state & PNG_FP_SAW_DOT) /* two dots */
+            goto PNG_FP_End;
+
+         else if (state & PNG_FP_SAW_DIGIT) /* trailing dot? */
+            png_fp_add(state, type);
+
+         else
+            png_fp_set(state, PNG_FP_FRACTION | type);
+
+         break;
+
+      case PNG_FP_INTEGER + PNG_FP_SAW_DIGIT:
+         if (state & PNG_FP_SAW_DOT) /* delayed fraction */
+            png_fp_set(state, PNG_FP_FRACTION | PNG_FP_SAW_DOT);
+
+         png_fp_add(state, type | PNG_FP_WAS_VALID);
+
+         break;
+
+      case PNG_FP_INTEGER + PNG_FP_SAW_E:
+         if ((state & PNG_FP_SAW_DIGIT) == 0)
+            goto PNG_FP_End;
+
+         png_fp_set(state, PNG_FP_EXPONENT);
+
+         break;
+
+   /* case PNG_FP_FRACTION + PNG_FP_SAW_SIGN:
+         goto PNG_FP_End; ** no sign in fraction */
+
+   /* case PNG_FP_FRACTION + PNG_FP_SAW_DOT:
+         goto PNG_FP_End; ** Because SAW_DOT is always set */
+
+      case PNG_FP_FRACTION + PNG_FP_SAW_DIGIT:
+         png_fp_add(state, type | PNG_FP_WAS_VALID);
+         break;
+
+      case PNG_FP_FRACTION + PNG_FP_SAW_E:
+         /* This is correct because the trailing '.' on an
+          * integer is handled above - so we can only get here
+          * with the sequence ".E" (with no preceding digits).
+          */
+         if ((state & PNG_FP_SAW_DIGIT) == 0)
+            goto PNG_FP_End;
+
+         png_fp_set(state, PNG_FP_EXPONENT);
+
+         break;
+
+      case PNG_FP_EXPONENT + PNG_FP_SAW_SIGN:
+         if (state & PNG_FP_SAW_ANY)
+            goto PNG_FP_End; /* not a part of the number */
+
+         png_fp_add(state, PNG_FP_SAW_SIGN);
+
+         break;
+
+   /* case PNG_FP_EXPONENT + PNG_FP_SAW_DOT:
+         goto PNG_FP_End; */
+
+      case PNG_FP_EXPONENT + PNG_FP_SAW_DIGIT:
+         png_fp_add(state, PNG_FP_SAW_DIGIT | PNG_FP_WAS_VALID);
+
+         break;
+
+   /* case PNG_FP_EXPONEXT + PNG_FP_SAW_E:
+         goto PNG_FP_End; */
+
+      default: goto PNG_FP_End; /* I.e. break 2 */
+      }
+
+      /* The character seems ok, continue. */
+      ++i;
+   }
+
+PNG_FP_End:
+   /* Here at the end, update the state and return the correct
+    * return code.
+    */
+   *statep = state;
+   *whereami = i;
+
+   return (state & PNG_FP_SAW_DIGIT) != 0;
+}
+
+
+/* The same but for a complete string. */
+int
+png_check_fp_string(png_const_charp string, png_size_t size)
+{
+   int        state=0;
+   png_size_t char_index=0;
+
+   if (png_check_fp_number(string, size, &state, &char_index) &&
+      (char_index == size || string[char_index] == 0))
+      return state /* must be non-zero - see above */;
+
+   return 0; /* i.e. fail */
+}
+#endif /* pCAL or sCAL */
+
+#ifdef PNG_sCAL_SUPPORTED
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+/* Utility used below - a simple accurate power of ten from an integral
+ * exponent.
+ */
+static double
+png_pow10(int power)
+{
+   int recip = 0;
+   double d = 1;
+
+   /* Handle negative exponent with a reciprocal at the end because
+    * 10 is exact whereas .1 is inexact in base 2
+    */
+   if (power < 0)
+   {
+      if (power < DBL_MIN_10_EXP) return 0;
+      recip = 1, power = -power;
+   }
+
+   if (power > 0)
+   {
+      /* Decompose power bitwise. */
+      double mult = 10;
+      do
+      {
+         if (power & 1) d *= mult;
+         mult *= mult;
+         power >>= 1;
+      }
+      while (power > 0);
+
+      if (recip) d = 1/d;
+   }
+   /* else power is 0 and d is 1 */
+
+   return d;
+}
+
+/* Function to format a floating point value in ASCII with a given
+ * precision.
+ */
+void /* PRIVATE */
+png_ascii_from_fp(png_const_structrp png_ptr, png_charp ascii, png_size_t size,
+    double fp, unsigned int precision)
+{
+   /* We use standard functions from math.h, but not printf because
+    * that would require stdio.  The caller must supply a buffer of
+    * sufficient size or we will png_error.  The tests on size and
+    * the space in ascii[] consumed are indicated below.
+    */
+   if (precision < 1)
+      precision = DBL_DIG;
+
+   /* Enforce the limit of the implementation precision too. */
+   if (precision > DBL_DIG+1)
+      precision = DBL_DIG+1;
+
+   /* Basic sanity checks */
+   if (size >= precision+5) /* See the requirements below. */
+   {
+      if (fp < 0)
+      {
+         fp = -fp;
+         *ascii++ = 45; /* '-'  PLUS 1 TOTAL 1 */
+         --size;
+      }
+
+      if (fp >= DBL_MIN && fp <= DBL_MAX)
+      {
+         int exp_b10;       /* A base 10 exponent */
+         double base;   /* 10^exp_b10 */
+
+         /* First extract a base 10 exponent of the number,
+          * the calculation below rounds down when converting
+          * from base 2 to base 10 (multiply by log10(2) -
+          * 0.3010, but 77/256 is 0.3008, so exp_b10 needs to
+          * be increased.  Note that the arithmetic shift
+          * performs a floor() unlike C arithmetic - using a
+          * C multiply would break the following for negative
+          * exponents.
+          */
+         (void)frexp(fp, &exp_b10); /* exponent to base 2 */
+
+         exp_b10 = (exp_b10 * 77) >> 8; /* <= exponent to base 10 */
+
+         /* Avoid underflow here. */
+         base = png_pow10(exp_b10); /* May underflow */
+
+         while (base < DBL_MIN || base < fp)
+         {
+            /* And this may overflow. */
+            double test = png_pow10(exp_b10+1);
+
+            if (test <= DBL_MAX)
+               ++exp_b10, base = test;
+
+            else
+               break;
+         }
+
+         /* Normalize fp and correct exp_b10, after this fp is in the
+          * range [.1,1) and exp_b10 is both the exponent and the digit
+          * *before* which the decimal point should be inserted
+          * (starting with 0 for the first digit).  Note that this
+          * works even if 10^exp_b10 is out of range because of the
+          * test on DBL_MAX above.
+          */
+         fp /= base;
+         while (fp >= 1) fp /= 10, ++exp_b10;
+
+         /* Because of the code above fp may, at this point, be
+          * less than .1, this is ok because the code below can
+          * handle the leading zeros this generates, so no attempt
+          * is made to correct that here.
+          */
+
+         {
+            int czero, clead, cdigits;
+            char exponent[10];
+
+            /* Allow up to two leading zeros - this will not lengthen
+             * the number compared to using E-n.
+             */
+            if (exp_b10 < 0 && exp_b10 > -3) /* PLUS 3 TOTAL 4 */
+            {
+               czero = -exp_b10; /* PLUS 2 digits: TOTAL 3 */
+               exp_b10 = 0;      /* Dot added below before first output. */
+            }
+            else
+               czero = 0;    /* No zeros to add */
+
+            /* Generate the digit list, stripping trailing zeros and
+             * inserting a '.' before a digit if the exponent is 0.
+             */
+            clead = czero; /* Count of leading zeros */
+            cdigits = 0;   /* Count of digits in list. */
+
+            do
+            {
+               double d;
+
+               fp *= 10;
+               /* Use modf here, not floor and subtract, so that
+                * the separation is done in one step.  At the end
+                * of the loop don't break the number into parts so
+                * that the final digit is rounded.
+                */
+               if (cdigits+czero-clead+1 < (int)precision)
+                  fp = modf(fp, &d);
+
+               else
+               {
+                  d = floor(fp + .5);
+
+                  if (d > 9)
+                  {
+                     /* Rounding up to 10, handle that here. */
+                     if (czero > 0)
+                     {
+                        --czero, d = 1;
+                        if (cdigits == 0) --clead;
+                     }
+                     else
+                     {
+                        while (cdigits > 0 && d > 9)
+                        {
+                           int ch = *--ascii;
+
+                           if (exp_b10 != (-1))
+                              ++exp_b10;
+
+                           else if (ch == 46)
+                           {
+                              ch = *--ascii, ++size;
+                              /* Advance exp_b10 to '1', so that the
+                               * decimal point happens after the
+                               * previous digit.
+                               */
+                              exp_b10 = 1;
+                           }
+
+                           --cdigits;
+                           d = ch - 47;  /* I.e. 1+(ch-48) */
+                        }
+
+                        /* Did we reach the beginning? If so adjust the
+                         * exponent but take into account the leading
+                         * decimal point.
+                         */
+                        if (d > 9)  /* cdigits == 0 */
+                        {
+                           if (exp_b10 == (-1))
+                           {
+                              /* Leading decimal point (plus zeros?), if
+                               * we lose the decimal point here it must
+                               * be reentered below.
+                               */
+                              int ch = *--ascii;
+
+                              if (ch == 46)
+                                 ++size, exp_b10 = 1;
+
+                              /* Else lost a leading zero, so 'exp_b10' is
+                               * still ok at (-1)
+                               */
+                           }
+                           else
+                              ++exp_b10;
+
+                           /* In all cases we output a '1' */
+                           d = 1;
+                        }
+                     }
+                  }
+                  fp = 0; /* Guarantees termination below. */
+               }
+
+               if (d == 0)
+               {
+                  ++czero;
+                  if (cdigits == 0) ++clead;
+               }
+               else
+               {
+                  /* Included embedded zeros in the digit count. */
+                  cdigits += czero - clead;
+                  clead = 0;
+
+                  while (czero > 0)
+                  {
+                     /* exp_b10 == (-1) means we just output the decimal
+                      * place - after the DP don't adjust 'exp_b10' any
+                      * more!
+                      */
+                     if (exp_b10 != (-1))
+                     {
+                        if (exp_b10 == 0) *ascii++ = 46, --size;
+                        /* PLUS 1: TOTAL 4 */
+                        --exp_b10;
+                     }
+                     *ascii++ = 48, --czero;
+                  }
+
+                  if (exp_b10 != (-1))
+                  {
+                     if (exp_b10 == 0) *ascii++ = 46, --size; /* counted
+                                                                 above */
+                     --exp_b10;
+                  }
+                  *ascii++ = (char)(48 + (int)d), ++cdigits;
+               }
+            }
+            while (cdigits+czero-clead < (int)precision && fp > DBL_MIN);
+
+            /* The total output count (max) is now 4+precision */
+
+            /* Check for an exponent, if we don't need one we are
+             * done and just need to terminate the string.  At
+             * this point exp_b10==(-1) is effectively if flag - it got
+             * to '-1' because of the decrement after outputing
+             * the decimal point above (the exponent required is
+             * *not* -1!)
+             */
+            if (exp_b10 >= (-1) && exp_b10 <= 2)
+            {
+               /* The following only happens if we didn't output the
+                * leading zeros above for negative exponent, so this
+                * doest add to the digit requirement.  Note that the
+                * two zeros here can only be output if the two leading
+                * zeros were *not* output, so this doesn't increase
+                * the output count.
+                */
+               while (--exp_b10 >= 0) *ascii++ = 48;
+
+               *ascii = 0;
+
+               /* Total buffer requirement (including the '\0') is
+                * 5+precision - see check at the start.
+                */
+               return;
+            }
+
+            /* Here if an exponent is required, adjust size for
+             * the digits we output but did not count.  The total
+             * digit output here so far is at most 1+precision - no
+             * decimal point and no leading or trailing zeros have
+             * been output.
+             */
+            size -= cdigits;
+
+            *ascii++ = 69, --size;    /* 'E': PLUS 1 TOTAL 2+precision */
+
+            /* The following use of an unsigned temporary avoids ambiguities in
+             * the signed arithmetic on exp_b10 and permits GCC at least to do
+             * better optimization.
+             */
+            {
+               unsigned int uexp_b10;
+
+               if (exp_b10 < 0)
+               {
+                  *ascii++ = 45, --size; /* '-': PLUS 1 TOTAL 3+precision */
+                  uexp_b10 = -exp_b10;
+               }
+
+               else
+                  uexp_b10 = exp_b10;
+
+               cdigits = 0;
+
+               while (uexp_b10 > 0)
+               {
+                  exponent[cdigits++] = (char)(48 + uexp_b10 % 10);
+                  uexp_b10 /= 10;
+               }
+            }
+
+            /* Need another size check here for the exponent digits, so
+             * this need not be considered above.
+             */
+            if ((int)size > cdigits)
+            {
+               while (cdigits > 0) *ascii++ = exponent[--cdigits];
+
+               *ascii = 0;
+
+               return;
+            }
+         }
+      }
+      else if (!(fp >= DBL_MIN))
+      {
+         *ascii++ = 48; /* '0' */
+         *ascii = 0;
+         return;
+      }
+      else
+      {
+         *ascii++ = 105; /* 'i' */
+         *ascii++ = 110; /* 'n' */
+         *ascii++ = 102; /* 'f' */
+         *ascii = 0;
+         return;
+      }
+   }
+
+   /* Here on buffer too small. */
+   png_error(png_ptr, "ASCII conversion buffer too small");
+}
+
+#  endif /* FLOATING_POINT */
+
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+/* Function to format a fixed point value in ASCII.
+ */
+void /* PRIVATE */
+png_ascii_from_fixed(png_const_structrp png_ptr, png_charp ascii,
+    png_size_t size, png_fixed_point fp)
+{
+   /* Require space for 10 decimal digits, a decimal point, a minus sign and a
+    * trailing \0, 13 characters:
+    */
+   if (size > 12)
+   {
+      png_uint_32 num;
+
+      /* Avoid overflow here on the minimum integer. */
+      if (fp < 0)
+         *ascii++ = 45, --size, num = -fp;
+      else
+         num = fp;
+
+      if (num <= 0x80000000) /* else overflowed */
+      {
+         unsigned int ndigits = 0, first = 16 /* flag value */;
+         char digits[10];
+
+         while (num)
+         {
+            /* Split the low digit off num: */
+            unsigned int tmp = num/10;
+            num -= tmp*10;
+            digits[ndigits++] = (char)(48 + num);
+            /* Record the first non-zero digit, note that this is a number
+             * starting at 1, it's not actually the array index.
+             */
+            if (first == 16 && num > 0)
+               first = ndigits;
+            num = tmp;
+         }
+
+         if (ndigits > 0)
+         {
+            while (ndigits > 5) *ascii++ = digits[--ndigits];
+            /* The remaining digits are fractional digits, ndigits is '5' or
+             * smaller at this point.  It is certainly not zero.  Check for a
+             * non-zero fractional digit:
+             */
+            if (first <= 5)
+            {
+               unsigned int i;
+               *ascii++ = 46; /* decimal point */
+               /* ndigits may be <5 for small numbers, output leading zeros
+                * then ndigits digits to first:
+                */
+               i = 5;
+               while (ndigits < i) *ascii++ = 48, --i;
+               while (ndigits >= first) *ascii++ = digits[--ndigits];
+               /* Don't output the trailing zeros! */
+            }
+         }
+         else
+            *ascii++ = 48;
+
+         /* And null terminate the string: */
+         *ascii = 0;
+         return;
+      }
+   }
+
+   /* Here on buffer too small. */
+   png_error(png_ptr, "ASCII conversion buffer too small");
+}
+#   endif /* FIXED_POINT */
+#endif /* READ_SCAL */
+
+#if defined(PNG_FLOATING_POINT_SUPPORTED) && \
+   !defined(PNG_FIXED_POINT_MACRO_SUPPORTED) && \
+   (defined(PNG_gAMA_SUPPORTED) || defined(PNG_cHRM_SUPPORTED) || \
+   defined(PNG_sCAL_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)) || \
+   (defined(PNG_sCAL_SUPPORTED) && \
+   defined(PNG_FLOATING_ARITHMETIC_SUPPORTED))
+png_fixed_point
+png_fixed(png_const_structrp png_ptr, double fp, png_const_charp text)
+{
+   double r = floor(100000 * fp + .5);
+
+   if (r > 2147483647. || r < -2147483648.)
+      png_fixed_error(png_ptr, text);
+
+   return (png_fixed_point)r;
+}
+#endif
+
+#if defined(PNG_READ_GAMMA_SUPPORTED) || \
+    defined(PNG_INCH_CONVERSIONS_SUPPORTED) || defined(PNG_READ_pHYs_SUPPORTED)
+/* muldiv functions */
+/* This API takes signed arguments and rounds the result to the nearest
+ * integer (or, for a fixed point number - the standard argument - to
+ * the nearest .00001).  Overflow and divide by zero are signalled in
+ * the result, a boolean - true on success, false on overflow.
+ */
+int
+png_muldiv(png_fixed_point_p res, png_fixed_point a, png_int_32 times,
+    png_int_32 divisor)
+{
+   /* Return a * times / divisor, rounded. */
+   if (divisor != 0)
+   {
+      if (a == 0 || times == 0)
+      {
+         *res = 0;
+         return 1;
+      }
+      else
+      {
+#ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+         double r = a;
+         r *= times;
+         r /= divisor;
+         r = floor(r+.5);
+
+         /* A png_fixed_point is a 32-bit integer. */
+         if (r <= 2147483647. && r >= -2147483648.)
+         {
+            *res = (png_fixed_point)r;
+            return 1;
+         }
+#else
+         int negative = 0;
+         png_uint_32 A, T, D;
+         png_uint_32 s16, s32, s00;
+
+         if (a < 0)
+            negative = 1, A = -a;
+         else
+            A = a;
+
+         if (times < 0)
+            negative = !negative, T = -times;
+         else
+            T = times;
+
+         if (divisor < 0)
+            negative = !negative, D = -divisor;
+         else
+            D = divisor;
+
+         /* Following can't overflow because the arguments only
+          * have 31 bits each, however the result may be 32 bits.
+          */
+         s16 = (A >> 16) * (T & 0xffff) +
+                           (A & 0xffff) * (T >> 16);
+         /* Can't overflow because the a*times bit is only 30
+          * bits at most.
+          */
+         s32 = (A >> 16) * (T >> 16) + (s16 >> 16);
+         s00 = (A & 0xffff) * (T & 0xffff);
+
+         s16 = (s16 & 0xffff) << 16;
+         s00 += s16;
+
+         if (s00 < s16)
+            ++s32; /* carry */
+
+         if (s32 < D) /* else overflow */
+         {
+            /* s32.s00 is now the 64-bit product, do a standard
+             * division, we know that s32 < D, so the maximum
+             * required shift is 31.
+             */
+            int bitshift = 32;
+            png_fixed_point result = 0; /* NOTE: signed */
+
+            while (--bitshift >= 0)
+            {
+               png_uint_32 d32, d00;
+
+               if (bitshift > 0)
+                  d32 = D >> (32-bitshift), d00 = D << bitshift;
+
+               else
+                  d32 = 0, d00 = D;
+
+               if (s32 > d32)
+               {
+                  if (s00 < d00) --s32; /* carry */
+                  s32 -= d32, s00 -= d00, result += 1<<bitshift;
+               }
+
+               else
+                  if (s32 == d32 && s00 >= d00)
+                     s32 = 0, s00 -= d00, result += 1<<bitshift;
+            }
+
+            /* Handle the rounding. */
+            if (s00 >= (D >> 1))
+               ++result;
+
+            if (negative)
+               result = -result;
+
+            /* Check for overflow. */
+            if ((negative && result <= 0) || (!negative && result >= 0))
+            {
+               *res = result;
+               return 1;
+            }
+         }
+#endif
+      }
+   }
+
+   return 0;
+}
+#endif /* READ_GAMMA || INCH_CONVERSIONS */
+
+#if defined(PNG_READ_GAMMA_SUPPORTED) || defined(PNG_INCH_CONVERSIONS_SUPPORTED)
+/* The following is for when the caller doesn't much care about the
+ * result.
+ */
+png_fixed_point
+png_muldiv_warn(png_const_structrp png_ptr, png_fixed_point a, png_int_32 times,
+    png_int_32 divisor)
+{
+   png_fixed_point result;
+
+   if (png_muldiv(&result, a, times, divisor))
+      return result;
+
+   png_warning(png_ptr, "fixed point overflow ignored");
+   return 0;
+}
+#endif
+
+#ifdef PNG_GAMMA_SUPPORTED /* more fixed point functions for gamma */
+/* Calculate a reciprocal, return 0 on div-by-zero or overflow. */
+png_fixed_point
+png_reciprocal(png_fixed_point a)
+{
+#ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+   double r = floor(1E10/a+.5);
+
+   if (r <= 2147483647. && r >= -2147483648.)
+      return (png_fixed_point)r;
+#else
+   png_fixed_point res;
+
+   if (png_muldiv(&res, 100000, 100000, a))
+      return res;
+#endif
+
+   return 0; /* error/overflow */
+}
+
+/* This is the shared test on whether a gamma value is 'significant' - whether
+ * it is worth doing gamma correction.
+ */
+int /* PRIVATE */
+png_gamma_significant(png_fixed_point gamma_val)
+{
+   return gamma_val < PNG_FP_1 - PNG_GAMMA_THRESHOLD_FIXED ||
+       gamma_val > PNG_FP_1 + PNG_GAMMA_THRESHOLD_FIXED;
+}
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* A local convenience routine. */
+static png_fixed_point
+png_product2(png_fixed_point a, png_fixed_point b)
+{
+   /* The required result is 1/a * 1/b; the following preserves accuracy. */
+#ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+   double r = a * 1E-5;
+   r *= b;
+   r = floor(r+.5);
+
+   if (r <= 2147483647. && r >= -2147483648.)
+      return (png_fixed_point)r;
+#else
+   png_fixed_point res;
+
+   if (png_muldiv(&res, a, b, 100000))
+      return res;
+#endif
+
+   return 0; /* overflow */
+}
+
+/* The inverse of the above. */
+png_fixed_point
+png_reciprocal2(png_fixed_point a, png_fixed_point b)
+{
+   /* The required result is 1/a * 1/b; the following preserves accuracy. */
+#ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+   double r = 1E15/a;
+   r /= b;
+   r = floor(r+.5);
+
+   if (r <= 2147483647. && r >= -2147483648.)
+      return (png_fixed_point)r;
+#else
+   /* This may overflow because the range of png_fixed_point isn't symmetric,
+    * but this API is only used for the product of file and screen gamma so it
+    * doesn't matter that the smallest number it can produce is 1/21474, not
+    * 1/100000
+    */
+   png_fixed_point res = png_product2(a, b);
+
+   if (res != 0)
+      return png_reciprocal(res);
+#endif
+
+   return 0; /* overflow */
+}
+#endif /* READ_GAMMA */
+
+#ifdef PNG_READ_GAMMA_SUPPORTED /* gamma table code */
+#ifndef PNG_FLOATING_ARITHMETIC_SUPPORTED
+/* Fixed point gamma.
+ *
+ * The code to calculate the tables used below can be found in the shell script
+ * contrib/tools/intgamma.sh
+ *
+ * To calculate gamma this code implements fast log() and exp() calls using only
+ * fixed point arithmetic.  This code has sufficient precision for either 8-bit
+ * or 16-bit sample values.
+ *
+ * The tables used here were calculated using simple 'bc' programs, but C double
+ * precision floating point arithmetic would work fine.
+ *
+ * 8-bit log table
+ *   This is a table of -log(value/255)/log(2) for 'value' in the range 128 to
+ *   255, so it's the base 2 logarithm of a normalized 8-bit floating point
+ *   mantissa.  The numbers are 32-bit fractions.
+ */
+static const png_uint_32
+png_8bit_l2[128] =
+{
+   4270715492U, 4222494797U, 4174646467U, 4127164793U, 4080044201U, 4033279239U,
+   3986864580U, 3940795015U, 3895065449U, 3849670902U, 3804606499U, 3759867474U,
+   3715449162U, 3671346997U, 3627556511U, 3584073329U, 3540893168U, 3498011834U,
+   3455425220U, 3413129301U, 3371120137U, 3329393864U, 3287946700U, 3246774933U,
+   3205874930U, 3165243125U, 3124876025U, 3084770202U, 3044922296U, 3005329011U,
+   2965987113U, 2926893432U, 2888044853U, 2849438323U, 2811070844U, 2772939474U,
+   2735041326U, 2697373562U, 2659933400U, 2622718104U, 2585724991U, 2548951424U,
+   2512394810U, 2476052606U, 2439922311U, 2404001468U, 2368287663U, 2332778523U,
+   2297471715U, 2262364947U, 2227455964U, 2192742551U, 2158222529U, 2123893754U,
+   2089754119U, 2055801552U, 2022034013U, 1988449497U, 1955046031U, 1921821672U,
+   1888774511U, 1855902668U, 1823204291U, 1790677560U, 1758320682U, 1726131893U,
+   1694109454U, 1662251657U, 1630556815U, 1599023271U, 1567649391U, 1536433567U,
+   1505374214U, 1474469770U, 1443718700U, 1413119487U, 1382670639U, 1352370686U,
+   1322218179U, 1292211689U, 1262349810U, 1232631153U, 1203054352U, 1173618059U,
+   1144320946U, 1115161701U, 1086139034U, 1057251672U, 1028498358U, 999877854U,
+   971388940U, 943030410U, 914801076U, 886699767U, 858725327U, 830876614U,
+   803152505U, 775551890U, 748073672U, 720716771U, 693480120U, 666362667U,
+   639363374U, 612481215U, 585715177U, 559064263U, 532527486U, 506103872U,
+   479792461U, 453592303U, 427502463U, 401522014U, 375650043U, 349885648U,
+   324227938U, 298676034U, 273229066U, 247886176U, 222646516U, 197509248U,
+   172473545U, 147538590U, 122703574U, 97967701U, 73330182U, 48790236U,
+   24347096U, 0U
+
+#if 0
+   /* The following are the values for 16-bit tables - these work fine for the
+    * 8-bit conversions but produce very slightly larger errors in the 16-bit
+    * log (about 1.2 as opposed to 0.7 absolute error in the final value).  To
+    * use these all the shifts below must be adjusted appropriately.
+    */
+   65166, 64430, 63700, 62976, 62257, 61543, 60835, 60132, 59434, 58741, 58054,
+   57371, 56693, 56020, 55352, 54689, 54030, 53375, 52726, 52080, 51439, 50803,
+   50170, 49542, 48918, 48298, 47682, 47070, 46462, 45858, 45257, 44661, 44068,
+   43479, 42894, 42312, 41733, 41159, 40587, 40020, 39455, 38894, 38336, 37782,
+   37230, 36682, 36137, 35595, 35057, 34521, 33988, 33459, 32932, 32408, 31887,
+   31369, 30854, 30341, 29832, 29325, 28820, 28319, 27820, 27324, 26830, 26339,
+   25850, 25364, 24880, 24399, 23920, 23444, 22970, 22499, 22029, 21562, 21098,
+   20636, 20175, 19718, 19262, 18808, 18357, 17908, 17461, 17016, 16573, 16132,
+   15694, 15257, 14822, 14390, 13959, 13530, 13103, 12678, 12255, 11834, 11415,
+   10997, 10582, 10168, 9756, 9346, 8937, 8531, 8126, 7723, 7321, 6921, 6523,
+   6127, 5732, 5339, 4947, 4557, 4169, 3782, 3397, 3014, 2632, 2251, 1872, 1495,
+   1119, 744, 372
+#endif
+};
+
+static png_int_32
+png_log8bit(unsigned int x)
+{
+   unsigned int lg2 = 0;
+   /* Each time 'x' is multiplied by 2, 1 must be subtracted off the final log,
+    * because the log is actually negate that means adding 1.  The final
+    * returned value thus has the range 0 (for 255 input) to 7.994 (for 1
+    * input), return -1 for the overflow (log 0) case, - so the result is
+    * always at most 19 bits.
+    */
+   if ((x &= 0xff) == 0)
+      return -1;
+
+   if ((x & 0xf0) == 0)
+      lg2  = 4, x <<= 4;
+
+   if ((x & 0xc0) == 0)
+      lg2 += 2, x <<= 2;
+
+   if ((x & 0x80) == 0)
+      lg2 += 1, x <<= 1;
+
+   /* result is at most 19 bits, so this cast is safe: */
+   return (png_int_32)((lg2 << 16) + ((png_8bit_l2[x-128]+32768)>>16));
+}
+
+/* The above gives exact (to 16 binary places) log2 values for 8-bit images,
+ * for 16-bit images we use the most significant 8 bits of the 16-bit value to
+ * get an approximation then multiply the approximation by a correction factor
+ * determined by the remaining up to 8 bits.  This requires an additional step
+ * in the 16-bit case.
+ *
+ * We want log2(value/65535), we have log2(v'/255), where:
+ *
+ *    value = v' * 256 + v''
+ *          = v' * f
+ *
+ * So f is value/v', which is equal to (256+v''/v') since v' is in the range 128
+ * to 255 and v'' is in the range 0 to 255 f will be in the range 256 to less
+ * than 258.  The final factor also needs to correct for the fact that our 8-bit
+ * value is scaled by 255, whereas the 16-bit values must be scaled by 65535.
+ *
+ * This gives a final formula using a calculated value 'x' which is value/v' and
+ * scaling by 65536 to match the above table:
+ *
+ *   log2(x/257) * 65536
+ *
+ * Since these numbers are so close to '1' we can use simple linear
+ * interpolation between the two end values 256/257 (result -368.61) and 258/257
+ * (result 367.179).  The values used below are scaled by a further 64 to give
+ * 16-bit precision in the interpolation:
+ *
+ * Start (256): -23591
+ * Zero  (257):      0
+ * End   (258):  23499
+ */
+static png_int_32
+png_log16bit(png_uint_32 x)
+{
+   unsigned int lg2 = 0;
+
+   /* As above, but now the input has 16 bits. */
+   if ((x &= 0xffff) == 0)
+      return -1;
+
+   if ((x & 0xff00) == 0)
+      lg2  = 8, x <<= 8;
+
+   if ((x & 0xf000) == 0)
+      lg2 += 4, x <<= 4;
+
+   if ((x & 0xc000) == 0)
+      lg2 += 2, x <<= 2;
+
+   if ((x & 0x8000) == 0)
+      lg2 += 1, x <<= 1;
+
+   /* Calculate the base logarithm from the top 8 bits as a 28-bit fractional
+    * value.
+    */
+   lg2 <<= 28;
+   lg2 += (png_8bit_l2[(x>>8)-128]+8) >> 4;
+
+   /* Now we need to interpolate the factor, this requires a division by the top
+    * 8 bits.  Do this with maximum precision.
+    */
+   x = ((x << 16) + (x >> 9)) / (x >> 8);
+
+   /* Since we divided by the top 8 bits of 'x' there will be a '1' at 1<<24,
+    * the value at 1<<16 (ignoring this) will be 0 or 1; this gives us exactly
+    * 16 bits to interpolate to get the low bits of the result.  Round the
+    * answer.  Note that the end point values are scaled by 64 to retain overall
+    * precision and that 'lg2' is current scaled by an extra 12 bits, so adjust
+    * the overall scaling by 6-12.  Round at every step.
+    */
+   x -= 1U << 24;
+
+   if (x <= 65536U) /* <= '257' */
+      lg2 += ((23591U * (65536U-x)) + (1U << (16+6-12-1))) >> (16+6-12);
+
+   else
+      lg2 -= ((23499U * (x-65536U)) + (1U << (16+6-12-1))) >> (16+6-12);
+
+   /* Safe, because the result can't have more than 20 bits: */
+   return (png_int_32)((lg2 + 2048) >> 12);
+}
+
+/* The 'exp()' case must invert the above, taking a 20-bit fixed point
+ * logarithmic value and returning a 16 or 8-bit number as appropriate.  In
+ * each case only the low 16 bits are relevant - the fraction - since the
+ * integer bits (the top 4) simply determine a shift.
+ *
+ * The worst case is the 16-bit distinction between 65535 and 65534, this
+ * requires perhaps spurious accuracty in the decoding of the logarithm to
+ * distinguish log2(65535/65534.5) - 10^-5 or 17 bits.  There is little chance
+ * of getting this accuracy in practice.
+ *
+ * To deal with this the following exp() function works out the exponent of the
+ * frational part of the logarithm by using an accurate 32-bit value from the
+ * top four fractional bits then multiplying in the remaining bits.
+ */
+static const png_uint_32
+png_32bit_exp[16] =
+{
+   /* NOTE: the first entry is deliberately set to the maximum 32-bit value. */
+   4294967295U, 4112874773U, 3938502376U, 3771522796U, 3611622603U, 3458501653U,
+   3311872529U, 3171459999U, 3037000500U, 2908241642U, 2784941738U, 2666869345U,
+   2553802834U, 2445529972U, 2341847524U, 2242560872U
+};
+
+/* Adjustment table; provided to explain the numbers in the code below. */
+#if 0
+for (i=11;i>=0;--i){ print i, " ", (1 - e(-(2^i)/65536*l(2))) * 2^(32-i), "\n"}
+   11 44937.64284865548751208448
+   10 45180.98734845585101160448
+    9 45303.31936980687359311872
+    8 45364.65110595323018870784
+    7 45395.35850361789624614912
+    6 45410.72259715102037508096
+    5 45418.40724413220722311168
+    4 45422.25021786898173001728
+    3 45424.17186732298419044352
+    2 45425.13273269940811464704
+    1 45425.61317555035558641664
+    0 45425.85339951654943850496
+#endif
+
+static png_uint_32
+png_exp(png_fixed_point x)
+{
+   if (x > 0 && x <= 0xfffff) /* Else overflow or zero (underflow) */
+   {
+      /* Obtain a 4-bit approximation */
+      png_uint_32 e = png_32bit_exp[(x >> 12) & 0xf];
+
+      /* Incorporate the low 12 bits - these decrease the returned value by
+       * multiplying by a number less than 1 if the bit is set.  The multiplier
+       * is determined by the above table and the shift. Notice that the values
+       * converge on 45426 and this is used to allow linear interpolation of the
+       * low bits.
+       */
+      if (x & 0x800)
+         e -= (((e >> 16) * 44938U) +  16U) >> 5;
+
+      if (x & 0x400)
+         e -= (((e >> 16) * 45181U) +  32U) >> 6;
+
+      if (x & 0x200)
+         e -= (((e >> 16) * 45303U) +  64U) >> 7;
+
+      if (x & 0x100)
+         e -= (((e >> 16) * 45365U) + 128U) >> 8;
+
+      if (x & 0x080)
+         e -= (((e >> 16) * 45395U) + 256U) >> 9;
+
+      if (x & 0x040)
+         e -= (((e >> 16) * 45410U) + 512U) >> 10;
+
+      /* And handle the low 6 bits in a single block. */
+      e -= (((e >> 16) * 355U * (x & 0x3fU)) + 256U) >> 9;
+
+      /* Handle the upper bits of x. */
+      e >>= x >> 16;
+      return e;
+   }
+
+   /* Check for overflow */
+   if (x <= 0)
+      return png_32bit_exp[0];
+
+   /* Else underflow */
+   return 0;
+}
+
+static png_byte
+png_exp8bit(png_fixed_point lg2)
+{
+   /* Get a 32-bit value: */
+   png_uint_32 x = png_exp(lg2);
+
+   /* Convert the 32-bit value to 0..255 by multiplying by 256-1, note that the
+    * second, rounding, step can't overflow because of the first, subtraction,
+    * step.
+    */
+   x -= x >> 8;
+   return (png_byte)((x + 0x7fffffU) >> 24);
+}
+
+static png_uint_16
+png_exp16bit(png_fixed_point lg2)
+{
+   /* Get a 32-bit value: */
+   png_uint_32 x = png_exp(lg2);
+
+   /* Convert the 32-bit value to 0..65535 by multiplying by 65536-1: */
+   x -= x >> 16;
+   return (png_uint_16)((x + 32767U) >> 16);
+}
+#endif /* FLOATING_ARITHMETIC */
+
+png_byte
+png_gamma_8bit_correct(unsigned int value, png_fixed_point gamma_val)
+{
+   if (value > 0 && value < 255)
+   {
+#     ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+         double r = floor(255*pow(value/255.,gamma_val*.00001)+.5);
+         return (png_byte)r;
+#     else
+         png_int_32 lg2 = png_log8bit(value);
+         png_fixed_point res;
+
+         if (png_muldiv(&res, gamma_val, lg2, PNG_FP_1))
+            return png_exp8bit(res);
+
+         /* Overflow. */
+         value = 0;
+#     endif
+   }
+
+   return (png_byte)value;
+}
+
+png_uint_16
+png_gamma_16bit_correct(unsigned int value, png_fixed_point gamma_val)
+{
+   if (value > 0 && value < 65535)
+   {
+#     ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+         double r = floor(65535*pow(value/65535.,gamma_val*.00001)+.5);
+         return (png_uint_16)r;
+#     else
+         png_int_32 lg2 = png_log16bit(value);
+         png_fixed_point res;
+
+         if (png_muldiv(&res, gamma_val, lg2, PNG_FP_1))
+            return png_exp16bit(res);
+
+         /* Overflow. */
+         value = 0;
+#     endif
+   }
+
+   return (png_uint_16)value;
+}
+
+/* This does the right thing based on the bit_depth field of the
+ * png_struct, interpreting values as 8-bit or 16-bit.  While the result
+ * is nominally a 16-bit value if bit depth is 8 then the result is
+ * 8-bit (as are the arguments.)
+ */
+png_uint_16 /* PRIVATE */
+png_gamma_correct(png_structrp png_ptr, unsigned int value,
+    png_fixed_point gamma_val)
+{
+   if (png_ptr->bit_depth == 8)
+      return png_gamma_8bit_correct(value, gamma_val);
+
+   else
+      return png_gamma_16bit_correct(value, gamma_val);
+}
+
+/* Internal function to build a single 16-bit table - the table consists of
+ * 'num' 256 entry subtables, where 'num' is determined by 'shift' - the amount
+ * to shift the input values right (or 16-number_of_signifiant_bits).
+ *
+ * The caller is responsible for ensuring that the table gets cleaned up on
+ * png_error (i.e. if one of the mallocs below fails) - i.e. the *table argument
+ * should be somewhere that will be cleaned.
+ */
+static void
+png_build_16bit_table(png_structrp png_ptr, png_uint_16pp *ptable,
+   PNG_CONST unsigned int shift, PNG_CONST png_fixed_point gamma_val)
+{
+   /* Various values derived from 'shift': */
+   PNG_CONST unsigned int num = 1U << (8U - shift);
+   PNG_CONST unsigned int max = (1U << (16U - shift))-1U;
+   PNG_CONST unsigned int max_by_2 = 1U << (15U-shift);
+   unsigned int i;
+
+   png_uint_16pp table = *ptable =
+       (png_uint_16pp)png_calloc(png_ptr, num * (sizeof (png_uint_16p)));
+
+   for (i = 0; i < num; i++)
+   {
+      png_uint_16p sub_table = table[i] =
+          (png_uint_16p)png_malloc(png_ptr, 256 * (sizeof (png_uint_16)));
+
+      /* The 'threshold' test is repeated here because it can arise for one of
+       * the 16-bit tables even if the others don't hit it.
+       */
+      if (png_gamma_significant(gamma_val))
+      {
+         /* The old code would overflow at the end and this would cause the
+          * 'pow' function to return a result >1, resulting in an
+          * arithmetic error.  This code follows the spec exactly; ig is
+          * the recovered input sample, it always has 8-16 bits.
+          *
+          * We want input * 65535/max, rounded, the arithmetic fits in 32
+          * bits (unsigned) so long as max <= 32767.
+          */
+         unsigned int j;
+         for (j = 0; j < 256; j++)
+         {
+            png_uint_32 ig = (j << (8-shift)) + i;
+#           ifdef PNG_FLOATING_ARITHMETIC_SUPPORTED
+               /* Inline the 'max' scaling operation: */
+               double d = floor(65535*pow(ig/(double)max, gamma_val*.00001)+.5);
+               sub_table[j] = (png_uint_16)d;
+#           else
+               if (shift)
+                  ig = (ig * 65535U + max_by_2)/max;
+
+               sub_table[j] = png_gamma_16bit_correct(ig, gamma_val);
+#           endif
+         }
+      }
+      else
+      {
+         /* We must still build a table, but do it the fast way. */
+         unsigned int j;
+
+         for (j = 0; j < 256; j++)
+         {
+            png_uint_32 ig = (j << (8-shift)) + i;
+
+            if (shift)
+               ig = (ig * 65535U + max_by_2)/max;
+
+            sub_table[j] = (png_uint_16)ig;
+         }
+      }
+   }
+}
+
+/* NOTE: this function expects the *inverse* of the overall gamma transformation
+ * required.
+ */
+static void
+png_build_16to8_table(png_structrp png_ptr, png_uint_16pp *ptable,
+   PNG_CONST unsigned int shift, PNG_CONST png_fixed_point gamma_val)
+{
+   PNG_CONST unsigned int num = 1U << (8U - shift);
+   PNG_CONST unsigned int max = (1U << (16U - shift))-1U;
+   unsigned int i;
+   png_uint_32 last;
+
+   png_uint_16pp table = *ptable =
+       (png_uint_16pp)png_calloc(png_ptr, num * (sizeof (png_uint_16p)));
+
+   /* 'num' is the number of tables and also the number of low bits of low
+    * bits of the input 16-bit value used to select a table.  Each table is
+    * itself index by the high 8 bits of the value.
+    */
+   for (i = 0; i < num; i++)
+      table[i] = (png_uint_16p)png_malloc(png_ptr,
+          256 * (sizeof (png_uint_16)));
+
+   /* 'gamma_val' is set to the reciprocal of the value calculated above, so
+    * pow(out,g) is an *input* value.  'last' is the last input value set.
+    *
+    * In the loop 'i' is used to find output values.  Since the output is
+    * 8-bit there are only 256 possible values.  The tables are set up to
+    * select the closest possible output value for each input by finding
+    * the input value at the boundary between each pair of output values
+    * and filling the table up to that boundary with the lower output
+    * value.
+    *
+    * The boundary values are 0.5,1.5..253.5,254.5.  Since these are 9-bit
+    * values the code below uses a 16-bit value in i; the values start at
+    * 128.5 (for 0.5) and step by 257, for a total of 254 values (the last
+    * entries are filled with 255).  Start i at 128 and fill all 'last'
+    * table entries <= 'max'
+    */
+   last = 0;
+   for (i = 0; i < 255; ++i) /* 8-bit output value */
+   {
+      /* Find the corresponding maximum input value */
+      png_uint_16 out = (png_uint_16)(i * 257U); /* 16-bit output value */
+
+      /* Find the boundary value in 16 bits: */
+      png_uint_32 bound = png_gamma_16bit_correct(out+128U, gamma_val);
+
+      /* Adjust (round) to (16-shift) bits: */
+      bound = (bound * max + 32768U)/65535U + 1U;
+
+      while (last < bound)
+      {
+         table[last & (0xffU >> shift)][last >> (8U - shift)] = out;
+         last++;
+      }
+   }
+
+   /* And fill in the final entries. */
+   while (last < (num << 8))
+   {
+      table[last & (0xff >> shift)][last >> (8U - shift)] = 65535U;
+      last++;
+   }
+}
+
+/* Build a single 8-bit table: same as the 16-bit case but much simpler (and
+ * typically much faster).  Note that libpng currently does no sBIT processing
+ * (apparently contrary to the spec) so a 256 entry table is always generated.
+ */
+static void
+png_build_8bit_table(png_structrp png_ptr, png_bytepp ptable,
+   PNG_CONST png_fixed_point gamma_val)
+{
+   unsigned int i;
+   png_bytep table = *ptable = (png_bytep)png_malloc(png_ptr, 256);
+
+   if (png_gamma_significant(gamma_val)) for (i=0; i<256; i++)
+      table[i] = png_gamma_8bit_correct(i, gamma_val);
+
+   else for (i=0; i<256; ++i)
+      table[i] = (png_byte)i;
+}
+
+/* Used from png_read_destroy and below to release the memory used by the gamma
+ * tables.
+ */
+void /* PRIVATE */
+png_destroy_gamma_table(png_structrp png_ptr)
+{
+   png_free(png_ptr, png_ptr->gamma_table);
+   png_ptr->gamma_table = NULL;
+
+   if (png_ptr->gamma_16_table != NULL)
+   {
+      int i;
+      int istop = (1 << (8 - png_ptr->gamma_shift));
+      for (i = 0; i < istop; i++)
+      {
+         png_free(png_ptr, png_ptr->gamma_16_table[i]);
+      }
+   png_free(png_ptr, png_ptr->gamma_16_table);
+   png_ptr->gamma_16_table = NULL;
+   }
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)
+   png_free(png_ptr, png_ptr->gamma_from_1);
+   png_ptr->gamma_from_1 = NULL;
+   png_free(png_ptr, png_ptr->gamma_to_1);
+   png_ptr->gamma_to_1 = NULL;
+
+   if (png_ptr->gamma_16_from_1 != NULL)
+   {
+      int i;
+      int istop = (1 << (8 - png_ptr->gamma_shift));
+      for (i = 0; i < istop; i++)
+      {
+         png_free(png_ptr, png_ptr->gamma_16_from_1[i]);
+      }
+   png_free(png_ptr, png_ptr->gamma_16_from_1);
+   png_ptr->gamma_16_from_1 = NULL;
+   }
+   if (png_ptr->gamma_16_to_1 != NULL)
+   {
+      int i;
+      int istop = (1 << (8 - png_ptr->gamma_shift));
+      for (i = 0; i < istop; i++)
+      {
+         png_free(png_ptr, png_ptr->gamma_16_to_1[i]);
+      }
+   png_free(png_ptr, png_ptr->gamma_16_to_1);
+   png_ptr->gamma_16_to_1 = NULL;
+   }
+#endif /* READ_BACKGROUND || READ_ALPHA_MODE || RGB_TO_GRAY */
+}
+
+/* We build the 8- or 16-bit gamma tables here.  Note that for 16-bit
+ * tables, we don't make a full table if we are reducing to 8-bit in
+ * the future.  Note also how the gamma_16 tables are segmented so that
+ * we don't need to allocate > 64K chunks for a full 16-bit table.
+ */
+void /* PRIVATE */
+png_build_gamma_table(png_structrp png_ptr, int bit_depth)
+{
+  png_debug(1, "in png_build_gamma_table");
+
+  /* Remove any existing table; this copes with multiple calls to
+   * png_read_update_info.  The warning is because building the gamma tables
+   * multiple times is a performance hit - it's harmless but the ability to call
+   * png_read_update_info() multiple times is new in 1.5.6 so it seems sensible
+   * to warn if the app introduces such a hit.
+   */
+  if (png_ptr->gamma_table != NULL || png_ptr->gamma_16_table != NULL)
+  {
+    png_warning(png_ptr, "gamma table being rebuilt");
+    png_destroy_gamma_table(png_ptr);
+  }
+
+  if (bit_depth <= 8)
+  {
+     png_build_8bit_table(png_ptr, &png_ptr->gamma_table,
+         png_ptr->screen_gamma > 0 ?  png_reciprocal2(png_ptr->colorspace.gamma,
+         png_ptr->screen_gamma) : PNG_FP_1);
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)
+     if (png_ptr->transformations & (PNG_COMPOSE | PNG_RGB_TO_GRAY))
+     {
+        png_build_8bit_table(png_ptr, &png_ptr->gamma_to_1,
+            png_reciprocal(png_ptr->colorspace.gamma));
+
+        png_build_8bit_table(png_ptr, &png_ptr->gamma_from_1,
+            png_ptr->screen_gamma > 0 ?  png_reciprocal(png_ptr->screen_gamma) :
+            png_ptr->colorspace.gamma/* Probably doing rgb_to_gray */);
+     }
+#endif /* READ_BACKGROUND || READ_ALPHA_MODE || RGB_TO_GRAY */
+  }
+  else
+  {
+     png_byte shift, sig_bit;
+
+     if (png_ptr->color_type & PNG_COLOR_MASK_COLOR)
+     {
+        sig_bit = png_ptr->sig_bit.red;
+
+        if (png_ptr->sig_bit.green > sig_bit)
+           sig_bit = png_ptr->sig_bit.green;
+
+        if (png_ptr->sig_bit.blue > sig_bit)
+           sig_bit = png_ptr->sig_bit.blue;
+     }
+     else
+        sig_bit = png_ptr->sig_bit.gray;
+
+     /* 16-bit gamma code uses this equation:
+      *
+      *   ov = table[(iv & 0xff) >> gamma_shift][iv >> 8]
+      *
+      * Where 'iv' is the input color value and 'ov' is the output value -
+      * pow(iv, gamma).
+      *
+      * Thus the gamma table consists of up to 256 256 entry tables.  The table
+      * is selected by the (8-gamma_shift) most significant of the low 8 bits of
+      * the color value then indexed by the upper 8 bits:
+      *
+      *   table[low bits][high 8 bits]
+      *
+      * So the table 'n' corresponds to all those 'iv' of:
+      *
+      *   <all high 8-bit values><n << gamma_shift>..<(n+1 << gamma_shift)-1>
+      *
+      */
+     if (sig_bit > 0 && sig_bit < 16U)
+        shift = (png_byte)(16U - sig_bit); /* shift == insignificant bits */
+
+     else
+        shift = 0; /* keep all 16 bits */
+
+     if (png_ptr->transformations & (PNG_16_TO_8 | PNG_SCALE_16_TO_8))
+     {
+        /* PNG_MAX_GAMMA_8 is the number of bits to keep - effectively
+         * the significant bits in the *input* when the output will
+         * eventually be 8 bits.  By default it is 11.
+         */
+        if (shift < (16U - PNG_MAX_GAMMA_8))
+           shift = (16U - PNG_MAX_GAMMA_8);
+     }
+
+     if (shift > 8U)
+        shift = 8U; /* Guarantees at least one table! */
+
+     png_ptr->gamma_shift = shift;
+
+#ifdef PNG_16BIT_SUPPORTED
+     /* NOTE: prior to 1.5.4 this test used to include PNG_BACKGROUND (now
+      * PNG_COMPOSE).  This effectively smashed the background calculation for
+      * 16-bit output because the 8-bit table assumes the result will be reduced
+      * to 8 bits.
+      */
+     if (png_ptr->transformations & (PNG_16_TO_8 | PNG_SCALE_16_TO_8))
+#endif
+         png_build_16to8_table(png_ptr, &png_ptr->gamma_16_table, shift,
+         png_ptr->screen_gamma > 0 ? png_product2(png_ptr->colorspace.gamma,
+         png_ptr->screen_gamma) : PNG_FP_1);
+
+#ifdef PNG_16BIT_SUPPORTED
+     else
+         png_build_16bit_table(png_ptr, &png_ptr->gamma_16_table, shift,
+         png_ptr->screen_gamma > 0 ? png_reciprocal2(png_ptr->colorspace.gamma,
+         png_ptr->screen_gamma) : PNG_FP_1);
+#endif
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)
+     if (png_ptr->transformations & (PNG_COMPOSE | PNG_RGB_TO_GRAY))
+     {
+        png_build_16bit_table(png_ptr, &png_ptr->gamma_16_to_1, shift,
+            png_reciprocal(png_ptr->colorspace.gamma));
+
+        /* Notice that the '16 from 1' table should be full precision, however
+         * the lookup on this table still uses gamma_shift, so it can't be.
+         * TODO: fix this.
+         */
+        png_build_16bit_table(png_ptr, &png_ptr->gamma_16_from_1, shift,
+            png_ptr->screen_gamma > 0 ? png_reciprocal(png_ptr->screen_gamma) :
+            png_ptr->colorspace.gamma/* Probably doing rgb_to_gray */);
+     }
+#endif /* READ_BACKGROUND || READ_ALPHA_MODE || RGB_TO_GRAY */
+  }
+}
+#endif /* READ_GAMMA */
+
+/* HARDWARE OPTION SUPPORT */
+#ifdef PNG_SET_OPTION_SUPPORTED
+int PNGAPI
+png_set_option(png_structrp png_ptr, int option, int onoff)
+{
+   if (png_ptr != NULL && option >= 0 && option < PNG_OPTION_NEXT &&
+      (option & 1) == 0)
+   {
+      int mask = 3 << option;
+      int setting = (2 + (onoff != 0)) << option;
+      int current = png_ptr->options;
+
+      png_ptr->options = (png_byte)((current & ~mask) | setting);
+
+      return (current & mask) >> option;
+   }
+
+   return PNG_OPTION_INVALID;
+}
+#endif
+
+/* sRGB support */
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) ||\
+   defined(PNG_SIMPLIFIED_WRITE_SUPPORTED)
+/* sRGB conversion tables; these are machine generated with the code in
+ * contrib/tools/makesRGB.c.  The actual sRGB transfer curve defined in the
+ * specification (see the article at http://en.wikipedia.org/wiki/SRGB)
+ * is used, not the gamma=1/2.2 approximation use elsewhere in libpng.
+ * The sRGB to linear table is exact (to the nearest 16 bit linear fraction).
+ * The inverse (linear to sRGB) table has accuracies as follows:
+ *
+ * For all possible (255*65535+1) input values:
+ *
+ *    error: -0.515566 - 0.625971, 79441 (0.475369%) of readings inexact
+ *
+ * For the input values corresponding to the 65536 16-bit values:
+ *
+ *    error: -0.513727 - 0.607759, 308 (0.469978%) of readings inexact
+ *
+ * In all cases the inexact readings are off by one.
+ */
+
+#ifdef PNG_SIMPLIFIED_READ_SUPPORTED
+/* The convert-to-sRGB table is only currently required for read. */
+const png_uint_16 png_sRGB_table[256] =
+{
+   0,20,40,60,80,99,119,139,
+   159,179,199,219,241,264,288,313,
+   340,367,396,427,458,491,526,562,
+   599,637,677,718,761,805,851,898,
+   947,997,1048,1101,1156,1212,1270,1330,
+   1391,1453,1517,1583,1651,1720,1790,1863,
+   1937,2013,2090,2170,2250,2333,2418,2504,
+   2592,2681,2773,2866,2961,3058,3157,3258,
+   3360,3464,3570,3678,3788,3900,4014,4129,
+   4247,4366,4488,4611,4736,4864,4993,5124,
+   5257,5392,5530,5669,5810,5953,6099,6246,
+   6395,6547,6700,6856,7014,7174,7335,7500,
+   7666,7834,8004,8177,8352,8528,8708,8889,
+   9072,9258,9445,9635,9828,10022,10219,10417,
+   10619,10822,11028,11235,11446,11658,11873,12090,
+   12309,12530,12754,12980,13209,13440,13673,13909,
+   14146,14387,14629,14874,15122,15371,15623,15878,
+   16135,16394,16656,16920,17187,17456,17727,18001,
+   18277,18556,18837,19121,19407,19696,19987,20281,
+   20577,20876,21177,21481,21787,22096,22407,22721,
+   23038,23357,23678,24002,24329,24658,24990,25325,
+   25662,26001,26344,26688,27036,27386,27739,28094,
+   28452,28813,29176,29542,29911,30282,30656,31033,
+   31412,31794,32179,32567,32957,33350,33745,34143,
+   34544,34948,35355,35764,36176,36591,37008,37429,
+   37852,38278,38706,39138,39572,40009,40449,40891,
+   41337,41785,42236,42690,43147,43606,44069,44534,
+   45002,45473,45947,46423,46903,47385,47871,48359,
+   48850,49344,49841,50341,50844,51349,51858,52369,
+   52884,53401,53921,54445,54971,55500,56032,56567,
+   57105,57646,58190,58737,59287,59840,60396,60955,
+   61517,62082,62650,63221,63795,64372,64952,65535
+};
+
+#endif /* simplified read only */
+
+/* The base/delta tables are required for both read and write (but currently
+ * only the simplified versions.)
+ */
+const png_uint_16 png_sRGB_base[512] =
+{
+   128,1782,3383,4644,5675,6564,7357,8074,
+   8732,9346,9921,10463,10977,11466,11935,12384,
+   12816,13233,13634,14024,14402,14769,15125,15473,
+   15812,16142,16466,16781,17090,17393,17690,17981,
+   18266,18546,18822,19093,19359,19621,19879,20133,
+   20383,20630,20873,21113,21349,21583,21813,22041,
+   22265,22487,22707,22923,23138,23350,23559,23767,
+   23972,24175,24376,24575,24772,24967,25160,25352,
+   25542,25730,25916,26101,26284,26465,26645,26823,
+   27000,27176,27350,27523,27695,27865,28034,28201,
+   28368,28533,28697,28860,29021,29182,29341,29500,
+   29657,29813,29969,30123,30276,30429,30580,30730,
+   30880,31028,31176,31323,31469,31614,31758,31902,
+   32045,32186,32327,32468,32607,32746,32884,33021,
+   33158,33294,33429,33564,33697,33831,33963,34095,
+   34226,34357,34486,34616,34744,34873,35000,35127,
+   35253,35379,35504,35629,35753,35876,35999,36122,
+   36244,36365,36486,36606,36726,36845,36964,37083,
+   37201,37318,37435,37551,37668,37783,37898,38013,
+   38127,38241,38354,38467,38580,38692,38803,38915,
+   39026,39136,39246,39356,39465,39574,39682,39790,
+   39898,40005,40112,40219,40325,40431,40537,40642,
+   40747,40851,40955,41059,41163,41266,41369,41471,
+   41573,41675,41777,41878,41979,42079,42179,42279,
+   42379,42478,42577,42676,42775,42873,42971,43068,
+   43165,43262,43359,43456,43552,43648,43743,43839,
+   43934,44028,44123,44217,44311,44405,44499,44592,
+   44685,44778,44870,44962,45054,45146,45238,45329,
+   45420,45511,45601,45692,45782,45872,45961,46051,
+   46140,46229,46318,46406,46494,46583,46670,46758,
+   46846,46933,47020,47107,47193,47280,47366,47452,
+   47538,47623,47709,47794,47879,47964,48048,48133,
+   48217,48301,48385,48468,48552,48635,48718,48801,
+   48884,48966,49048,49131,49213,49294,49376,49458,
+   49539,49620,49701,49782,49862,49943,50023,50103,
+   50183,50263,50342,50422,50501,50580,50659,50738,
+   50816,50895,50973,51051,51129,51207,51285,51362,
+   51439,51517,51594,51671,51747,51824,51900,51977,
+   52053,52129,52205,52280,52356,52432,52507,52582,
+   52657,52732,52807,52881,52956,53030,53104,53178,
+   53252,53326,53400,53473,53546,53620,53693,53766,
+   53839,53911,53984,54056,54129,54201,54273,54345,
+   54417,54489,54560,54632,54703,54774,54845,54916,
+   54987,55058,55129,55199,55269,55340,55410,55480,
+   55550,55620,55689,55759,55828,55898,55967,56036,
+   56105,56174,56243,56311,56380,56448,56517,56585,
+   56653,56721,56789,56857,56924,56992,57059,57127,
+   57194,57261,57328,57395,57462,57529,57595,57662,
+   57728,57795,57861,57927,57993,58059,58125,58191,
+   58256,58322,58387,58453,58518,58583,58648,58713,
+   58778,58843,58908,58972,59037,59101,59165,59230,
+   59294,59358,59422,59486,59549,59613,59677,59740,
+   59804,59867,59930,59993,60056,60119,60182,60245,
+   60308,60370,60433,60495,60558,60620,60682,60744,
+   60806,60868,60930,60992,61054,61115,61177,61238,
+   61300,61361,61422,61483,61544,61605,61666,61727,
+   61788,61848,61909,61969,62030,62090,62150,62211,
+   62271,62331,62391,62450,62510,62570,62630,62689,
+   62749,62808,62867,62927,62986,63045,63104,63163,
+   63222,63281,63340,63398,63457,63515,63574,63632,
+   63691,63749,63807,63865,63923,63981,64039,64097,
+   64155,64212,64270,64328,64385,64443,64500,64557,
+   64614,64672,64729,64786,64843,64900,64956,65013,
+   65070,65126,65183,65239,65296,65352,65409,65465
+};
+
+const png_byte png_sRGB_delta[512] =
+{
+   207,201,158,129,113,100,90,82,77,72,68,64,61,59,56,54,
+   52,50,49,47,46,45,43,42,41,40,39,39,38,37,36,36,
+   35,34,34,33,33,32,32,31,31,30,30,30,29,29,28,28,
+   28,27,27,27,27,26,26,26,25,25,25,25,24,24,24,24,
+   23,23,23,23,23,22,22,22,22,22,22,21,21,21,21,21,
+   21,20,20,20,20,20,20,20,20,19,19,19,19,19,19,19,
+   19,18,18,18,18,18,18,18,18,18,18,17,17,17,17,17,
+   17,17,17,17,17,17,16,16,16,16,16,16,16,16,16,16,
+   16,16,16,16,15,15,15,15,15,15,15,15,15,15,15,15,
+   15,15,15,15,14,14,14,14,14,14,14,14,14,14,14,14,
+   14,14,14,14,14,14,14,13,13,13,13,13,13,13,13,13,
+   13,13,13,13,13,13,13,13,13,13,13,13,13,13,12,12,
+   12,12,12,12,12,12,12,12,12,12,12,12,12,12,12,12,
+   12,12,12,12,12,12,12,12,12,12,12,12,11,11,11,11,
+   11,11,11,11,11,11,11,11,11,11,11,11,11,11,11,11,
+   11,11,11,11,11,11,11,11,11,11,11,11,11,11,11,11,
+   11,10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,
+   10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,
+   10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,
+   10,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,7,7,7,7,7,7,7,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
+};
+#endif /* SIMPLIFIED READ/WRITE sRGB support */
+
+/* SIMPLIFIED READ/WRITE SUPPORT */
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) ||\
+   defined(PNG_SIMPLIFIED_WRITE_SUPPORTED)
+static int
+png_image_free_function(png_voidp argument)
+{
+   png_imagep image = png_voidcast(png_imagep, argument);
+   png_controlp cp = image->opaque;
+   png_control c;
+
+   /* Double check that we have a png_ptr - it should be impossible to get here
+    * without one.
+    */
+   if (cp->png_ptr == NULL)
+      return 0;
+
+   /* First free any data held in the control structure. */
+#  ifdef PNG_STDIO_SUPPORTED
+      if (cp->owned_file)
+      {
+         FILE *fp = png_voidcast(FILE*, cp->png_ptr->io_ptr);
+         cp->owned_file = 0;
+
+         /* Ignore errors here. */
+         if (fp != NULL)
+         {
+            cp->png_ptr->io_ptr = NULL;
+            (void)fclose(fp);
+         }
+      }
+#  endif
+
+   /* Copy the control structure so that the original, allocated, version can be
+    * safely freed.  Notice that a png_error here stops the remainder of the
+    * cleanup, but this is probably fine because that would indicate bad memory
+    * problems anyway.
+    */
+   c = *cp;
+   image->opaque = &c;
+   png_free(c.png_ptr, cp);
+
+   /* Then the structures, calling the correct API. */
+   if (c.for_write)
+   {
+#     ifdef PNG_SIMPLIFIED_WRITE_SUPPORTED
+         png_destroy_write_struct(&c.png_ptr, &c.info_ptr);
+#     else
+         png_error(c.png_ptr, "simplified write not supported");
+#     endif
+   }
+   else
+   {
+#     ifdef PNG_SIMPLIFIED_READ_SUPPORTED
+         png_destroy_read_struct(&c.png_ptr, &c.info_ptr, NULL);
+#     else
+         png_error(c.png_ptr, "simplified read not supported");
+#     endif
+   }
+
+   /* Success. */
+   return 1;
+}
+
+void PNGAPI
+png_image_free(png_imagep image)
+{
+   /* Safely call the real function, but only if doing so is safe at this point
+    * (if not inside an error handling context).  Otherwise assume
+    * png_safe_execute will call this API after the return.
+    */
+   if (image != NULL && image->opaque != NULL &&
+      image->opaque->error_buf == NULL)
+   {
+      /* Ignore errors here: */
+      (void)png_safe_execute(image, png_image_free_function, image);
+      image->opaque = NULL;
+   }
+}
+
+int /* PRIVATE */
+png_image_error(png_imagep image, png_const_charp error_message)
+{
+   /* Utility to log an error. */
+   png_safecat(image->message, (sizeof image->message), 0, error_message);
+   image->warning_or_error |= PNG_IMAGE_ERROR;
+   png_image_free(image);
+   return 0;
+}
+
+#endif /* SIMPLIFIED READ/WRITE */
+#endif /* defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED) */
diff --git a/ml/dlib/dlib/external/libpng/png.h b/ml/dlib/dlib/external/libpng/png.h
new file mode 100644
index 000000000..527392738
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/png.h
@@ -0,0 +1,3319 @@
+
+/* png.h - header file for PNG reference library
+ *
+ * libpng version 1.6.7 - November 14, 2013
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license (See LICENSE, below)
+ *
+ * Authors and maintainers:
+ *   libpng versions 0.71, May 1995, through 0.88, January 1996: Guy Schalnat
+ *   libpng versions 0.89c, June 1996, through 0.96, May 1997: Andreas Dilger
+ *   libpng versions 0.97, January 1998, through 1.6.7 - November 14, 2013: Glenn
+ *   See also "Contributing Authors", below.
+ *
+ * Note about libpng version numbers:
+ *
+ *   Due to various miscommunications, unforeseen code incompatibilities
+ *   and occasional factors outside the authors' control, version numbering
+ *   on the library has not always been consistent and straightforward.
+ *   The following table summarizes matters since version 0.89c, which was
+ *   the first widely used release:
+ *
+ *    source                 png.h  png.h  shared-lib
+ *    version                string   int  version
+ *    -------                ------ -----  ----------
+ *    0.89c "1.0 beta 3"     0.89      89  1.0.89
+ *    0.90  "1.0 beta 4"     0.90      90  0.90  [should have been 2.0.90]
+ *    0.95  "1.0 beta 5"     0.95      95  0.95  [should have been 2.0.95]
+ *    0.96  "1.0 beta 6"     0.96      96  0.96  [should have been 2.0.96]
+ *    0.97b "1.00.97 beta 7" 1.00.97   97  1.0.1 [should have been 2.0.97]
+ *    0.97c                  0.97      97  2.0.97
+ *    0.98                   0.98      98  2.0.98
+ *    0.99                   0.99      98  2.0.99
+ *    0.99a-m                0.99      99  2.0.99
+ *    1.00                   1.00     100  2.1.0 [100 should be 10000]
+ *    1.0.0      (from here on, the   100  2.1.0 [100 should be 10000]
+ *    1.0.1       png.h string is   10001  2.1.0
+ *    1.0.1a-e    identical to the  10002  from here on, the shared library
+ *    1.0.2       source version)   10002  is 2.V where V is the source code
+ *    1.0.2a-b                      10003  version, except as noted.
+ *    1.0.3                         10003
+ *    1.0.3a-d                      10004
+ *    1.0.4                         10004
+ *    1.0.4a-f                      10005
+ *    1.0.5 (+ 2 patches)           10005
+ *    1.0.5a-d                      10006
+ *    1.0.5e-r                      10100 (not source compatible)
+ *    1.0.5s-v                      10006 (not binary compatible)
+ *    1.0.6 (+ 3 patches)           10006 (still binary incompatible)
+ *    1.0.6d-f                      10007 (still binary incompatible)
+ *    1.0.6g                        10007
+ *    1.0.6h                        10007  10.6h (testing xy.z so-numbering)
+ *    1.0.6i                        10007  10.6i
+ *    1.0.6j                        10007  2.1.0.6j (incompatible with 1.0.0)
+ *    1.0.7beta11-14        DLLNUM  10007  2.1.0.7beta11-14 (binary compatible)
+ *    1.0.7beta15-18           1    10007  2.1.0.7beta15-18 (binary compatible)
+ *    1.0.7rc1-2               1    10007  2.1.0.7rc1-2 (binary compatible)
+ *    1.0.7                    1    10007  (still compatible)
+ *    1.0.8beta1-4             1    10008  2.1.0.8beta1-4
+ *    1.0.8rc1                 1    10008  2.1.0.8rc1
+ *    1.0.8                    1    10008  2.1.0.8
+ *    1.0.9beta1-6             1    10009  2.1.0.9beta1-6
+ *    1.0.9rc1                 1    10009  2.1.0.9rc1
+ *    1.0.9beta7-10            1    10009  2.1.0.9beta7-10
+ *    1.0.9rc2                 1    10009  2.1.0.9rc2
+ *    1.0.9                    1    10009  2.1.0.9
+ *    1.0.10beta1              1    10010  2.1.0.10beta1
+ *    1.0.10rc1                1    10010  2.1.0.10rc1
+ *    1.0.10                   1    10010  2.1.0.10
+ *    1.0.11beta1-3            1    10011  2.1.0.11beta1-3
+ *    1.0.11rc1                1    10011  2.1.0.11rc1
+ *    1.0.11                   1    10011  2.1.0.11
+ *    1.0.12beta1-2            2    10012  2.1.0.12beta1-2
+ *    1.0.12rc1                2    10012  2.1.0.12rc1
+ *    1.0.12                   2    10012  2.1.0.12
+ *    1.1.0a-f                 -    10100  2.1.1.0a-f (branch abandoned)
+ *    1.2.0beta1-2             2    10200  2.1.2.0beta1-2
+ *    1.2.0beta3-5             3    10200  3.1.2.0beta3-5
+ *    1.2.0rc1                 3    10200  3.1.2.0rc1
+ *    1.2.0                    3    10200  3.1.2.0
+ *    1.2.1beta1-4             3    10201  3.1.2.1beta1-4
+ *    1.2.1rc1-2               3    10201  3.1.2.1rc1-2
+ *    1.2.1                    3    10201  3.1.2.1
+ *    1.2.2beta1-6            12    10202  12.so.0.1.2.2beta1-6
+ *    1.0.13beta1             10    10013  10.so.0.1.0.13beta1
+ *    1.0.13rc1               10    10013  10.so.0.1.0.13rc1
+ *    1.2.2rc1                12    10202  12.so.0.1.2.2rc1
+ *    1.0.13                  10    10013  10.so.0.1.0.13
+ *    1.2.2                   12    10202  12.so.0.1.2.2
+ *    1.2.3rc1-6              12    10203  12.so.0.1.2.3rc1-6
+ *    1.2.3                   12    10203  12.so.0.1.2.3
+ *    1.2.4beta1-3            13    10204  12.so.0.1.2.4beta1-3
+ *    1.0.14rc1               13    10014  10.so.0.1.0.14rc1
+ *    1.2.4rc1                13    10204  12.so.0.1.2.4rc1
+ *    1.0.14                  10    10014  10.so.0.1.0.14
+ *    1.2.4                   13    10204  12.so.0.1.2.4
+ *    1.2.5beta1-2            13    10205  12.so.0.1.2.5beta1-2
+ *    1.0.15rc1-3             10    10015  10.so.0.1.0.15rc1-3
+ *    1.2.5rc1-3              13    10205  12.so.0.1.2.5rc1-3
+ *    1.0.15                  10    10015  10.so.0.1.0.15
+ *    1.2.5                   13    10205  12.so.0.1.2.5
+ *    1.2.6beta1-4            13    10206  12.so.0.1.2.6beta1-4
+ *    1.0.16                  10    10016  10.so.0.1.0.16
+ *    1.2.6                   13    10206  12.so.0.1.2.6
+ *    1.2.7beta1-2            13    10207  12.so.0.1.2.7beta1-2
+ *    1.0.17rc1               10    10017  12.so.0.1.0.17rc1
+ *    1.2.7rc1                13    10207  12.so.0.1.2.7rc1
+ *    1.0.17                  10    10017  12.so.0.1.0.17
+ *    1.2.7                   13    10207  12.so.0.1.2.7
+ *    1.2.8beta1-5            13    10208  12.so.0.1.2.8beta1-5
+ *    1.0.18rc1-5             10    10018  12.so.0.1.0.18rc1-5
+ *    1.2.8rc1-5              13    10208  12.so.0.1.2.8rc1-5
+ *    1.0.18                  10    10018  12.so.0.1.0.18
+ *    1.2.8                   13    10208  12.so.0.1.2.8
+ *    1.2.9beta1-3            13    10209  12.so.0.1.2.9beta1-3
+ *    1.2.9beta4-11           13    10209  12.so.0.9[.0]
+ *    1.2.9rc1                13    10209  12.so.0.9[.0]
+ *    1.2.9                   13    10209  12.so.0.9[.0]
+ *    1.2.10beta1-7           13    10210  12.so.0.10[.0]
+ *    1.2.10rc1-2             13    10210  12.so.0.10[.0]
+ *    1.2.10                  13    10210  12.so.0.10[.0]
+ *    1.4.0beta1-5            14    10400  14.so.0.0[.0]
+ *    1.2.11beta1-4           13    10211  12.so.0.11[.0]
+ *    1.4.0beta7-8            14    10400  14.so.0.0[.0]
+ *    1.2.11                  13    10211  12.so.0.11[.0]
+ *    1.2.12                  13    10212  12.so.0.12[.0]
+ *    1.4.0beta9-14           14    10400  14.so.0.0[.0]
+ *    1.2.13                  13    10213  12.so.0.13[.0]
+ *    1.4.0beta15-36          14    10400  14.so.0.0[.0]
+ *    1.4.0beta37-87          14    10400  14.so.14.0[.0]
+ *    1.4.0rc01               14    10400  14.so.14.0[.0]
+ *    1.4.0beta88-109         14    10400  14.so.14.0[.0]
+ *    1.4.0rc02-08            14    10400  14.so.14.0[.0]
+ *    1.4.0                   14    10400  14.so.14.0[.0]
+ *    1.4.1beta01-03          14    10401  14.so.14.1[.0]
+ *    1.4.1rc01               14    10401  14.so.14.1[.0]
+ *    1.4.1beta04-12          14    10401  14.so.14.1[.0]
+ *    1.4.1                   14    10401  14.so.14.1[.0]
+ *    1.4.2                   14    10402  14.so.14.2[.0]
+ *    1.4.3                   14    10403  14.so.14.3[.0]
+ *    1.4.4                   14    10404  14.so.14.4[.0]
+ *    1.5.0beta01-58          15    10500  15.so.15.0[.0]
+ *    1.5.0rc01-07            15    10500  15.so.15.0[.0]
+ *    1.5.0                   15    10500  15.so.15.0[.0]
+ *    1.5.1beta01-11          15    10501  15.so.15.1[.0]
+ *    1.5.1rc01-02            15    10501  15.so.15.1[.0]
+ *    1.5.1                   15    10501  15.so.15.1[.0]
+ *    1.5.2beta01-03          15    10502  15.so.15.2[.0]
+ *    1.5.2rc01-03            15    10502  15.so.15.2[.0]
+ *    1.5.2                   15    10502  15.so.15.2[.0]
+ *    1.5.3beta01-10          15    10503  15.so.15.3[.0]
+ *    1.5.3rc01-02            15    10503  15.so.15.3[.0]
+ *    1.5.3beta11             15    10503  15.so.15.3[.0]
+ *    1.5.3 [omitted]
+ *    1.5.4beta01-08          15    10504  15.so.15.4[.0]
+ *    1.5.4rc01               15    10504  15.so.15.4[.0]
+ *    1.5.4                   15    10504  15.so.15.4[.0]
+ *    1.5.5beta01-08          15    10505  15.so.15.5[.0]
+ *    1.5.5rc01               15    10505  15.so.15.5[.0]
+ *    1.5.5                   15    10505  15.so.15.5[.0]
+ *    1.5.6beta01-07          15    10506  15.so.15.6[.0]
+ *    1.5.6rc01-03            15    10506  15.so.15.6[.0]
+ *    1.5.6                   15    10506  15.so.15.6[.0]
+ *    1.5.7beta01-05          15    10507  15.so.15.7[.0]
+ *    1.5.7rc01-03            15    10507  15.so.15.7[.0]
+ *    1.5.7                   15    10507  15.so.15.7[.0]
+ *    1.6.0beta01-40          16    10600  16.so.16.0[.0]
+ *    1.6.0rc01-08            16    10600  16.so.16.0[.0]
+ *    1.6.0                   16    10600  16.so.16.0[.0]
+ *    1.6.1beta01-09          16    10601  16.so.16.1[.0]
+ *    1.6.1rc01               16    10601  16.so.16.1[.0]
+ *    1.6.1                   16    10601  16.so.16.1[.0]
+ *    1.6.2beta01             16    10602  16.so.16.2[.0]
+ *    1.6.2rc01-06            16    10602  16.so.16.2[.0]
+ *    1.6.2                   16    10602  16.so.16.2[.0]
+ *    1.6.3beta01-11          16    10603  16.so.16.3[.0]
+ *    1.6.3rc01               16    10603  16.so.16.3[.0]
+ *    1.6.3                   16    10603  16.so.16.3[.0]
+ *    1.6.4beta01-02          16    10604  16.so.16.4[.0]
+ *    1.6.4rc01               16    10604  16.so.16.4[.0]
+ *    1.6.4                   16    10604  16.so.16.4[.0]
+ *    1.6.5                   16    10605  16.so.16.5[.0]
+ *    1.6.6                   16    10606  16.so.16.6[.0]
+ *    1.6.7beta01-04          16    10607  16.so.16.7[.0]
+ *    1.6.7rc01-02            16    10607  16.so.16.7[.0]
+ *    1.6.7                   16    10607  16.so.16.7[.0]
+ *
+ *   Henceforth the source version will match the shared-library major
+ *   and minor numbers; the shared-library major version number will be
+ *   used for changes in backward compatibility, as it is intended.  The
+ *   PNG_LIBPNG_VER macro, which is not used within libpng but is available
+ *   for applications, is an unsigned integer of the form xyyzz corresponding
+ *   to the source version x.y.z (leading zeros in y and z).  Beta versions
+ *   were given the previous public release number plus a letter, until
+ *   version 1.0.6j; from then on they were given the upcoming public
+ *   release number plus "betaNN" or "rcNN".
+ *
+ *   Binary incompatibility exists only when applications make direct access
+ *   to the info_ptr or png_ptr members through png.h, and the compiled
+ *   application is loaded with a different version of the library.
+ *
+ *   DLLNUM will change each time there are forward or backward changes
+ *   in binary compatibility (e.g., when a new feature is added).
+ *
+ * See libpng-manual.txt or libpng.3 for more information.  The PNG
+ * specification is available as a W3C Recommendation and as an ISO
+ * Specification, <http://www.w3.org/TR/2003/REC-PNG-20031110/
+ */
+
+/*
+ * COPYRIGHT NOTICE, DISCLAIMER, and LICENSE:
+ *
+ * If you modify libpng you may insert additional notices immediately following
+ * this sentence.
+ *
+ * This code is released under the libpng license.
+ *
+ * libpng versions 1.2.6, August 15, 2004, through 1.6.7, November 14, 2013, are
+ * Copyright (c) 2004, 2006-2013 Glenn Randers-Pehrson, and are
+ * distributed according to the same disclaimer and license as libpng-1.2.5
+ * with the following individual added to the list of Contributing Authors:
+ *
+ *    Cosmin Truta
+ *
+ * libpng versions 1.0.7, July 1, 2000, through 1.2.5, October 3, 2002, are
+ * Copyright (c) 2000-2002 Glenn Randers-Pehrson, and are
+ * distributed according to the same disclaimer and license as libpng-1.0.6
+ * with the following individuals added to the list of Contributing Authors:
+ *
+ *    Simon-Pierre Cadieux
+ *    Eric S. Raymond
+ *    Gilles Vollant
+ *
+ * and with the following additions to the disclaimer:
+ *
+ *    There is no warranty against interference with your enjoyment of the
+ *    library or against infringement.  There is no warranty that our
+ *    efforts or the library will fulfill any of your particular purposes
+ *    or needs.  This library is provided with all faults, and the entire
+ *    risk of satisfactory quality, performance, accuracy, and effort is with
+ *    the user.
+ *
+ * libpng versions 0.97, January 1998, through 1.0.6, March 20, 2000, are
+ * Copyright (c) 1998, 1999, 2000 Glenn Randers-Pehrson, and are
+ * distributed according to the same disclaimer and license as libpng-0.96,
+ * with the following individuals added to the list of Contributing Authors:
+ *
+ *    Tom Lane
+ *    Glenn Randers-Pehrson
+ *    Willem van Schaik
+ *
+ * libpng versions 0.89, June 1996, through 0.96, May 1997, are
+ * Copyright (c) 1996, 1997 Andreas Dilger
+ * Distributed according to the same disclaimer and license as libpng-0.88,
+ * with the following individuals added to the list of Contributing Authors:
+ *
+ *    John Bowler
+ *    Kevin Bracey
+ *    Sam Bushell
+ *    Magnus Holmgren
+ *    Greg Roelofs
+ *    Tom Tanner
+ *
+ * libpng versions 0.5, May 1995, through 0.88, January 1996, are
+ * Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.
+ *
+ * For the purposes of this copyright and license, "Contributing Authors"
+ * is defined as the following set of individuals:
+ *
+ *    Andreas Dilger
+ *    Dave Martindale
+ *    Guy Eric Schalnat
+ *    Paul Schmidt
+ *    Tim Wegner
+ *
+ * The PNG Reference Library is supplied "AS IS".  The Contributing Authors
+ * and Group 42, Inc. disclaim all warranties, expressed or implied,
+ * including, without limitation, the warranties of merchantability and of
+ * fitness for any purpose.  The Contributing Authors and Group 42, Inc.
+ * assume no liability for direct, indirect, incidental, special, exemplary,
+ * or consequential damages, which may result from the use of the PNG
+ * Reference Library, even if advised of the possibility of such damage.
+ *
+ * Permission is hereby granted to use, copy, modify, and distribute this
+ * source code, or portions hereof, for any purpose, without fee, subject
+ * to the following restrictions:
+ *
+ *   1. The origin of this source code must not be misrepresented.
+ *
+ *   2. Altered versions must be plainly marked as such and must not
+ *      be misrepresented as being the original source.
+ *
+ *   3. This Copyright notice may not be removed or altered from
+ *      any source or altered source distribution.
+ *
+ * The Contributing Authors and Group 42, Inc. specifically permit, without
+ * fee, and encourage the use of this source code as a component to
+ * supporting the PNG file format in commercial products.  If you use this
+ * source code in a product, acknowledgment is not required but would be
+ * appreciated.
+ */
+
+/*
+ * A "png_get_copyright" function is available, for convenient use in "about"
+ * boxes and the like:
+ *
+ *     printf("%s", png_get_copyright(NULL));
+ *
+ * Also, the PNG logo (in PNG format, of course) is supplied in the
+ * files "pngbar.png" and "pngbar.jpg (88x31) and "pngnow.png" (98x31).
+ */
+
+/*
+ * Libpng is OSI Certified Open Source Software.  OSI Certified is a
+ * certification mark of the Open Source Initiative.
+ */
+
+/*
+ * The contributing authors would like to thank all those who helped
+ * with testing, bug fixes, and patience.  This wouldn't have been
+ * possible without all of you.
+ *
+ * Thanks to Frank J. T. Wojcik for helping with the documentation.
+ */
+
+/*
+ * Y2K compliance in libpng:
+ * =========================
+ *
+ *    November 14, 2013
+ *
+ *    Since the PNG Development group is an ad-hoc body, we can't make
+ *    an official declaration.
+ *
+ *    This is your unofficial assurance that libpng from version 0.71 and
+ *    upward through 1.6.7 are Y2K compliant.  It is my belief that
+ *    earlier versions were also Y2K compliant.
+ *
+ *    Libpng only has two year fields.  One is a 2-byte unsigned integer
+ *    that will hold years up to 65535.  The other, which is deprecated,
+ *    holds the date in text format, and will hold years up to 9999.
+ *
+ *    The integer is
+ *        "png_uint_16 year" in png_time_struct.
+ *
+ *    The string is
+ *        "char time_buffer[29]" in png_struct.  This is no longer used
+ *    in libpng-1.6.x and will be removed from libpng-1.7.0.
+ *
+ *    There are seven time-related functions:
+ *        png.c: png_convert_to_rfc_1123_buffer() in png.c
+ *          (formerly png_convert_to_rfc_1123() prior to libpng-1.5.x and
+ *          png_convert_to_rfc_1152() in error prior to libpng-0.98)
+ *        png_convert_from_struct_tm() in pngwrite.c, called in pngwrite.c
+ *        png_convert_from_time_t() in pngwrite.c
+ *        png_get_tIME() in pngget.c
+ *        png_handle_tIME() in pngrutil.c, called in pngread.c
+ *        png_set_tIME() in pngset.c
+ *        png_write_tIME() in pngwutil.c, called in pngwrite.c
+ *
+ *    All handle dates properly in a Y2K environment.  The
+ *    png_convert_from_time_t() function calls gmtime() to convert from system
+ *    clock time, which returns (year - 1900), which we properly convert to
+ *    the full 4-digit year.  There is a possibility that libpng applications
+ *    are not passing 4-digit years into the png_convert_to_rfc_1123_buffer()
+ *    function, or that they are incorrectly passing only a 2-digit year
+ *    instead of "year - 1900" into the png_convert_from_struct_tm() function,
+ *    but this is not under our control.  The libpng documentation has always
+ *    stated that it works with 4-digit years, and the APIs have been
+ *    documented as such.
+ *
+ *    The tIME chunk itself is also Y2K compliant.  It uses a 2-byte unsigned
+ *    integer to hold the year, and can hold years as large as 65535.
+ *
+ *    zlib, upon which libpng depends, is also Y2K compliant.  It contains
+ *    no date-related code.
+ *
+ *       Glenn Randers-Pehrson
+ *       libpng maintainer
+ *       PNG Development Group
+ */
+
+#ifndef PNG_H
+#define PNG_H
+
+/* This is not the place to learn how to use libpng. The file libpng-manual.txt
+ * describes how to use libpng, and the file example.c summarizes it
+ * with some code on which to build.  This file is useful for looking
+ * at the actual function definitions and structure components.
+ *
+ * If you just need to read a PNG file and don't want to read the documentation
+ * skip to the end of this file and read the section entitled 'simplified API'.
+ */
+
+/* Version information for png.h - this should match the version in png.c */
+#define PNG_LIBPNG_VER_STRING "1.6.7"
+#define PNG_HEADER_VERSION_STRING \
+     " libpng version 1.6.7 - November 14, 2013\n"
+
+#define PNG_LIBPNG_VER_SONUM   16
+#define PNG_LIBPNG_VER_DLLNUM  16
+
+/* These should match the first 3 components of PNG_LIBPNG_VER_STRING: */
+#define PNG_LIBPNG_VER_MAJOR   1
+#define PNG_LIBPNG_VER_MINOR   6
+#define PNG_LIBPNG_VER_RELEASE 7
+
+/* This should match the numeric part of the final component of
+ * PNG_LIBPNG_VER_STRING, omitting any leading zero:
+ */
+
+#define PNG_LIBPNG_VER_BUILD  0
+
+/* Release Status */
+#define PNG_LIBPNG_BUILD_ALPHA    1
+#define PNG_LIBPNG_BUILD_BETA     2
+#define PNG_LIBPNG_BUILD_RC       3
+#define PNG_LIBPNG_BUILD_STABLE   4
+#define PNG_LIBPNG_BUILD_RELEASE_STATUS_MASK 7
+
+/* Release-Specific Flags */
+#define PNG_LIBPNG_BUILD_PATCH    8 /* Can be OR'ed with
+                                       PNG_LIBPNG_BUILD_STABLE only */
+#define PNG_LIBPNG_BUILD_PRIVATE 16 /* Cannot be OR'ed with
+                                       PNG_LIBPNG_BUILD_SPECIAL */
+#define PNG_LIBPNG_BUILD_SPECIAL 32 /* Cannot be OR'ed with
+                                       PNG_LIBPNG_BUILD_PRIVATE */
+
+#define PNG_LIBPNG_BUILD_BASE_TYPE PNG_LIBPNG_BUILD_STABLE
+
+/* Careful here.  At one time, Guy wanted to use 082, but that would be octal.
+ * We must not include leading zeros.
+ * Versions 0.7 through 1.0.0 were in the range 0 to 100 here (only
+ * version 1.0.0 was mis-numbered 100 instead of 10000).  From
+ * version 1.0.1 it's    xxyyzz, where x=major, y=minor, z=release
+ */
+#define PNG_LIBPNG_VER 10607 /* 1.6.7 */
+
+/* Library configuration: these options cannot be changed after
+ * the library has been built.
+ */
+#ifndef PNGLCONF_H
+    /* If pnglibconf.h is missing, you can
+     * copy scripts/pnglibconf.h.prebuilt to pnglibconf.h
+     */
+#   include "pnglibconf.h"
+#endif
+
+#ifndef PNG_VERSION_INFO_ONLY
+   /* Machine specific configuration. */
+#  include "pngconf.h"
+#endif
+
+/*
+ * Added at libpng-1.2.8
+ *
+ * Ref MSDN: Private as priority over Special
+ * VS_FF_PRIVATEBUILD File *was not* built using standard release
+ * procedures. If this value is given, the StringFileInfo block must
+ * contain a PrivateBuild string.
+ *
+ * VS_FF_SPECIALBUILD File *was* built by the original company using
+ * standard release procedures but is a variation of the standard
+ * file of the same version number. If this value is given, the
+ * StringFileInfo block must contain a SpecialBuild string.
+ */
+
+#ifdef PNG_USER_PRIVATEBUILD /* From pnglibconf.h */
+#  define PNG_LIBPNG_BUILD_TYPE \
+       (PNG_LIBPNG_BUILD_BASE_TYPE | PNG_LIBPNG_BUILD_PRIVATE)
+#else
+#  ifdef PNG_LIBPNG_SPECIALBUILD
+#    define PNG_LIBPNG_BUILD_TYPE \
+         (PNG_LIBPNG_BUILD_BASE_TYPE | PNG_LIBPNG_BUILD_SPECIAL)
+#  else
+#    define PNG_LIBPNG_BUILD_TYPE (PNG_LIBPNG_BUILD_BASE_TYPE)
+#  endif
+#endif
+
+#ifndef PNG_VERSION_INFO_ONLY
+
+/* Inhibit C++ name-mangling for libpng functions but not for system calls. */
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+/* Version information for C files, stored in png.c.  This had better match
+ * the version above.
+ */
+#define png_libpng_ver png_get_header_ver(NULL)
+
+/* This file is arranged in several sections:
+ *
+ * 1. Any configuration options that can be specified by for the application
+ *    code when it is built.  (Build time configuration is in pnglibconf.h)
+ * 2. Type definitions (base types are defined in pngconf.h), structure
+ *    definitions.
+ * 3. Exported library functions.
+ * 4. Simplified API.
+ *
+ * The library source code has additional files (principally pngpriv.h) that
+ * allow configuration of the library.
+ */
+/* Section 1: run time configuration
+ * See pnglibconf.h for build time configuration
+ *
+ * Run time configuration allows the application to choose between
+ * implementations of certain arithmetic APIs.  The default is set
+ * at build time and recorded in pnglibconf.h, but it is safe to
+ * override these (and only these) settings.  Note that this won't
+ * change what the library does, only application code, and the
+ * settings can (and probably should) be made on a per-file basis
+ * by setting the #defines before including png.h
+ *
+ * Use macros to read integers from PNG data or use the exported
+ * functions?
+ *   PNG_USE_READ_MACROS: use the macros (see below)  Note that
+ *     the macros evaluate their argument multiple times.
+ *   PNG_NO_USE_READ_MACROS: call the relevant library function.
+ *
+ * Use the alternative algorithm for compositing alpha samples that
+ * does not use division?
+ *   PNG_READ_COMPOSITE_NODIV_SUPPORTED: use the 'no division'
+ *      algorithm.
+ *   PNG_NO_READ_COMPOSITE_NODIV: use the 'division' algorithm.
+ *
+ * How to handle benign errors if PNG_ALLOW_BENIGN_ERRORS is
+ * false?
+ *   PNG_ALLOW_BENIGN_ERRORS: map calls to the benign error
+ *      APIs to png_warning.
+ * Otherwise the calls are mapped to png_error.
+ */
+
+/* Section 2: type definitions, including structures and compile time
+ * constants.
+ * See pngconf.h for base types that vary by machine/system
+ */
+
+/* This triggers a compiler error in png.c, if png.c and png.h
+ * do not agree upon the version number.
+ */
+typedef char* png_libpng_version_1_6_7;
+
+/* Basic control structions.  Read libpng-manual.txt or libpng.3 for more info.
+ *
+ * png_struct is the cache of information used while reading or writing a single
+ * PNG file.  One of these is always required, although the simplified API
+ * (below) hides the creation and destruction of it.
+ */
+typedef struct png_struct_def png_struct;
+typedef const png_struct * png_const_structp;
+typedef png_struct * png_structp;
+typedef png_struct * * png_structpp;
+
+/* png_info contains information read from or to be written to a PNG file.  One
+ * or more of these must exist while reading or creating a PNG file.  The
+ * information is not used by libpng during read but is used to control what
+ * gets written when a PNG file is created.  "png_get_" function calls read
+ * information during read and "png_set_" functions calls write information
+ * when creating a PNG.
+ * been moved into a separate header file that is not accessible to
+ * applications.  Read libpng-manual.txt or libpng.3 for more info.
+ */
+typedef struct png_info_def png_info;
+typedef png_info * png_infop;
+typedef const png_info * png_const_infop;
+typedef png_info * * png_infopp;
+
+/* Types with names ending 'p' are pointer types.  The corresponding types with
+ * names ending 'rp' are identical pointer types except that the pointer is
+ * marked 'restrict', which means that it is the only pointer to the object
+ * passed to the function.  Applications should not use the 'restrict' types;
+ * it is always valid to pass 'p' to a pointer with a function argument of the
+ * corresponding 'rp' type.  Different compilers have different rules with
+ * regard to type matching in the presence of 'restrict'.  For backward
+ * compatibility libpng callbacks never have 'restrict' in their parameters and,
+ * consequentially, writing portable application code is extremely difficult if
+ * an attempt is made to use 'restrict'.
+ */
+typedef png_struct * PNG_RESTRICT png_structrp;
+typedef const png_struct * PNG_RESTRICT png_const_structrp;
+typedef png_info * PNG_RESTRICT png_inforp;
+typedef const png_info * PNG_RESTRICT png_const_inforp;
+
+/* Three color definitions.  The order of the red, green, and blue, (and the
+ * exact size) is not important, although the size of the fields need to
+ * be png_byte or png_uint_16 (as defined below).
+ */
+typedef struct png_color_struct
+{
+   png_byte red;
+   png_byte green;
+   png_byte blue;
+} png_color;
+typedef png_color * png_colorp;
+typedef const png_color * png_const_colorp;
+typedef png_color * * png_colorpp;
+
+typedef struct png_color_16_struct
+{
+   png_byte index;    /* used for palette files */
+   png_uint_16 red;   /* for use in red green blue files */
+   png_uint_16 green;
+   png_uint_16 blue;
+   png_uint_16 gray;  /* for use in grayscale files */
+} png_color_16;
+typedef png_color_16 * png_color_16p;
+typedef const png_color_16 * png_const_color_16p;
+typedef png_color_16 * * png_color_16pp;
+
+typedef struct png_color_8_struct
+{
+   png_byte red;   /* for use in red green blue files */
+   png_byte green;
+   png_byte blue;
+   png_byte gray;  /* for use in grayscale files */
+   png_byte alpha; /* for alpha channel files */
+} png_color_8;
+typedef png_color_8 * png_color_8p;
+typedef const png_color_8 * png_const_color_8p;
+typedef png_color_8 * * png_color_8pp;
+
+/*
+ * The following two structures are used for the in-core representation
+ * of sPLT chunks.
+ */
+typedef struct png_sPLT_entry_struct
+{
+   png_uint_16 red;
+   png_uint_16 green;
+   png_uint_16 blue;
+   png_uint_16 alpha;
+   png_uint_16 frequency;
+} png_sPLT_entry;
+typedef png_sPLT_entry * png_sPLT_entryp;
+typedef const png_sPLT_entry * png_const_sPLT_entryp;
+typedef png_sPLT_entry * * png_sPLT_entrypp;
+
+/*  When the depth of the sPLT palette is 8 bits, the color and alpha samples
+ *  occupy the LSB of their respective members, and the MSB of each member
+ *  is zero-filled.  The frequency member always occupies the full 16 bits.
+ */
+
+typedef struct png_sPLT_struct
+{
+   png_charp name;           /* palette name */
+   png_byte depth;           /* depth of palette samples */
+   png_sPLT_entryp entries;  /* palette entries */
+   png_int_32 nentries;      /* number of palette entries */
+} png_sPLT_t;
+typedef png_sPLT_t * png_sPLT_tp;
+typedef const png_sPLT_t * png_const_sPLT_tp;
+typedef png_sPLT_t * * png_sPLT_tpp;
+
+#ifdef PNG_TEXT_SUPPORTED
+/* png_text holds the contents of a text/ztxt/itxt chunk in a PNG file,
+ * and whether that contents is compressed or not.  The "key" field
+ * points to a regular zero-terminated C string.  The "text" fields can be a
+ * regular C string, an empty string, or a NULL pointer.
+ * However, the structure returned by png_get_text() will always contain
+ * the "text" field as a regular zero-terminated C string (possibly
+ * empty), never a NULL pointer, so it can be safely used in printf() and
+ * other string-handling functions.  Note that the "itxt_length", "lang", and
+ * "lang_key" members of the structure only exist when the library is built
+ * with iTXt chunk support.  Prior to libpng-1.4.0 the library was built by
+ * default without iTXt support. Also note that when iTXt *is* supported,
+ * the "lang" and "lang_key" fields contain NULL pointers when the
+ * "compression" field contains * PNG_TEXT_COMPRESSION_NONE or
+ * PNG_TEXT_COMPRESSION_zTXt. Note that the "compression value" is not the
+ * same as what appears in the PNG tEXt/zTXt/iTXt chunk's "compression flag"
+ * which is always 0 or 1, or its "compression method" which is always 0.
+ */
+typedef struct png_text_struct
+{
+   int  compression;       /* compression value:
+                             -1: tEXt, none
+                              0: zTXt, deflate
+                              1: iTXt, none
+                              2: iTXt, deflate  */
+   png_charp key;          /* keyword, 1-79 character description of "text" */
+   png_charp text;         /* comment, may be an empty string (ie "")
+                              or a NULL pointer */
+   png_size_t text_length; /* length of the text string */
+   png_size_t itxt_length; /* length of the itxt string */
+   png_charp lang;         /* language code, 0-79 characters
+                              or a NULL pointer */
+   png_charp lang_key;     /* keyword translated UTF-8 string, 0 or more
+                              chars or a NULL pointer */
+} png_text;
+typedef png_text * png_textp;
+typedef const png_text * png_const_textp;
+typedef png_text * * png_textpp;
+#endif
+
+/* Supported compression types for text in PNG files (tEXt, and zTXt).
+ * The values of the PNG_TEXT_COMPRESSION_ defines should NOT be changed. */
+#define PNG_TEXT_COMPRESSION_NONE_WR -3
+#define PNG_TEXT_COMPRESSION_zTXt_WR -2
+#define PNG_TEXT_COMPRESSION_NONE    -1
+#define PNG_TEXT_COMPRESSION_zTXt     0
+#define PNG_ITXT_COMPRESSION_NONE     1
+#define PNG_ITXT_COMPRESSION_zTXt     2
+#define PNG_TEXT_COMPRESSION_LAST     3  /* Not a valid value */
+
+/* png_time is a way to hold the time in an machine independent way.
+ * Two conversions are provided, both from time_t and struct tm.  There
+ * is no portable way to convert to either of these structures, as far
+ * as I know.  If you know of a portable way, send it to me.  As a side
+ * note - PNG has always been Year 2000 compliant!
+ */
+typedef struct png_time_struct
+{
+   png_uint_16 year; /* full year, as in, 1995 */
+   png_byte month;   /* month of year, 1 - 12 */
+   png_byte day;     /* day of month, 1 - 31 */
+   png_byte hour;    /* hour of day, 0 - 23 */
+   png_byte minute;  /* minute of hour, 0 - 59 */
+   png_byte second;  /* second of minute, 0 - 60 (for leap seconds) */
+} png_time;
+typedef png_time * png_timep;
+typedef const png_time * png_const_timep;
+typedef png_time * * png_timepp;
+
+#if defined(PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED) ||\
+   defined(PNG_USER_CHUNKS_SUPPORTED)
+/* png_unknown_chunk is a structure to hold queued chunks for which there is
+ * no specific support.  The idea is that we can use this to queue
+ * up private chunks for output even though the library doesn't actually
+ * know about their semantics.
+ *
+ * The data in the structure is set by libpng on read and used on write.
+ */
+typedef struct png_unknown_chunk_t
+{
+    png_byte name[5]; /* Textual chunk name with '\0' terminator */
+    png_byte *data;   /* Data, should not be modified on read! */
+    png_size_t size;
+
+    /* On write 'location' must be set using the flag values listed below.
+     * Notice that on read it is set by libpng however the values stored have
+     * more bits set than are listed below.  Always treat the value as a
+     * bitmask.  On write set only one bit - setting multiple bits may cause the
+     * chunk to be written in multiple places.
+     */
+    png_byte location; /* mode of operation at read time */
+}
+png_unknown_chunk;
+
+typedef png_unknown_chunk * png_unknown_chunkp;
+typedef const png_unknown_chunk * png_const_unknown_chunkp;
+typedef png_unknown_chunk * * png_unknown_chunkpp;
+#endif
+
+/* Flag values for the unknown chunk location byte. */
+#define PNG_HAVE_IHDR  0x01
+#define PNG_HAVE_PLTE  0x02
+#define PNG_AFTER_IDAT 0x08
+
+/* Maximum positive integer used in PNG is (2^31)-1 */
+#define PNG_UINT_31_MAX ((png_uint_32)0x7fffffffL)
+#define PNG_UINT_32_MAX ((png_uint_32)(-1))
+#define PNG_SIZE_MAX ((png_size_t)(-1))
+
+/* These are constants for fixed point values encoded in the
+ * PNG specification manner (x100000)
+ */
+#define PNG_FP_1    100000
+#define PNG_FP_HALF  50000
+#define PNG_FP_MAX  ((png_fixed_point)0x7fffffffL)
+#define PNG_FP_MIN  (-PNG_FP_MAX)
+
+/* These describe the color_type field in png_info. */
+/* color type masks */
+#define PNG_COLOR_MASK_PALETTE    1
+#define PNG_COLOR_MASK_COLOR      2
+#define PNG_COLOR_MASK_ALPHA      4
+
+/* color types.  Note that not all combinations are legal */
+#define PNG_COLOR_TYPE_GRAY 0
+#define PNG_COLOR_TYPE_PALETTE  (PNG_COLOR_MASK_COLOR | PNG_COLOR_MASK_PALETTE)
+#define PNG_COLOR_TYPE_RGB        (PNG_COLOR_MASK_COLOR)
+#define PNG_COLOR_TYPE_RGB_ALPHA  (PNG_COLOR_MASK_COLOR | PNG_COLOR_MASK_ALPHA)
+#define PNG_COLOR_TYPE_GRAY_ALPHA (PNG_COLOR_MASK_ALPHA)
+/* aliases */
+#define PNG_COLOR_TYPE_RGBA  PNG_COLOR_TYPE_RGB_ALPHA
+#define PNG_COLOR_TYPE_GA  PNG_COLOR_TYPE_GRAY_ALPHA
+
+/* This is for compression type. PNG 1.0-1.2 only define the single type. */
+#define PNG_COMPRESSION_TYPE_BASE 0 /* Deflate method 8, 32K window */
+#define PNG_COMPRESSION_TYPE_DEFAULT PNG_COMPRESSION_TYPE_BASE
+
+/* This is for filter type. PNG 1.0-1.2 only define the single type. */
+#define PNG_FILTER_TYPE_BASE      0 /* Single row per-byte filtering */
+#define PNG_INTRAPIXEL_DIFFERENCING 64 /* Used only in MNG datastreams */
+#define PNG_FILTER_TYPE_DEFAULT   PNG_FILTER_TYPE_BASE
+
+/* These are for the interlacing type.  These values should NOT be changed. */
+#define PNG_INTERLACE_NONE        0 /* Non-interlaced image */
+#define PNG_INTERLACE_ADAM7       1 /* Adam7 interlacing */
+#define PNG_INTERLACE_LAST        2 /* Not a valid value */
+
+/* These are for the oFFs chunk.  These values should NOT be changed. */
+#define PNG_OFFSET_PIXEL          0 /* Offset in pixels */
+#define PNG_OFFSET_MICROMETER     1 /* Offset in micrometers (1/10^6 meter) */
+#define PNG_OFFSET_LAST           2 /* Not a valid value */
+
+/* These are for the pCAL chunk.  These values should NOT be changed. */
+#define PNG_EQUATION_LINEAR       0 /* Linear transformation */
+#define PNG_EQUATION_BASE_E       1 /* Exponential base e transform */
+#define PNG_EQUATION_ARBITRARY    2 /* Arbitrary base exponential transform */
+#define PNG_EQUATION_HYPERBOLIC   3 /* Hyperbolic sine transformation */
+#define PNG_EQUATION_LAST         4 /* Not a valid value */
+
+/* These are for the sCAL chunk.  These values should NOT be changed. */
+#define PNG_SCALE_UNKNOWN         0 /* unknown unit (image scale) */
+#define PNG_SCALE_METER           1 /* meters per pixel */
+#define PNG_SCALE_RADIAN          2 /* radians per pixel */
+#define PNG_SCALE_LAST            3 /* Not a valid value */
+
+/* These are for the pHYs chunk.  These values should NOT be changed. */
+#define PNG_RESOLUTION_UNKNOWN    0 /* pixels/unknown unit (aspect ratio) */
+#define PNG_RESOLUTION_METER      1 /* pixels/meter */
+#define PNG_RESOLUTION_LAST       2 /* Not a valid value */
+
+/* These are for the sRGB chunk.  These values should NOT be changed. */
+#define PNG_sRGB_INTENT_PERCEPTUAL 0
+#define PNG_sRGB_INTENT_RELATIVE   1
+#define PNG_sRGB_INTENT_SATURATION 2
+#define PNG_sRGB_INTENT_ABSOLUTE   3
+#define PNG_sRGB_INTENT_LAST       4 /* Not a valid value */
+
+/* This is for text chunks */
+#define PNG_KEYWORD_MAX_LENGTH     79
+
+/* Maximum number of entries in PLTE/sPLT/tRNS arrays */
+#define PNG_MAX_PALETTE_LENGTH    256
+
+/* These determine if an ancillary chunk's data has been successfully read
+ * from the PNG header, or if the application has filled in the corresponding
+ * data in the info_struct to be written into the output file.  The values
+ * of the PNG_INFO_<chunk> defines should NOT be changed.
+ */
+#define PNG_INFO_gAMA 0x0001
+#define PNG_INFO_sBIT 0x0002
+#define PNG_INFO_cHRM 0x0004
+#define PNG_INFO_PLTE 0x0008
+#define PNG_INFO_tRNS 0x0010
+#define PNG_INFO_bKGD 0x0020
+#define PNG_INFO_hIST 0x0040
+#define PNG_INFO_pHYs 0x0080
+#define PNG_INFO_oFFs 0x0100
+#define PNG_INFO_tIME 0x0200
+#define PNG_INFO_pCAL 0x0400
+#define PNG_INFO_sRGB 0x0800   /* GR-P, 0.96a */
+#define PNG_INFO_iCCP 0x1000   /* ESR, 1.0.6 */
+#define PNG_INFO_sPLT 0x2000   /* ESR, 1.0.6 */
+#define PNG_INFO_sCAL 0x4000   /* ESR, 1.0.6 */
+#define PNG_INFO_IDAT 0x8000   /* ESR, 1.0.6 */
+
+/* This is used for the transformation routines, as some of them
+ * change these values for the row.  It also should enable using
+ * the routines for other purposes.
+ */
+typedef struct png_row_info_struct
+{
+   png_uint_32 width;    /* width of row */
+   png_size_t rowbytes;  /* number of bytes in row */
+   png_byte color_type;  /* color type of row */
+   png_byte bit_depth;   /* bit depth of row */
+   png_byte channels;    /* number of channels (1, 2, 3, or 4) */
+   png_byte pixel_depth; /* bits per pixel (depth * channels) */
+} png_row_info;
+
+typedef png_row_info * png_row_infop;
+typedef png_row_info * * png_row_infopp;
+
+/* These are the function types for the I/O functions and for the functions
+ * that allow the user to override the default I/O functions with his or her
+ * own.  The png_error_ptr type should match that of user-supplied warning
+ * and error functions, while the png_rw_ptr type should match that of the
+ * user read/write data functions.  Note that the 'write' function must not
+ * modify the buffer it is passed. The 'read' function, on the other hand, is
+ * expected to return the read data in the buffer.
+ */
+typedef PNG_CALLBACK(void, *png_error_ptr, (png_structp, png_const_charp));
+typedef PNG_CALLBACK(void, *png_rw_ptr, (png_structp, png_bytep, png_size_t));
+typedef PNG_CALLBACK(void, *png_flush_ptr, (png_structp));
+typedef PNG_CALLBACK(void, *png_read_status_ptr, (png_structp, png_uint_32,
+    int));
+typedef PNG_CALLBACK(void, *png_write_status_ptr, (png_structp, png_uint_32,
+    int));
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+typedef PNG_CALLBACK(void, *png_progressive_info_ptr, (png_structp, png_infop));
+typedef PNG_CALLBACK(void, *png_progressive_end_ptr, (png_structp, png_infop));
+
+/* The following callback receives png_uint_32 row_number, int pass for the
+ * png_bytep data of the row.  When transforming an interlaced image the
+ * row number is the row number within the sub-image of the interlace pass, so
+ * the value will increase to the height of the sub-image (not the full image)
+ * then reset to 0 for the next pass.
+ *
+ * Use PNG_ROW_FROM_PASS_ROW(row, pass) and PNG_COL_FROM_PASS_COL(col, pass) to
+ * find the output pixel (x,y) given an interlaced sub-image pixel
+ * (row,col,pass).  (See below for these macros.)
+ */
+typedef PNG_CALLBACK(void, *png_progressive_row_ptr, (png_structp, png_bytep,
+    png_uint_32, int));
+#endif
+
+#if defined(PNG_READ_USER_TRANSFORM_SUPPORTED) || \
+    defined(PNG_WRITE_USER_TRANSFORM_SUPPORTED)
+typedef PNG_CALLBACK(void, *png_user_transform_ptr, (png_structp, png_row_infop,
+    png_bytep));
+#endif
+
+#ifdef PNG_USER_CHUNKS_SUPPORTED
+typedef PNG_CALLBACK(int, *png_user_chunk_ptr, (png_structp,
+    png_unknown_chunkp));
+#endif
+#ifdef PNG_UNKNOWN_CHUNKS_SUPPORTED
+/* not used anywhere */
+/* typedef PNG_CALLBACK(void, *png_unknown_chunk_ptr, (png_structp)); */
+#endif
+
+#ifdef PNG_SETJMP_SUPPORTED
+/* This must match the function definition in <setjmp.h>, and the application
+ * must include this before png.h to obtain the definition of jmp_buf.  The
+ * function is required to be PNG_NORETURN, but this is not checked.  If the
+ * function does return the application will crash via an abort() or similar
+ * system level call.
+ *
+ * If you get a warning here while building the library you may need to make
+ * changes to ensure that pnglibconf.h records the calling convention used by
+ * your compiler.  This may be very difficult - try using a different compiler
+ * to build the library!
+ */
+PNG_FUNCTION(void, (PNGCAPI *png_longjmp_ptr), PNGARG((jmp_buf, int)), typedef);
+#endif
+
+/* Transform masks for the high-level interface */
+#define PNG_TRANSFORM_IDENTITY       0x0000    /* read and write */
+#define PNG_TRANSFORM_STRIP_16       0x0001    /* read only */
+#define PNG_TRANSFORM_STRIP_ALPHA    0x0002    /* read only */
+#define PNG_TRANSFORM_PACKING        0x0004    /* read and write */
+#define PNG_TRANSFORM_PACKSWAP       0x0008    /* read and write */
+#define PNG_TRANSFORM_EXPAND         0x0010    /* read only */
+#define PNG_TRANSFORM_INVERT_MONO    0x0020    /* read and write */
+#define PNG_TRANSFORM_SHIFT          0x0040    /* read and write */
+#define PNG_TRANSFORM_BGR            0x0080    /* read and write */
+#define PNG_TRANSFORM_SWAP_ALPHA     0x0100    /* read and write */
+#define PNG_TRANSFORM_SWAP_ENDIAN    0x0200    /* read and write */
+#define PNG_TRANSFORM_INVERT_ALPHA   0x0400    /* read and write */
+#define PNG_TRANSFORM_STRIP_FILLER   0x0800    /* write only */
+/* Added to libpng-1.2.34 */
+#define PNG_TRANSFORM_STRIP_FILLER_BEFORE PNG_TRANSFORM_STRIP_FILLER
+#define PNG_TRANSFORM_STRIP_FILLER_AFTER 0x1000 /* write only */
+/* Added to libpng-1.4.0 */
+#define PNG_TRANSFORM_GRAY_TO_RGB   0x2000      /* read only */
+/* Added to libpng-1.5.4 */
+#define PNG_TRANSFORM_EXPAND_16     0x4000      /* read only */
+#define PNG_TRANSFORM_SCALE_16      0x8000      /* read only */
+
+/* Flags for MNG supported features */
+#define PNG_FLAG_MNG_EMPTY_PLTE     0x01
+#define PNG_FLAG_MNG_FILTER_64      0x04
+#define PNG_ALL_MNG_FEATURES        0x05
+
+/* NOTE: prior to 1.5 these functions had no 'API' style declaration,
+ * this allowed the zlib default functions to be used on Windows
+ * platforms.  In 1.5 the zlib default malloc (which just calls malloc and
+ * ignores the first argument) should be completely compatible with the
+ * following.
+ */
+typedef PNG_CALLBACK(png_voidp, *png_malloc_ptr, (png_structp,
+    png_alloc_size_t));
+typedef PNG_CALLBACK(void, *png_free_ptr, (png_structp, png_voidp));
+
+/* Section 3: exported functions
+ * Here are the function definitions most commonly used.  This is not
+ * the place to find out how to use libpng.  See libpng-manual.txt for the
+ * full explanation, see example.c for the summary.  This just provides
+ * a simple one line description of the use of each function.
+ *
+ * The PNG_EXPORT() and PNG_EXPORTA() macros used below are defined in
+ * pngconf.h and in the *.dfn files in the scripts directory.
+ *
+ *   PNG_EXPORT(ordinal, type, name, (args));
+ *
+ *       ordinal:    ordinal that is used while building
+ *                   *.def files. The ordinal value is only
+ *                   relevant when preprocessing png.h with
+ *                   the *.dfn files for building symbol table
+ *                   entries, and are removed by pngconf.h.
+ *       type:       return type of the function
+ *       name:       function name
+ *       args:       function arguments, with types
+ *
+ * When we wish to append attributes to a function prototype we use
+ * the PNG_EXPORTA() macro instead.
+ *
+ *   PNG_EXPORTA(ordinal, type, name, (args), attributes);
+ *
+ *       ordinal, type, name, and args: same as in PNG_EXPORT().
+ *       attributes: function attributes
+ */
+
+/* Returns the version number of the library */
+PNG_EXPORT(1, png_uint_32, png_access_version_number, (void));
+
+/* Tell lib we have already handled the first <num_bytes> magic bytes.
+ * Handling more than 8 bytes from the beginning of the file is an error.
+ */
+PNG_EXPORT(2, void, png_set_sig_bytes, (png_structrp png_ptr, int num_bytes));
+
+/* Check sig[start] through sig[start + num_to_check - 1] to see if it's a
+ * PNG file.  Returns zero if the supplied bytes match the 8-byte PNG
+ * signature, and non-zero otherwise.  Having num_to_check == 0 or
+ * start > 7 will always fail (ie return non-zero).
+ */
+PNG_EXPORT(3, int, png_sig_cmp, (png_const_bytep sig, png_size_t start,
+    png_size_t num_to_check));
+
+/* Simple signature checking function.  This is the same as calling
+ * png_check_sig(sig, n) := !png_sig_cmp(sig, 0, n).
+ */
+#define png_check_sig(sig, n) !png_sig_cmp((sig), 0, (n))
+
+/* Allocate and initialize png_ptr struct for reading, and any other memory. */
+PNG_EXPORTA(4, png_structp, png_create_read_struct,
+    (png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn),
+    PNG_ALLOCATED);
+
+/* Allocate and initialize png_ptr struct for writing, and any other memory */
+PNG_EXPORTA(5, png_structp, png_create_write_struct,
+    (png_const_charp user_png_ver, png_voidp error_ptr, png_error_ptr error_fn,
+    png_error_ptr warn_fn),
+    PNG_ALLOCATED);
+
+PNG_EXPORT(6, png_size_t, png_get_compression_buffer_size,
+    (png_const_structrp png_ptr));
+
+PNG_EXPORT(7, void, png_set_compression_buffer_size, (png_structrp png_ptr,
+    png_size_t size));
+
+/* Moved from pngconf.h in 1.4.0 and modified to ensure setjmp/longjmp
+ * match up.
+ */
+#ifdef PNG_SETJMP_SUPPORTED
+/* This function returns the jmp_buf built in to *png_ptr.  It must be
+ * supplied with an appropriate 'longjmp' function to use on that jmp_buf
+ * unless the default error function is overridden in which case NULL is
+ * acceptable.  The size of the jmp_buf is checked against the actual size
+ * allocated by the library - the call will return NULL on a mismatch
+ * indicating an ABI mismatch.
+ */
+PNG_EXPORT(8, jmp_buf*, png_set_longjmp_fn, (png_structrp png_ptr,
+    png_longjmp_ptr longjmp_fn, size_t jmp_buf_size));
+#  define png_jmpbuf(png_ptr) \
+      (*png_set_longjmp_fn((png_ptr), longjmp, (sizeof (jmp_buf))))
+#else
+#  define png_jmpbuf(png_ptr) \
+      (LIBPNG_WAS_COMPILED_WITH__PNG_NO_SETJMP)
+#endif
+/* This function should be used by libpng applications in place of
+ * longjmp(png_ptr->jmpbuf, val).  If longjmp_fn() has been set, it
+ * will use it; otherwise it will call PNG_ABORT().  This function was
+ * added in libpng-1.5.0.
+ */
+PNG_EXPORTA(9, void, png_longjmp, (png_const_structrp png_ptr, int val),
+    PNG_NORETURN);
+
+#ifdef PNG_READ_SUPPORTED
+/* Reset the compression stream */
+PNG_EXPORTA(10, int, png_reset_zstream, (png_structrp png_ptr), PNG_DEPRECATED);
+#endif
+
+/* New functions added in libpng-1.0.2 (not enabled by default until 1.2.0) */
+#ifdef PNG_USER_MEM_SUPPORTED
+PNG_EXPORTA(11, png_structp, png_create_read_struct_2,
+    (png_const_charp user_png_ver, png_voidp error_ptr, png_error_ptr error_fn,
+    png_error_ptr warn_fn,
+    png_voidp mem_ptr, png_malloc_ptr malloc_fn, png_free_ptr free_fn),
+    PNG_ALLOCATED);
+PNG_EXPORTA(12, png_structp, png_create_write_struct_2,
+    (png_const_charp user_png_ver, png_voidp error_ptr, png_error_ptr error_fn,
+    png_error_ptr warn_fn,
+    png_voidp mem_ptr, png_malloc_ptr malloc_fn, png_free_ptr free_fn),
+    PNG_ALLOCATED);
+#endif
+
+/* Write the PNG file signature. */
+PNG_EXPORT(13, void, png_write_sig, (png_structrp png_ptr));
+
+/* Write a PNG chunk - size, type, (optional) data, CRC. */
+PNG_EXPORT(14, void, png_write_chunk, (png_structrp png_ptr, png_const_bytep
+    chunk_name, png_const_bytep data, png_size_t length));
+
+/* Write the start of a PNG chunk - length and chunk name. */
+PNG_EXPORT(15, void, png_write_chunk_start, (png_structrp png_ptr,
+    png_const_bytep chunk_name, png_uint_32 length));
+
+/* Write the data of a PNG chunk started with png_write_chunk_start(). */
+PNG_EXPORT(16, void, png_write_chunk_data, (png_structrp png_ptr,
+    png_const_bytep data, png_size_t length));
+
+/* Finish a chunk started with png_write_chunk_start() (includes CRC). */
+PNG_EXPORT(17, void, png_write_chunk_end, (png_structrp png_ptr));
+
+/* Allocate and initialize the info structure */
+PNG_EXPORTA(18, png_infop, png_create_info_struct, (png_const_structrp png_ptr),
+    PNG_ALLOCATED);
+
+/* DEPRECATED: this function allowed init structures to be created using the
+ * default allocation method (typically malloc).  Use is deprecated in 1.6.0 and
+ * the API will be removed in the future.
+ */
+PNG_EXPORTA(19, void, png_info_init_3, (png_infopp info_ptr,
+    png_size_t png_info_struct_size), PNG_DEPRECATED);
+
+/* Writes all the PNG information before the image. */
+PNG_EXPORT(20, void, png_write_info_before_PLTE,
+    (png_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(21, void, png_write_info,
+    (png_structrp png_ptr, png_const_inforp info_ptr));
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the information before the actual image data. */
+PNG_EXPORT(22, void, png_read_info,
+    (png_structrp png_ptr, png_inforp info_ptr));
+#endif
+
+#ifdef PNG_TIME_RFC1123_SUPPORTED
+   /* Convert to a US string format: there is no localization support in this
+    * routine.  The original implementation used a 29 character buffer in
+    * png_struct, this will be removed in future versions.
+    */
+#if PNG_LIBPNG_VER < 10700
+/* To do: remove this from libpng17 (and from libpng17/png.c and pngstruct.h) */
+PNG_EXPORTA(23, png_const_charp, png_convert_to_rfc1123, (png_structrp png_ptr,
+    png_const_timep ptime),PNG_DEPRECATED);
+#endif
+PNG_EXPORT(241, int, png_convert_to_rfc1123_buffer, (char out[29],
+    png_const_timep ptime));
+#endif
+
+#ifdef PNG_CONVERT_tIME_SUPPORTED
+/* Convert from a struct tm to png_time */
+PNG_EXPORT(24, void, png_convert_from_struct_tm, (png_timep ptime,
+    const struct tm * ttime));
+
+/* Convert from time_t to png_time.  Uses gmtime() */
+PNG_EXPORT(25, void, png_convert_from_time_t, (png_timep ptime, time_t ttime));
+#endif /* PNG_CONVERT_tIME_SUPPORTED */
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+/* Expand data to 24-bit RGB, or 8-bit grayscale, with alpha if available. */
+PNG_EXPORT(26, void, png_set_expand, (png_structrp png_ptr));
+PNG_EXPORT(27, void, png_set_expand_gray_1_2_4_to_8, (png_structrp png_ptr));
+PNG_EXPORT(28, void, png_set_palette_to_rgb, (png_structrp png_ptr));
+PNG_EXPORT(29, void, png_set_tRNS_to_alpha, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+/* Expand to 16-bit channels, forces conversion of palette to RGB and expansion
+ * of a tRNS chunk if present.
+ */
+PNG_EXPORT(221, void, png_set_expand_16, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_BGR_SUPPORTED) || defined(PNG_WRITE_BGR_SUPPORTED)
+/* Use blue, green, red order for pixels. */
+PNG_EXPORT(30, void, png_set_bgr, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+/* Expand the grayscale to 24-bit RGB if necessary. */
+PNG_EXPORT(31, void, png_set_gray_to_rgb, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+/* Reduce RGB to grayscale. */
+#define PNG_ERROR_ACTION_NONE  1
+#define PNG_ERROR_ACTION_WARN  2
+#define PNG_ERROR_ACTION_ERROR 3
+#define PNG_RGB_TO_GRAY_DEFAULT (-1)/*for red/green coefficients*/
+
+PNG_FP_EXPORT(32, void, png_set_rgb_to_gray, (png_structrp png_ptr,
+    int error_action, double red, double green))
+PNG_FIXED_EXPORT(33, void, png_set_rgb_to_gray_fixed, (png_structrp png_ptr,
+    int error_action, png_fixed_point red, png_fixed_point green))
+
+PNG_EXPORT(34, png_byte, png_get_rgb_to_gray_status, (png_const_structrp
+    png_ptr));
+#endif
+
+#ifdef PNG_BUILD_GRAYSCALE_PALETTE_SUPPORTED
+PNG_EXPORT(35, void, png_build_grayscale_palette, (int bit_depth,
+    png_colorp palette));
+#endif
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+/* How the alpha channel is interpreted - this affects how the color channels of
+ * a PNG file are returned when an alpha channel, or tRNS chunk in a palette
+ * file, is present.
+ *
+ * This has no effect on the way pixels are written into a PNG output
+ * datastream. The color samples in a PNG datastream are never premultiplied
+ * with the alpha samples.
+ *
+ * The default is to return data according to the PNG specification: the alpha
+ * channel is a linear measure of the contribution of the pixel to the
+ * corresponding composited pixel.  The gamma encoded color channels must be
+ * scaled according to the contribution and to do this it is necessary to undo
+ * the encoding, scale the color values, perform the composition and reencode
+ * the values.  This is the 'PNG' mode.
+ *
+ * The alternative is to 'associate' the alpha with the color information by
+ * storing color channel values that have been scaled by the alpha.  The
+ * advantage is that the color channels can be resampled (the image can be
+ * scaled) in this form.  The disadvantage is that normal practice is to store
+ * linear, not (gamma) encoded, values and this requires 16-bit channels for
+ * still images rather than the 8-bit channels that are just about sufficient if
+ * gamma encoding is used.  In addition all non-transparent pixel values,
+ * including completely opaque ones, must be gamma encoded to produce the final
+ * image.  This is the 'STANDARD', 'ASSOCIATED' or 'PREMULTIPLIED' mode (the
+ * latter being the two common names for associated alpha color channels.)
+ *
+ * Since it is not necessary to perform arithmetic on opaque color values so
+ * long as they are not to be resampled and are in the final color space it is
+ * possible to optimize the handling of alpha by storing the opaque pixels in
+ * the PNG format (adjusted for the output color space) while storing partially
+ * opaque pixels in the standard, linear, format.  The accuracy required for
+ * standard alpha composition is relatively low, because the pixels are
+ * isolated, therefore typically the accuracy loss in storing 8-bit linear
+ * values is acceptable.  (This is not true if the alpha channel is used to
+ * simulate transparency over large areas - use 16 bits or the PNG mode in
+ * this case!)  This is the 'OPTIMIZED' mode.  For this mode a pixel is
+ * treated as opaque only if the alpha value is equal to the maximum value.
+ *
+ * The final choice is to gamma encode the alpha channel as well.  This is
+ * broken because, in practice, no implementation that uses this choice
+ * correctly undoes the encoding before handling alpha composition.  Use this
+ * choice only if other serious errors in the software or hardware you use
+ * mandate it; the typical serious error is for dark halos to appear around
+ * opaque areas of the composited PNG image because of arithmetic overflow.
+ *
+ * The API function png_set_alpha_mode specifies which of these choices to use
+ * with an enumerated 'mode' value and the gamma of the required output:
+ */
+#define PNG_ALPHA_PNG           0 /* according to the PNG standard */
+#define PNG_ALPHA_STANDARD      1 /* according to Porter/Duff */
+#define PNG_ALPHA_ASSOCIATED    1 /* as above; this is the normal practice */
+#define PNG_ALPHA_PREMULTIPLIED 1 /* as above */
+#define PNG_ALPHA_OPTIMIZED     2 /* 'PNG' for opaque pixels, else 'STANDARD' */
+#define PNG_ALPHA_BROKEN        3 /* the alpha channel is gamma encoded */
+
+PNG_FP_EXPORT(227, void, png_set_alpha_mode, (png_structrp png_ptr, int mode,
+    double output_gamma))
+PNG_FIXED_EXPORT(228, void, png_set_alpha_mode_fixed, (png_structrp png_ptr,
+    int mode, png_fixed_point output_gamma))
+#endif
+
+#if defined(PNG_GAMMA_SUPPORTED) || defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+/* The output_gamma value is a screen gamma in libpng terminology: it expresses
+ * how to decode the output values, not how they are encoded.  The values used
+ * correspond to the normal numbers used to describe the overall gamma of a
+ * computer display system; for example 2.2 for an sRGB conformant system.  The
+ * values are scaled by 100000 in the _fixed version of the API (so 220000 for
+ * sRGB.)
+ *
+ * The inverse of the value is always used to provide a default for the PNG file
+ * encoding if it has no gAMA chunk and if png_set_gamma() has not been called
+ * to override the PNG gamma information.
+ *
+ * When the ALPHA_OPTIMIZED mode is selected the output gamma is used to encode
+ * opaque pixels however pixels with lower alpha values are not encoded,
+ * regardless of the output gamma setting.
+ *
+ * When the standard Porter Duff handling is requested with mode 1 the output
+ * encoding is set to be linear and the output_gamma value is only relevant
+ * as a default for input data that has no gamma information.  The linear output
+ * encoding will be overridden if png_set_gamma() is called - the results may be
+ * highly unexpected!
+ *
+ * The following numbers are derived from the sRGB standard and the research
+ * behind it.  sRGB is defined to be approximated by a PNG gAMA chunk value of
+ * 0.45455 (1/2.2) for PNG.  The value implicitly includes any viewing
+ * correction required to take account of any differences in the color
+ * environment of the original scene and the intended display environment; the
+ * value expresses how to *decode* the image for display, not how the original
+ * data was *encoded*.
+ *
+ * sRGB provides a peg for the PNG standard by defining a viewing environment.
+ * sRGB itself, and earlier TV standards, actually use a more complex transform
+ * (a linear portion then a gamma 2.4 power law) than PNG can express.  (PNG is
+ * limited to simple power laws.)  By saying that an image for direct display on
+ * an sRGB conformant system should be stored with a gAMA chunk value of 45455
+ * (11.3.3.2 and 11.3.3.5 of the ISO PNG specification) the PNG specification
+ * makes it possible to derive values for other display systems and
+ * environments.
+ *
+ * The Mac value is deduced from the sRGB based on an assumption that the actual
+ * extra viewing correction used in early Mac display systems was implemented as
+ * a power 1.45 lookup table.
+ *
+ * Any system where a programmable lookup table is used or where the behavior of
+ * the final display device characteristics can be changed requires system
+ * specific code to obtain the current characteristic.  However this can be
+ * difficult and most PNG gamma correction only requires an approximate value.
+ *
+ * By default, if png_set_alpha_mode() is not called, libpng assumes that all
+ * values are unencoded, linear, values and that the output device also has a
+ * linear characteristic.  This is only very rarely correct - it is invariably
+ * better to call png_set_alpha_mode() with PNG_DEFAULT_sRGB than rely on the
+ * default if you don't know what the right answer is!
+ *
+ * The special value PNG_GAMMA_MAC_18 indicates an older Mac system (pre Mac OS
+ * 10.6) which used a correction table to implement a somewhat lower gamma on an
+ * otherwise sRGB system.
+ *
+ * Both these values are reserved (not simple gamma values) in order to allow
+ * more precise correction internally in the future.
+ *
+ * NOTE: the following values can be passed to either the fixed or floating
+ * point APIs, but the floating point API will also accept floating point
+ * values.
+ */
+#define PNG_DEFAULT_sRGB -1       /* sRGB gamma and color space */
+#define PNG_GAMMA_MAC_18 -2       /* Old Mac '1.8' gamma and color space */
+#define PNG_GAMMA_sRGB   220000   /* Television standards--matches sRGB gamma */
+#define PNG_GAMMA_LINEAR PNG_FP_1 /* Linear */
+#endif
+
+/* The following are examples of calls to png_set_alpha_mode to achieve the
+ * required overall gamma correction and, where necessary, alpha
+ * premultiplication.
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_PNG, PNG_DEFAULT_sRGB);
+ *    This is the default libpng handling of the alpha channel - it is not
+ *    pre-multiplied into the color components.  In addition the call states
+ *    that the output is for a sRGB system and causes all PNG files without gAMA
+ *    chunks to be assumed to be encoded using sRGB.
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_PNG, PNG_GAMMA_MAC);
+ *    In this case the output is assumed to be something like an sRGB conformant
+ *    display preceeded by a power-law lookup table of power 1.45.  This is how
+ *    early Mac systems behaved.
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_STANDARD, PNG_GAMMA_LINEAR);
+ *    This is the classic Jim Blinn approach and will work in academic
+ *    environments where everything is done by the book.  It has the shortcoming
+ *    of assuming that input PNG data with no gamma information is linear - this
+ *    is unlikely to be correct unless the PNG files where generated locally.
+ *    Most of the time the output precision will be so low as to show
+ *    significant banding in dark areas of the image.
+ *
+ * png_set_expand_16(pp);
+ * png_set_alpha_mode(pp, PNG_ALPHA_STANDARD, PNG_DEFAULT_sRGB);
+ *    This is a somewhat more realistic Jim Blinn inspired approach.  PNG files
+ *    are assumed to have the sRGB encoding if not marked with a gamma value and
+ *    the output is always 16 bits per component.  This permits accurate scaling
+ *    and processing of the data.  If you know that your input PNG files were
+ *    generated locally you might need to replace PNG_DEFAULT_sRGB with the
+ *    correct value for your system.
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_OPTIMIZED, PNG_DEFAULT_sRGB);
+ *    If you just need to composite the PNG image onto an existing background
+ *    and if you control the code that does this you can use the optimization
+ *    setting.  In this case you just copy completely opaque pixels to the
+ *    output.  For pixels that are not completely transparent (you just skip
+ *    those) you do the composition math using png_composite or png_composite_16
+ *    below then encode the resultant 8-bit or 16-bit values to match the output
+ *    encoding.
+ *
+ * Other cases
+ *    If neither the PNG nor the standard linear encoding work for you because
+ *    of the software or hardware you use then you have a big problem.  The PNG
+ *    case will probably result in halos around the image.  The linear encoding
+ *    will probably result in a washed out, too bright, image (it's actually too
+ *    contrasty.)  Try the ALPHA_OPTIMIZED mode above - this will probably
+ *    substantially reduce the halos.  Alternatively try:
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_BROKEN, PNG_DEFAULT_sRGB);
+ *    This option will also reduce the halos, but there will be slight dark
+ *    halos round the opaque parts of the image where the background is light.
+ *    In the OPTIMIZED mode the halos will be light halos where the background
+ *    is dark.  Take your pick - the halos are unavoidable unless you can get
+ *    your hardware/software fixed!  (The OPTIMIZED approach is slightly
+ *    faster.)
+ *
+ * When the default gamma of PNG files doesn't match the output gamma.
+ *    If you have PNG files with no gamma information png_set_alpha_mode allows
+ *    you to provide a default gamma, but it also sets the ouput gamma to the
+ *    matching value.  If you know your PNG files have a gamma that doesn't
+ *    match the output you can take advantage of the fact that
+ *    png_set_alpha_mode always sets the output gamma but only sets the PNG
+ *    default if it is not already set:
+ *
+ * png_set_alpha_mode(pp, PNG_ALPHA_PNG, PNG_DEFAULT_sRGB);
+ * png_set_alpha_mode(pp, PNG_ALPHA_PNG, PNG_GAMMA_MAC);
+ *    The first call sets both the default and the output gamma values, the
+ *    second call overrides the output gamma without changing the default.  This
+ *    is easier than achieving the same effect with png_set_gamma.  You must use
+ *    PNG_ALPHA_PNG for the first call - internal checking in png_set_alpha will
+ *    fire if more than one call to png_set_alpha_mode and png_set_background is
+ *    made in the same read operation, however multiple calls with PNG_ALPHA_PNG
+ *    are ignored.
+ */
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+PNG_EXPORT(36, void, png_set_strip_alpha, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_SWAP_ALPHA_SUPPORTED) || \
+    defined(PNG_WRITE_SWAP_ALPHA_SUPPORTED)
+PNG_EXPORT(37, void, png_set_swap_alpha, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_INVERT_ALPHA_SUPPORTED) || \
+    defined(PNG_WRITE_INVERT_ALPHA_SUPPORTED)
+PNG_EXPORT(38, void, png_set_invert_alpha, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_FILLER_SUPPORTED) || defined(PNG_WRITE_FILLER_SUPPORTED)
+/* Add a filler byte to 8-bit Gray or 24-bit RGB images. */
+PNG_EXPORT(39, void, png_set_filler, (png_structrp png_ptr, png_uint_32 filler,
+    int flags));
+/* The values of the PNG_FILLER_ defines should NOT be changed */
+#  define PNG_FILLER_BEFORE 0
+#  define PNG_FILLER_AFTER 1
+/* Add an alpha byte to 8-bit Gray or 24-bit RGB images. */
+PNG_EXPORT(40, void, png_set_add_alpha, (png_structrp png_ptr,
+    png_uint_32 filler, int flags));
+#endif /* PNG_READ_FILLER_SUPPORTED || PNG_WRITE_FILLER_SUPPORTED */
+
+#if defined(PNG_READ_SWAP_SUPPORTED) || defined(PNG_WRITE_SWAP_SUPPORTED)
+/* Swap bytes in 16-bit depth files. */
+PNG_EXPORT(41, void, png_set_swap, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_PACK_SUPPORTED) || defined(PNG_WRITE_PACK_SUPPORTED)
+/* Use 1 byte per pixel in 1, 2, or 4-bit depth files. */
+PNG_EXPORT(42, void, png_set_packing, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_PACKSWAP_SUPPORTED) || \
+    defined(PNG_WRITE_PACKSWAP_SUPPORTED)
+/* Swap packing order of pixels in bytes. */
+PNG_EXPORT(43, void, png_set_packswap, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_SHIFT_SUPPORTED) || defined(PNG_WRITE_SHIFT_SUPPORTED)
+/* Converts files to legal bit depths. */
+PNG_EXPORT(44, void, png_set_shift, (png_structrp png_ptr, png_const_color_8p
+    true_bits));
+#endif
+
+#if defined(PNG_READ_INTERLACING_SUPPORTED) || \
+    defined(PNG_WRITE_INTERLACING_SUPPORTED)
+/* Have the code handle the interlacing.  Returns the number of passes.
+ * MUST be called before png_read_update_info or png_start_read_image,
+ * otherwise it will not have the desired effect.  Note that it is still
+ * necessary to call png_read_row or png_read_rows png_get_image_height
+ * times for each pass.
+*/
+PNG_EXPORT(45, int, png_set_interlace_handling, (png_structrp png_ptr));
+#endif
+
+#if defined(PNG_READ_INVERT_SUPPORTED) || defined(PNG_WRITE_INVERT_SUPPORTED)
+/* Invert monochrome files */
+PNG_EXPORT(46, void, png_set_invert_mono, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+/* Handle alpha and tRNS by replacing with a background color.  Prior to
+ * libpng-1.5.4 this API must not be called before the PNG file header has been
+ * read.  Doing so will result in unexpected behavior and possible warnings or
+ * errors if the PNG file contains a bKGD chunk.
+ */
+PNG_FP_EXPORT(47, void, png_set_background, (png_structrp png_ptr,
+    png_const_color_16p background_color, int background_gamma_code,
+    int need_expand, double background_gamma))
+PNG_FIXED_EXPORT(215, void, png_set_background_fixed, (png_structrp png_ptr,
+    png_const_color_16p background_color, int background_gamma_code,
+    int need_expand, png_fixed_point background_gamma))
+#endif
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+#  define PNG_BACKGROUND_GAMMA_UNKNOWN 0
+#  define PNG_BACKGROUND_GAMMA_SCREEN  1
+#  define PNG_BACKGROUND_GAMMA_FILE    2
+#  define PNG_BACKGROUND_GAMMA_UNIQUE  3
+#endif
+
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+/* Scale a 16-bit depth file down to 8-bit, accurately. */
+PNG_EXPORT(229, void, png_set_scale_16, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+#define PNG_READ_16_TO_8 SUPPORTED /* Name prior to 1.5.4 */
+/* Strip the second byte of information from a 16-bit depth file. */
+PNG_EXPORT(48, void, png_set_strip_16, (png_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+/* Turn on quantizing, and reduce the palette to the number of colors
+ * available.
+ */
+PNG_EXPORT(49, void, png_set_quantize, (png_structrp png_ptr,
+    png_colorp palette, int num_palette, int maximum_colors,
+    png_const_uint_16p histogram, int full_quantize));
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* The threshold on gamma processing is configurable but hard-wired into the
+ * library.  The following is the floating point variant.
+ */
+#define PNG_GAMMA_THRESHOLD (PNG_GAMMA_THRESHOLD_FIXED*.00001)
+
+/* Handle gamma correction. Screen_gamma=(display_exponent).
+ * NOTE: this API simply sets the screen and file gamma values. It will
+ * therefore override the value for gamma in a PNG file if it is called after
+ * the file header has been read - use with care  - call before reading the PNG
+ * file for best results!
+ *
+ * These routines accept the same gamma values as png_set_alpha_mode (described
+ * above).  The PNG_GAMMA_ defines and PNG_DEFAULT_sRGB can be passed to either
+ * API (floating point or fixed.)  Notice, however, that the 'file_gamma' value
+ * is the inverse of a 'screen gamma' value.
+ */
+PNG_FP_EXPORT(50, void, png_set_gamma, (png_structrp png_ptr,
+    double screen_gamma, double override_file_gamma))
+PNG_FIXED_EXPORT(208, void, png_set_gamma_fixed, (png_structrp png_ptr,
+    png_fixed_point screen_gamma, png_fixed_point override_file_gamma))
+#endif
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+/* Set how many lines between output flushes - 0 for no flushing */
+PNG_EXPORT(51, void, png_set_flush, (png_structrp png_ptr, int nrows));
+/* Flush the current PNG output buffer */
+PNG_EXPORT(52, void, png_write_flush, (png_structrp png_ptr));
+#endif
+
+/* Optional update palette with requested transformations */
+PNG_EXPORT(53, void, png_start_read_image, (png_structrp png_ptr));
+
+/* Optional call to update the users info structure */
+PNG_EXPORT(54, void, png_read_update_info, (png_structrp png_ptr,
+    png_inforp info_ptr));
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read one or more rows of image data. */
+PNG_EXPORT(55, void, png_read_rows, (png_structrp png_ptr, png_bytepp row,
+    png_bytepp display_row, png_uint_32 num_rows));
+#endif
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read a row of data. */
+PNG_EXPORT(56, void, png_read_row, (png_structrp png_ptr, png_bytep row,
+    png_bytep display_row));
+#endif
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the whole image into memory at once. */
+PNG_EXPORT(57, void, png_read_image, (png_structrp png_ptr, png_bytepp image));
+#endif
+
+/* Write a row of image data */
+PNG_EXPORT(58, void, png_write_row, (png_structrp png_ptr,
+    png_const_bytep row));
+
+/* Write a few rows of image data: (*row) is not written; however, the type
+ * is declared as writeable to maintain compatibility with previous versions
+ * of libpng and to allow the 'display_row' array from read_rows to be passed
+ * unchanged to write_rows.
+ */
+PNG_EXPORT(59, void, png_write_rows, (png_structrp png_ptr, png_bytepp row,
+    png_uint_32 num_rows));
+
+/* Write the image data */
+PNG_EXPORT(60, void, png_write_image, (png_structrp png_ptr, png_bytepp image));
+
+/* Write the end of the PNG file. */
+PNG_EXPORT(61, void, png_write_end, (png_structrp png_ptr,
+    png_inforp info_ptr));
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the end of the PNG file. */
+PNG_EXPORT(62, void, png_read_end, (png_structrp png_ptr, png_inforp info_ptr));
+#endif
+
+/* Free any memory associated with the png_info_struct */
+PNG_EXPORT(63, void, png_destroy_info_struct, (png_const_structrp png_ptr,
+    png_infopp info_ptr_ptr));
+
+/* Free any memory associated with the png_struct and the png_info_structs */
+PNG_EXPORT(64, void, png_destroy_read_struct, (png_structpp png_ptr_ptr,
+    png_infopp info_ptr_ptr, png_infopp end_info_ptr_ptr));
+
+/* Free any memory associated with the png_struct and the png_info_structs */
+PNG_EXPORT(65, void, png_destroy_write_struct, (png_structpp png_ptr_ptr,
+    png_infopp info_ptr_ptr));
+
+/* Set the libpng method of handling chunk CRC errors */
+PNG_EXPORT(66, void, png_set_crc_action, (png_structrp png_ptr, int crit_action,
+    int ancil_action));
+
+/* Values for png_set_crc_action() say how to handle CRC errors in
+ * ancillary and critical chunks, and whether to use the data contained
+ * therein.  Note that it is impossible to "discard" data in a critical
+ * chunk.  For versions prior to 0.90, the action was always error/quit,
+ * whereas in version 0.90 and later, the action for CRC errors in ancillary
+ * chunks is warn/discard.  These values should NOT be changed.
+ *
+ *      value                       action:critical     action:ancillary
+ */
+#define PNG_CRC_DEFAULT       0  /* error/quit          warn/discard data */
+#define PNG_CRC_ERROR_QUIT    1  /* error/quit          error/quit        */
+#define PNG_CRC_WARN_DISCARD  2  /* (INVALID)           warn/discard data */
+#define PNG_CRC_WARN_USE      3  /* warn/use data       warn/use data     */
+#define PNG_CRC_QUIET_USE     4  /* quiet/use data      quiet/use data    */
+#define PNG_CRC_NO_CHANGE     5  /* use current value   use current value */
+
+/* These functions give the user control over the scan-line filtering in
+ * libpng and the compression methods used by zlib.  These functions are
+ * mainly useful for testing, as the defaults should work with most users.
+ * Those users who are tight on memory or want faster performance at the
+ * expense of compression can modify them.  See the compression library
+ * header file (zlib.h) for an explination of the compression functions.
+ */
+
+/* Set the filtering method(s) used by libpng.  Currently, the only valid
+ * value for "method" is 0.
+ */
+PNG_EXPORT(67, void, png_set_filter, (png_structrp png_ptr, int method,
+    int filters));
+
+/* Flags for png_set_filter() to say which filters to use.  The flags
+ * are chosen so that they don't conflict with real filter types
+ * below, in case they are supplied instead of the #defined constants.
+ * These values should NOT be changed.
+ */
+#define PNG_NO_FILTERS     0x00
+#define PNG_FILTER_NONE    0x08
+#define PNG_FILTER_SUB     0x10
+#define PNG_FILTER_UP      0x20
+#define PNG_FILTER_AVG     0x40
+#define PNG_FILTER_PAETH   0x80
+#define PNG_ALL_FILTERS (PNG_FILTER_NONE | PNG_FILTER_SUB | PNG_FILTER_UP | \
+                         PNG_FILTER_AVG | PNG_FILTER_PAETH)
+
+/* Filter values (not flags) - used in pngwrite.c, pngwutil.c for now.
+ * These defines should NOT be changed.
+ */
+#define PNG_FILTER_VALUE_NONE  0
+#define PNG_FILTER_VALUE_SUB   1
+#define PNG_FILTER_VALUE_UP    2
+#define PNG_FILTER_VALUE_AVG   3
+#define PNG_FILTER_VALUE_PAETH 4
+#define PNG_FILTER_VALUE_LAST  5
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED /* EXPERIMENTAL */
+/* The "heuristic_method" is given by one of the PNG_FILTER_HEURISTIC_
+ * defines, either the default (minimum-sum-of-absolute-differences), or
+ * the experimental method (weighted-minimum-sum-of-absolute-differences).
+ *
+ * Weights are factors >= 1.0, indicating how important it is to keep the
+ * filter type consistent between rows.  Larger numbers mean the current
+ * filter is that many times as likely to be the same as the "num_weights"
+ * previous filters.  This is cumulative for each previous row with a weight.
+ * There needs to be "num_weights" values in "filter_weights", or it can be
+ * NULL if the weights aren't being specified.  Weights have no influence on
+ * the selection of the first row filter.  Well chosen weights can (in theory)
+ * improve the compression for a given image.
+ *
+ * Costs are factors >= 1.0 indicating the relative decoding costs of a
+ * filter type.  Higher costs indicate more decoding expense, and are
+ * therefore less likely to be selected over a filter with lower computational
+ * costs.  There needs to be a value in "filter_costs" for each valid filter
+ * type (given by PNG_FILTER_VALUE_LAST), or it can be NULL if you aren't
+ * setting the costs.  Costs try to improve the speed of decompression without
+ * unduly increasing the compressed image size.
+ *
+ * A negative weight or cost indicates the default value is to be used, and
+ * values in the range [0.0, 1.0) indicate the value is to remain unchanged.
+ * The default values for both weights and costs are currently 1.0, but may
+ * change if good general weighting/cost heuristics can be found.  If both
+ * the weights and costs are set to 1.0, this degenerates the WEIGHTED method
+ * to the UNWEIGHTED method, but with added encoding time/computation.
+ */
+PNG_FP_EXPORT(68, void, png_set_filter_heuristics, (png_structrp png_ptr,
+    int heuristic_method, int num_weights, png_const_doublep filter_weights,
+    png_const_doublep filter_costs))
+PNG_FIXED_EXPORT(209, void, png_set_filter_heuristics_fixed,
+    (png_structrp png_ptr, int heuristic_method, int num_weights,
+    png_const_fixed_point_p filter_weights,
+    png_const_fixed_point_p filter_costs))
+#endif /*  PNG_WRITE_WEIGHTED_FILTER_SUPPORTED */
+
+/* Heuristic used for row filter selection.  These defines should NOT be
+ * changed.
+ */
+#define PNG_FILTER_HEURISTIC_DEFAULT    0  /* Currently "UNWEIGHTED" */
+#define PNG_FILTER_HEURISTIC_UNWEIGHTED 1  /* Used by libpng < 0.95 */
+#define PNG_FILTER_HEURISTIC_WEIGHTED   2  /* Experimental feature */
+#define PNG_FILTER_HEURISTIC_LAST       3  /* Not a valid value */
+
+#ifdef PNG_WRITE_SUPPORTED
+/* Set the library compression level.  Currently, valid values range from
+ * 0 - 9, corresponding directly to the zlib compression levels 0 - 9
+ * (0 - no compression, 9 - "maximal" compression).  Note that tests have
+ * shown that zlib compression levels 3-6 usually perform as well as level 9
+ * for PNG images, and do considerably fewer caclulations.  In the future,
+ * these values may not correspond directly to the zlib compression levels.
+ */
+PNG_EXPORT(69, void, png_set_compression_level, (png_structrp png_ptr,
+    int level));
+
+PNG_EXPORT(70, void, png_set_compression_mem_level, (png_structrp png_ptr,
+    int mem_level));
+
+PNG_EXPORT(71, void, png_set_compression_strategy, (png_structrp png_ptr,
+    int strategy));
+
+/* If PNG_WRITE_OPTIMIZE_CMF_SUPPORTED is defined, libpng will use a
+ * smaller value of window_bits if it can do so safely.
+ */
+PNG_EXPORT(72, void, png_set_compression_window_bits, (png_structrp png_ptr,
+    int window_bits));
+
+PNG_EXPORT(73, void, png_set_compression_method, (png_structrp png_ptr,
+    int method));
+#endif
+
+#ifdef PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED
+/* Also set zlib parameters for compressing non-IDAT chunks */
+PNG_EXPORT(222, void, png_set_text_compression_level, (png_structrp png_ptr,
+    int level));
+
+PNG_EXPORT(223, void, png_set_text_compression_mem_level, (png_structrp png_ptr,
+    int mem_level));
+
+PNG_EXPORT(224, void, png_set_text_compression_strategy, (png_structrp png_ptr,
+    int strategy));
+
+/* If PNG_WRITE_OPTIMIZE_CMF_SUPPORTED is defined, libpng will use a
+ * smaller value of window_bits if it can do so safely.
+ */
+PNG_EXPORT(225, void, png_set_text_compression_window_bits,
+    (png_structrp png_ptr, int window_bits));
+
+PNG_EXPORT(226, void, png_set_text_compression_method, (png_structrp png_ptr,
+    int method));
+#endif /* PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED */
+
+/* These next functions are called for input/output, memory, and error
+ * handling.  They are in the file pngrio.c, pngwio.c, and pngerror.c,
+ * and call standard C I/O routines such as fread(), fwrite(), and
+ * fprintf().  These functions can be made to use other I/O routines
+ * at run time for those applications that need to handle I/O in a
+ * different manner by calling png_set_???_fn().  See libpng-manual.txt for
+ * more information.
+ */
+
+#ifdef PNG_STDIO_SUPPORTED
+/* Initialize the input/output for the PNG file to the default functions. */
+PNG_EXPORT(74, void, png_init_io, (png_structrp png_ptr, png_FILE_p fp));
+#endif
+
+/* Replace the (error and abort), and warning functions with user
+ * supplied functions.  If no messages are to be printed you must still
+ * write and use replacement functions. The replacement error_fn should
+ * still do a longjmp to the last setjmp location if you are using this
+ * method of error handling.  If error_fn or warning_fn is NULL, the
+ * default function will be used.
+ */
+
+PNG_EXPORT(75, void, png_set_error_fn, (png_structrp png_ptr,
+    png_voidp error_ptr, png_error_ptr error_fn, png_error_ptr warning_fn));
+
+/* Return the user pointer associated with the error functions */
+PNG_EXPORT(76, png_voidp, png_get_error_ptr, (png_const_structrp png_ptr));
+
+/* Replace the default data output functions with a user supplied one(s).
+ * If buffered output is not used, then output_flush_fn can be set to NULL.
+ * If PNG_WRITE_FLUSH_SUPPORTED is not defined at libpng compile time
+ * output_flush_fn will be ignored (and thus can be NULL).
+ * It is probably a mistake to use NULL for output_flush_fn if
+ * write_data_fn is not also NULL unless you have built libpng with
+ * PNG_WRITE_FLUSH_SUPPORTED undefined, because in this case libpng's
+ * default flush function, which uses the standard *FILE structure, will
+ * be used.
+ */
+PNG_EXPORT(77, void, png_set_write_fn, (png_structrp png_ptr, png_voidp io_ptr,
+    png_rw_ptr write_data_fn, png_flush_ptr output_flush_fn));
+
+/* Replace the default data input function with a user supplied one. */
+PNG_EXPORT(78, void, png_set_read_fn, (png_structrp png_ptr, png_voidp io_ptr,
+    png_rw_ptr read_data_fn));
+
+/* Return the user pointer associated with the I/O functions */
+PNG_EXPORT(79, png_voidp, png_get_io_ptr, (png_const_structrp png_ptr));
+
+PNG_EXPORT(80, void, png_set_read_status_fn, (png_structrp png_ptr,
+    png_read_status_ptr read_row_fn));
+
+PNG_EXPORT(81, void, png_set_write_status_fn, (png_structrp png_ptr,
+    png_write_status_ptr write_row_fn));
+
+#ifdef PNG_USER_MEM_SUPPORTED
+/* Replace the default memory allocation functions with user supplied one(s). */
+PNG_EXPORT(82, void, png_set_mem_fn, (png_structrp png_ptr, png_voidp mem_ptr,
+    png_malloc_ptr malloc_fn, png_free_ptr free_fn));
+/* Return the user pointer associated with the memory functions */
+PNG_EXPORT(83, png_voidp, png_get_mem_ptr, (png_const_structrp png_ptr));
+#endif
+
+#ifdef PNG_READ_USER_TRANSFORM_SUPPORTED
+PNG_EXPORT(84, void, png_set_read_user_transform_fn, (png_structrp png_ptr,
+    png_user_transform_ptr read_user_transform_fn));
+#endif
+
+#ifdef PNG_WRITE_USER_TRANSFORM_SUPPORTED
+PNG_EXPORT(85, void, png_set_write_user_transform_fn, (png_structrp png_ptr,
+    png_user_transform_ptr write_user_transform_fn));
+#endif
+
+#ifdef PNG_USER_TRANSFORM_PTR_SUPPORTED
+PNG_EXPORT(86, void, png_set_user_transform_info, (png_structrp png_ptr,
+    png_voidp user_transform_ptr, int user_transform_depth,
+    int user_transform_channels));
+/* Return the user pointer associated with the user transform functions */
+PNG_EXPORT(87, png_voidp, png_get_user_transform_ptr,
+    (png_const_structrp png_ptr));
+#endif
+
+#ifdef PNG_USER_TRANSFORM_INFO_SUPPORTED
+/* Return information about the row currently being processed.  Note that these
+ * APIs do not fail but will return unexpected results if called outside a user
+ * transform callback.  Also note that when transforming an interlaced image the
+ * row number is the row number within the sub-image of the interlace pass, so
+ * the value will increase to the height of the sub-image (not the full image)
+ * then reset to 0 for the next pass.
+ *
+ * Use PNG_ROW_FROM_PASS_ROW(row, pass) and PNG_COL_FROM_PASS_COL(col, pass) to
+ * find the output pixel (x,y) given an interlaced sub-image pixel
+ * (row,col,pass).  (See below for these macros.)
+ */
+PNG_EXPORT(217, png_uint_32, png_get_current_row_number, (png_const_structrp));
+PNG_EXPORT(218, png_byte, png_get_current_pass_number, (png_const_structrp));
+#endif
+
+#ifdef PNG_READ_USER_CHUNKS_SUPPORTED
+/* This callback is called only for *unknown* chunks.  If
+ * PNG_HANDLE_AS_UNKNOWN_SUPPORTED is set then it is possible to set known
+ * chunks to be treated as unknown, however in this case the callback must do
+ * any processing required by the chunk (e.g. by calling the appropriate
+ * png_set_ APIs.)
+ *
+ * There is no write support - on write, by default, all the chunks in the
+ * 'unknown' list are written in the specified position.
+ *
+ * The integer return from the callback function is interpreted thus:
+ *
+ * negative: An error occured, png_chunk_error will be called.
+ *     zero: The chunk was not handled, the chunk will be saved. A critical
+ *           chunk will cause an error at this point unless it is to be saved.
+ * positive: The chunk was handled, libpng will ignore/discard it.
+ *
+ * See "INTERACTION WTIH USER CHUNK CALLBACKS" below for important notes about
+ * how this behavior will change in libpng 1.7
+ */
+PNG_EXPORT(88, void, png_set_read_user_chunk_fn, (png_structrp png_ptr,
+    png_voidp user_chunk_ptr, png_user_chunk_ptr read_user_chunk_fn));
+#endif
+
+#ifdef PNG_USER_CHUNKS_SUPPORTED
+PNG_EXPORT(89, png_voidp, png_get_user_chunk_ptr, (png_const_structrp png_ptr));
+#endif
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+/* Sets the function callbacks for the push reader, and a pointer to a
+ * user-defined structure available to the callback functions.
+ */
+PNG_EXPORT(90, void, png_set_progressive_read_fn, (png_structrp png_ptr,
+    png_voidp progressive_ptr, png_progressive_info_ptr info_fn,
+    png_progressive_row_ptr row_fn, png_progressive_end_ptr end_fn));
+
+/* Returns the user pointer associated with the push read functions */
+PNG_EXPORT(91, png_voidp, png_get_progressive_ptr,
+    (png_const_structrp png_ptr));
+
+/* Function to be called when data becomes available */
+PNG_EXPORT(92, void, png_process_data, (png_structrp png_ptr,
+    png_inforp info_ptr, png_bytep buffer, png_size_t buffer_size));
+
+/* A function which may be called *only* within png_process_data to stop the
+ * processing of any more data.  The function returns the number of bytes
+ * remaining, excluding any that libpng has cached internally.  A subsequent
+ * call to png_process_data must supply these bytes again.  If the argument
+ * 'save' is set to true the routine will first save all the pending data and
+ * will always return 0.
+ */
+PNG_EXPORT(219, png_size_t, png_process_data_pause, (png_structrp, int save));
+
+/* A function which may be called *only* outside (after) a call to
+ * png_process_data.  It returns the number of bytes of data to skip in the
+ * input.  Normally it will return 0, but if it returns a non-zero value the
+ * application must skip than number of bytes of input data and pass the
+ * following data to the next call to png_process_data.
+ */
+PNG_EXPORT(220, png_uint_32, png_process_data_skip, (png_structrp));
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+/* Function that combines rows.  'new_row' is a flag that should come from
+ * the callback and be non-NULL if anything needs to be done; the library
+ * stores its own version of the new data internally and ignores the passed
+ * in value.
+ */
+PNG_EXPORT(93, void, png_progressive_combine_row, (png_const_structrp png_ptr,
+    png_bytep old_row, png_const_bytep new_row));
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+#endif /* PNG_PROGRESSIVE_READ_SUPPORTED */
+
+PNG_EXPORTA(94, png_voidp, png_malloc, (png_const_structrp png_ptr,
+    png_alloc_size_t size), PNG_ALLOCATED);
+/* Added at libpng version 1.4.0 */
+PNG_EXPORTA(95, png_voidp, png_calloc, (png_const_structrp png_ptr,
+    png_alloc_size_t size), PNG_ALLOCATED);
+
+/* Added at libpng version 1.2.4 */
+PNG_EXPORTA(96, png_voidp, png_malloc_warn, (png_const_structrp png_ptr,
+    png_alloc_size_t size), PNG_ALLOCATED);
+
+/* Frees a pointer allocated by png_malloc() */
+PNG_EXPORT(97, void, png_free, (png_const_structrp png_ptr, png_voidp ptr));
+
+/* Free data that was allocated internally */
+PNG_EXPORT(98, void, png_free_data, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 free_me, int num));
+
+/* Reassign responsibility for freeing existing data, whether allocated
+ * by libpng or by the application; this works on the png_info structure passed
+ * in, it does not change the state for other png_info structures.
+ *
+ * It is unlikely that this function works correctly as of 1.6.0 and using it
+ * may result either in memory leaks or double free of allocated data.
+ */
+PNG_EXPORTA(99, void, png_data_freer, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int freer, png_uint_32 mask), PNG_DEPRECATED);
+
+/* Assignments for png_data_freer */
+#define PNG_DESTROY_WILL_FREE_DATA 1
+#define PNG_SET_WILL_FREE_DATA 1
+#define PNG_USER_WILL_FREE_DATA 2
+/* Flags for png_ptr->free_me and info_ptr->free_me */
+#define PNG_FREE_HIST 0x0008
+#define PNG_FREE_ICCP 0x0010
+#define PNG_FREE_SPLT 0x0020
+#define PNG_FREE_ROWS 0x0040
+#define PNG_FREE_PCAL 0x0080
+#define PNG_FREE_SCAL 0x0100
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+#  define PNG_FREE_UNKN 0x0200
+#endif
+/*      PNG_FREE_LIST 0x0400    removed in 1.6.0 because it is ignored */
+#define PNG_FREE_PLTE 0x1000
+#define PNG_FREE_TRNS 0x2000
+#define PNG_FREE_TEXT 0x4000
+#define PNG_FREE_ALL  0x7fff
+#define PNG_FREE_MUL  0x4220 /* PNG_FREE_SPLT|PNG_FREE_TEXT|PNG_FREE_UNKN */
+
+#ifdef PNG_USER_MEM_SUPPORTED
+PNG_EXPORTA(100, png_voidp, png_malloc_default, (png_const_structrp png_ptr,
+    png_alloc_size_t size), PNG_ALLOCATED PNG_DEPRECATED);
+PNG_EXPORTA(101, void, png_free_default, (png_const_structrp png_ptr,
+    png_voidp ptr), PNG_DEPRECATED);
+#endif
+
+#ifdef PNG_ERROR_TEXT_SUPPORTED
+/* Fatal error in PNG image of libpng - can't continue */
+PNG_EXPORTA(102, void, png_error, (png_const_structrp png_ptr,
+    png_const_charp error_message), PNG_NORETURN);
+
+/* The same, but the chunk name is prepended to the error string. */
+PNG_EXPORTA(103, void, png_chunk_error, (png_const_structrp png_ptr,
+    png_const_charp error_message), PNG_NORETURN);
+
+#else
+/* Fatal error in PNG image of libpng - can't continue */
+PNG_EXPORTA(104, void, png_err, (png_const_structrp png_ptr), PNG_NORETURN);
+#endif
+
+#ifdef PNG_WARNINGS_SUPPORTED
+/* Non-fatal error in libpng.  Can continue, but may have a problem. */
+PNG_EXPORT(105, void, png_warning, (png_const_structrp png_ptr,
+    png_const_charp warning_message));
+
+/* Non-fatal error in libpng, chunk name is prepended to message. */
+PNG_EXPORT(106, void, png_chunk_warning, (png_const_structrp png_ptr,
+    png_const_charp warning_message));
+#endif
+
+#ifdef PNG_BENIGN_ERRORS_SUPPORTED
+/* Benign error in libpng.  Can continue, but may have a problem.
+ * User can choose whether to handle as a fatal error or as a warning. */
+PNG_EXPORT(107, void, png_benign_error, (png_const_structrp png_ptr,
+    png_const_charp warning_message));
+
+#ifdef PNG_READ_SUPPORTED
+/* Same, chunk name is prepended to message (only during read) */
+PNG_EXPORT(108, void, png_chunk_benign_error, (png_const_structrp png_ptr,
+    png_const_charp warning_message));
+#endif
+
+PNG_EXPORT(109, void, png_set_benign_errors,
+    (png_structrp png_ptr, int allowed));
+#else
+#  ifdef PNG_ALLOW_BENIGN_ERRORS
+#    define png_benign_error png_warning
+#    define png_chunk_benign_error png_chunk_warning
+#  else
+#    define png_benign_error png_error
+#    define png_chunk_benign_error png_chunk_error
+#  endif
+#endif
+
+/* The png_set_<chunk> functions are for storing values in the png_info_struct.
+ * Similarly, the png_get_<chunk> calls are used to read values from the
+ * png_info_struct, either storing the parameters in the passed variables, or
+ * setting pointers into the png_info_struct where the data is stored.  The
+ * png_get_<chunk> functions return a non-zero value if the data was available
+ * in info_ptr, or return zero and do not change any of the parameters if the
+ * data was not available.
+ *
+ * These functions should be used instead of directly accessing png_info
+ * to avoid problems with future changes in the size and internal layout of
+ * png_info_struct.
+ */
+/* Returns "flag" if chunk data is valid in info_ptr. */
+PNG_EXPORT(110, png_uint_32, png_get_valid, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, png_uint_32 flag));
+
+/* Returns number of bytes needed to hold a transformed row. */
+PNG_EXPORT(111, png_size_t, png_get_rowbytes, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+/* Returns row_pointers, which is an array of pointers to scanlines that was
+ * returned from png_read_png().
+ */
+PNG_EXPORT(112, png_bytepp, png_get_rows, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Set row_pointers, which is an array of pointers to scanlines for use
+ * by png_write_png().
+ */
+PNG_EXPORT(113, void, png_set_rows, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_bytepp row_pointers));
+#endif
+
+/* Returns number of color channels in image. */
+PNG_EXPORT(114, png_byte, png_get_channels, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+#ifdef PNG_EASY_ACCESS_SUPPORTED
+/* Returns image width in pixels. */
+PNG_EXPORT(115, png_uint_32, png_get_image_width, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image height in pixels. */
+PNG_EXPORT(116, png_uint_32, png_get_image_height, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image bit_depth. */
+PNG_EXPORT(117, png_byte, png_get_bit_depth, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image color_type. */
+PNG_EXPORT(118, png_byte, png_get_color_type, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image filter_type. */
+PNG_EXPORT(119, png_byte, png_get_filter_type, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image interlace_type. */
+PNG_EXPORT(120, png_byte, png_get_interlace_type, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image compression_type. */
+PNG_EXPORT(121, png_byte, png_get_compression_type, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+
+/* Returns image resolution in pixels per meter, from pHYs chunk data. */
+PNG_EXPORT(122, png_uint_32, png_get_pixels_per_meter,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(123, png_uint_32, png_get_x_pixels_per_meter,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(124, png_uint_32, png_get_y_pixels_per_meter,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+
+/* Returns pixel aspect ratio, computed from pHYs chunk data.  */
+PNG_FP_EXPORT(125, float, png_get_pixel_aspect_ratio,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr))
+PNG_FIXED_EXPORT(210, png_fixed_point, png_get_pixel_aspect_ratio_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr))
+
+/* Returns image x, y offset in pixels or microns, from oFFs chunk data. */
+PNG_EXPORT(126, png_int_32, png_get_x_offset_pixels,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(127, png_int_32, png_get_y_offset_pixels,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(128, png_int_32, png_get_x_offset_microns,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+PNG_EXPORT(129, png_int_32, png_get_y_offset_microns,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+
+#endif /* PNG_EASY_ACCESS_SUPPORTED */
+
+#ifdef PNG_READ_SUPPORTED
+/* Returns pointer to signature string read from PNG header */
+PNG_EXPORT(130, png_const_bytep, png_get_signature, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr));
+#endif
+
+#ifdef PNG_bKGD_SUPPORTED
+PNG_EXPORT(131, png_uint_32, png_get_bKGD, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_color_16p *background));
+#endif
+
+#ifdef PNG_bKGD_SUPPORTED
+PNG_EXPORT(132, void, png_set_bKGD, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_color_16p background));
+#endif
+
+#ifdef PNG_cHRM_SUPPORTED
+PNG_FP_EXPORT(133, png_uint_32, png_get_cHRM, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, double *white_x, double *white_y, double *red_x,
+    double *red_y, double *green_x, double *green_y, double *blue_x,
+    double *blue_y))
+PNG_FP_EXPORT(230, png_uint_32, png_get_cHRM_XYZ, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, double *red_X, double *red_Y, double *red_Z,
+    double *green_X, double *green_Y, double *green_Z, double *blue_X,
+    double *blue_Y, double *blue_Z))
+PNG_FIXED_EXPORT(134, png_uint_32, png_get_cHRM_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *int_white_x, png_fixed_point *int_white_y,
+    png_fixed_point *int_red_x, png_fixed_point *int_red_y,
+    png_fixed_point *int_green_x, png_fixed_point *int_green_y,
+    png_fixed_point *int_blue_x, png_fixed_point *int_blue_y))
+PNG_FIXED_EXPORT(231, png_uint_32, png_get_cHRM_XYZ_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *int_red_X, png_fixed_point *int_red_Y,
+    png_fixed_point *int_red_Z, png_fixed_point *int_green_X,
+    png_fixed_point *int_green_Y, png_fixed_point *int_green_Z,
+    png_fixed_point *int_blue_X, png_fixed_point *int_blue_Y,
+    png_fixed_point *int_blue_Z))
+#endif
+
+#ifdef PNG_cHRM_SUPPORTED
+PNG_FP_EXPORT(135, void, png_set_cHRM, (png_const_structrp png_ptr,
+    png_inforp info_ptr,
+    double white_x, double white_y, double red_x, double red_y, double green_x,
+    double green_y, double blue_x, double blue_y))
+PNG_FP_EXPORT(232, void, png_set_cHRM_XYZ, (png_const_structrp png_ptr,
+    png_inforp info_ptr, double red_X, double red_Y, double red_Z,
+    double green_X, double green_Y, double green_Z, double blue_X,
+    double blue_Y, double blue_Z))
+PNG_FIXED_EXPORT(136, void, png_set_cHRM_fixed, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_fixed_point int_white_x,
+    png_fixed_point int_white_y, png_fixed_point int_red_x,
+    png_fixed_point int_red_y, png_fixed_point int_green_x,
+    png_fixed_point int_green_y, png_fixed_point int_blue_x,
+    png_fixed_point int_blue_y))
+PNG_FIXED_EXPORT(233, void, png_set_cHRM_XYZ_fixed, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_fixed_point int_red_X, png_fixed_point int_red_Y,
+    png_fixed_point int_red_Z, png_fixed_point int_green_X,
+    png_fixed_point int_green_Y, png_fixed_point int_green_Z,
+    png_fixed_point int_blue_X, png_fixed_point int_blue_Y,
+    png_fixed_point int_blue_Z))
+#endif
+
+#ifdef PNG_gAMA_SUPPORTED
+PNG_FP_EXPORT(137, png_uint_32, png_get_gAMA, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, double *file_gamma))
+PNG_FIXED_EXPORT(138, png_uint_32, png_get_gAMA_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *int_file_gamma))
+#endif
+
+#ifdef PNG_gAMA_SUPPORTED
+PNG_FP_EXPORT(139, void, png_set_gAMA, (png_const_structrp png_ptr,
+    png_inforp info_ptr, double file_gamma))
+PNG_FIXED_EXPORT(140, void, png_set_gAMA_fixed, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_fixed_point int_file_gamma))
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+PNG_EXPORT(141, png_uint_32, png_get_hIST, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_uint_16p *hist));
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+PNG_EXPORT(142, void, png_set_hIST, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_uint_16p hist));
+#endif
+
+PNG_EXPORT(143, png_uint_32, png_get_IHDR, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, png_uint_32 *width, png_uint_32 *height,
+    int *bit_depth, int *color_type, int *interlace_method,
+    int *compression_method, int *filter_method));
+
+PNG_EXPORT(144, void, png_set_IHDR, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 width, png_uint_32 height, int bit_depth,
+    int color_type, int interlace_method, int compression_method,
+    int filter_method));
+
+#ifdef PNG_oFFs_SUPPORTED
+PNG_EXPORT(145, png_uint_32, png_get_oFFs, (png_const_structrp png_ptr,
+   png_const_inforp info_ptr, png_int_32 *offset_x, png_int_32 *offset_y,
+   int *unit_type));
+#endif
+
+#ifdef PNG_oFFs_SUPPORTED
+PNG_EXPORT(146, void, png_set_oFFs, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_int_32 offset_x, png_int_32 offset_y,
+    int unit_type));
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+PNG_EXPORT(147, png_uint_32, png_get_pCAL, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_charp *purpose, png_int_32 *X0,
+    png_int_32 *X1, int *type, int *nparams, png_charp *units,
+    png_charpp *params));
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+PNG_EXPORT(148, void, png_set_pCAL, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_charp purpose, png_int_32 X0, png_int_32 X1,
+    int type, int nparams, png_const_charp units, png_charpp params));
+#endif
+
+#ifdef PNG_pHYs_SUPPORTED
+PNG_EXPORT(149, png_uint_32, png_get_pHYs, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, png_uint_32 *res_x, png_uint_32 *res_y,
+    int *unit_type));
+#endif
+
+#ifdef PNG_pHYs_SUPPORTED
+PNG_EXPORT(150, void, png_set_pHYs, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 res_x, png_uint_32 res_y, int unit_type));
+#endif
+
+PNG_EXPORT(151, png_uint_32, png_get_PLTE, (png_const_structrp png_ptr,
+   png_inforp info_ptr, png_colorp *palette, int *num_palette));
+
+PNG_EXPORT(152, void, png_set_PLTE, (png_structrp png_ptr,
+    png_inforp info_ptr, png_const_colorp palette, int num_palette));
+
+#ifdef PNG_sBIT_SUPPORTED
+PNG_EXPORT(153, png_uint_32, png_get_sBIT, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_color_8p *sig_bit));
+#endif
+
+#ifdef PNG_sBIT_SUPPORTED
+PNG_EXPORT(154, void, png_set_sBIT, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_color_8p sig_bit));
+#endif
+
+#ifdef PNG_sRGB_SUPPORTED
+PNG_EXPORT(155, png_uint_32, png_get_sRGB, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, int *file_srgb_intent));
+#endif
+
+#ifdef PNG_sRGB_SUPPORTED
+PNG_EXPORT(156, void, png_set_sRGB, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int srgb_intent));
+PNG_EXPORT(157, void, png_set_sRGB_gAMA_and_cHRM, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int srgb_intent));
+#endif
+
+#ifdef PNG_iCCP_SUPPORTED
+PNG_EXPORT(158, png_uint_32, png_get_iCCP, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_charpp name, int *compression_type,
+    png_bytepp profile, png_uint_32 *proflen));
+#endif
+
+#ifdef PNG_iCCP_SUPPORTED
+PNG_EXPORT(159, void, png_set_iCCP, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_charp name, int compression_type,
+    png_const_bytep profile, png_uint_32 proflen));
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+PNG_EXPORT(160, int, png_get_sPLT, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_sPLT_tpp entries));
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+PNG_EXPORT(161, void, png_set_sPLT, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_sPLT_tp entries, int nentries));
+#endif
+
+#ifdef PNG_TEXT_SUPPORTED
+/* png_get_text also returns the number of text chunks in *num_text */
+PNG_EXPORT(162, int, png_get_text, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_textp *text_ptr, int *num_text));
+#endif
+
+/* Note while png_set_text() will accept a structure whose text,
+ * language, and  translated keywords are NULL pointers, the structure
+ * returned by png_get_text will always contain regular
+ * zero-terminated C strings.  They might be empty strings but
+ * they will never be NULL pointers.
+ */
+
+#ifdef PNG_TEXT_SUPPORTED
+PNG_EXPORT(163, void, png_set_text, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_textp text_ptr, int num_text));
+#endif
+
+#ifdef PNG_tIME_SUPPORTED
+PNG_EXPORT(164, png_uint_32, png_get_tIME, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_timep *mod_time));
+#endif
+
+#ifdef PNG_tIME_SUPPORTED
+PNG_EXPORT(165, void, png_set_tIME, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_timep mod_time));
+#endif
+
+#ifdef PNG_tRNS_SUPPORTED
+PNG_EXPORT(166, png_uint_32, png_get_tRNS, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_bytep *trans_alpha, int *num_trans,
+    png_color_16p *trans_color));
+#endif
+
+#ifdef PNG_tRNS_SUPPORTED
+PNG_EXPORT(167, void, png_set_tRNS, (png_structrp png_ptr,
+    png_inforp info_ptr, png_const_bytep trans_alpha, int num_trans,
+    png_const_color_16p trans_color));
+#endif
+
+#ifdef PNG_sCAL_SUPPORTED
+PNG_FP_EXPORT(168, png_uint_32, png_get_sCAL, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, int *unit, double *width, double *height))
+#if defined(PNG_FLOATING_ARITHMETIC_SUPPORTED) || \
+   defined(PNG_FLOATING_POINT_SUPPORTED)
+/* NOTE: this API is currently implemented using floating point arithmetic,
+ * consequently it can only be used on systems with floating point support.
+ * In any case the range of values supported by png_fixed_point is small and it
+ * is highly recommended that png_get_sCAL_s be used instead.
+ */
+PNG_FIXED_EXPORT(214, png_uint_32, png_get_sCAL_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr, int *unit,
+    png_fixed_point *width, png_fixed_point *height))
+#endif
+PNG_EXPORT(169, png_uint_32, png_get_sCAL_s,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr, int *unit,
+    png_charpp swidth, png_charpp sheight));
+
+PNG_FP_EXPORT(170, void, png_set_sCAL, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int unit, double width, double height))
+PNG_FIXED_EXPORT(213, void, png_set_sCAL_fixed, (png_const_structrp png_ptr,
+   png_inforp info_ptr, int unit, png_fixed_point width,
+   png_fixed_point height))
+PNG_EXPORT(171, void, png_set_sCAL_s, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int unit,
+    png_const_charp swidth, png_const_charp sheight));
+#endif /* PNG_sCAL_SUPPORTED */
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+/* Provide the default handling for all unknown chunks or, optionally, for
+ * specific unknown chunks.
+ *
+ * NOTE: prior to 1.6.0 the handling specified for particular chunks on read was
+ * ignored and the default was used, the per-chunk setting only had an effect on
+ * write.  If you wish to have chunk-specific handling on read in code that must
+ * work on earlier versions you must use a user chunk callback to specify the
+ * desired handling (keep or discard.)
+ *
+ * The 'keep' parameter is a PNG_HANDLE_CHUNK_ value as listed below.  The
+ * parameter is interpreted as follows:
+ *
+ * READ:
+ *    PNG_HANDLE_CHUNK_AS_DEFAULT:
+ *       Known chunks: do normal libpng processing, do not keep the chunk (but
+ *          see the comments below about PNG_HANDLE_AS_UNKNOWN_SUPPORTED)
+ *       Unknown chunks: for a specific chunk use the global default, when used
+ *          as the default discard the chunk data.
+ *    PNG_HANDLE_CHUNK_NEVER:
+ *       Discard the chunk data.
+ *    PNG_HANDLE_CHUNK_IF_SAFE:
+ *       Keep the chunk data if the chunk is not critical else raise a chunk
+ *       error.
+ *    PNG_HANDLE_CHUNK_ALWAYS:
+ *       Keep the chunk data.
+ *
+ * If the chunk data is saved it can be retrieved using png_get_unknown_chunks,
+ * below.  Notice that specifying "AS_DEFAULT" as a global default is equivalent
+ * to specifying "NEVER", however when "AS_DEFAULT" is used for specific chunks
+ * it simply resets the behavior to the libpng default.
+ *
+ * INTERACTION WTIH USER CHUNK CALLBACKS:
+ * The per-chunk handling is always used when there is a png_user_chunk_ptr
+ * callback and the callback returns 0; the chunk is then always stored *unless*
+ * it is critical and the per-chunk setting is other than ALWAYS.  Notice that
+ * the global default is *not* used in this case.  (In effect the per-chunk
+ * value is incremented to at least IF_SAFE.)
+ *
+ * IMPORTANT NOTE: this behavior will change in libpng 1.7 - the global and
+ * per-chunk defaults will be honored.  If you want to preserve the current
+ * behavior when your callback returns 0 you must set PNG_HANDLE_CHUNK_IF_SAFE
+ * as the default - if you don't do this libpng 1.6 will issue a warning.
+ *
+ * If you want unhandled unknown chunks to be discarded in libpng 1.6 and
+ * earlier simply return '1' (handled).
+ *
+ * PNG_HANDLE_AS_UNKNOWN_SUPPORTED:
+ *    If this is *not* set known chunks will always be handled by libpng and
+ *    will never be stored in the unknown chunk list.  Known chunks listed to
+ *    png_set_keep_unknown_chunks will have no effect.  If it is set then known
+ *    chunks listed with a keep other than AS_DEFAULT will *never* be processed
+ *    by libpng, in addition critical chunks must either be processed by the
+ *    callback or saved.
+ *
+ *    The IHDR and IEND chunks must not be listed.  Because this turns off the
+ *    default handling for chunks that would otherwise be recognized the
+ *    behavior of libpng transformations may well become incorrect!
+ *
+ * WRITE:
+ *    When writing chunks the options only apply to the chunks specified by
+ *    png_set_unknown_chunks (below), libpng will *always* write known chunks
+ *    required by png_set_ calls and will always write the core critical chunks
+ *    (as required for PLTE).
+ *
+ *    Each chunk in the png_set_unknown_chunks list is looked up in the
+ *    png_set_keep_unknown_chunks list to find the keep setting, this is then
+ *    interpreted as follows:
+ *
+ *    PNG_HANDLE_CHUNK_AS_DEFAULT:
+ *       Write safe-to-copy chunks and write other chunks if the global
+ *       default is set to _ALWAYS, otherwise don't write this chunk.
+ *    PNG_HANDLE_CHUNK_NEVER:
+ *       Do not write the chunk.
+ *    PNG_HANDLE_CHUNK_IF_SAFE:
+ *       Write the chunk if it is safe-to-copy, otherwise do not write it.
+ *    PNG_HANDLE_CHUNK_ALWAYS:
+ *       Write the chunk.
+ *
+ * Note that the default behavior is effectively the opposite of the read case -
+ * in read unknown chunks are not stored by default, in write they are written
+ * by default.  Also the behavior of PNG_HANDLE_CHUNK_IF_SAFE is very different
+ * - on write the safe-to-copy bit is checked, on read the critical bit is
+ * checked and on read if the chunk is critical an error will be raised.
+ *
+ * num_chunks:
+ * ===========
+ *    If num_chunks is positive, then the "keep" parameter specifies the manner
+ *    for handling only those chunks appearing in the chunk_list array,
+ *    otherwise the chunk list array is ignored.
+ *
+ *    If num_chunks is 0 the "keep" parameter specifies the default behavior for
+ *    unknown chunks, as described above.
+ *
+ *    If num_chunks is negative, then the "keep" parameter specifies the manner
+ *    for handling all unknown chunks plus all chunks recognized by libpng
+ *    except for the IHDR, PLTE, tRNS, IDAT, and IEND chunks (which continue to
+ *    be processed by libpng.
+ */
+PNG_EXPORT(172, void, png_set_keep_unknown_chunks, (png_structrp png_ptr,
+    int keep, png_const_bytep chunk_list, int num_chunks));
+
+/* The "keep" PNG_HANDLE_CHUNK_ parameter for the specified chunk is returned;
+ * the result is therefore true (non-zero) if special handling is required,
+ * false for the default handling.
+ */
+PNG_EXPORT(173, int, png_handle_as_unknown, (png_const_structrp png_ptr,
+    png_const_bytep chunk_name));
+#endif
+
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+PNG_EXPORT(174, void, png_set_unknown_chunks, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_unknown_chunkp unknowns,
+    int num_unknowns));
+   /* NOTE: prior to 1.6.0 this routine set the 'location' field of the added
+    * unknowns to the location currently stored in the png_struct.  This is
+    * invariably the wrong value on write.  To fix this call the following API
+    * for each chunk in the list with the correct location.  If you know your
+    * code won't be compiled on earlier versions you can rely on
+    * png_set_unknown_chunks(write-ptr, png_get_unknown_chunks(read-ptr)) doing
+    * the correct thing.
+    */
+
+PNG_EXPORT(175, void, png_set_unknown_chunk_location,
+    (png_const_structrp png_ptr, png_inforp info_ptr, int chunk, int location));
+
+PNG_EXPORT(176, int, png_get_unknown_chunks, (png_const_structrp png_ptr,
+    png_inforp info_ptr, png_unknown_chunkpp entries));
+#endif
+
+/* Png_free_data() will turn off the "valid" flag for anything it frees.
+ * If you need to turn it off for a chunk that your application has freed,
+ * you can use png_set_invalid(png_ptr, info_ptr, PNG_INFO_CHNK);
+ */
+PNG_EXPORT(177, void, png_set_invalid, (png_const_structrp png_ptr,
+    png_inforp info_ptr, int mask));
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+/* The "params" pointer is currently not used and is for future expansion. */
+PNG_EXPORT(178, void, png_read_png, (png_structrp png_ptr, png_inforp info_ptr,
+    int transforms, png_voidp params));
+PNG_EXPORT(179, void, png_write_png, (png_structrp png_ptr, png_inforp info_ptr,
+    int transforms, png_voidp params));
+#endif
+
+PNG_EXPORT(180, png_const_charp, png_get_copyright,
+    (png_const_structrp png_ptr));
+PNG_EXPORT(181, png_const_charp, png_get_header_ver,
+    (png_const_structrp png_ptr));
+PNG_EXPORT(182, png_const_charp, png_get_header_version,
+    (png_const_structrp png_ptr));
+PNG_EXPORT(183, png_const_charp, png_get_libpng_ver,
+    (png_const_structrp png_ptr));
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+PNG_EXPORT(184, png_uint_32, png_permit_mng_features, (png_structrp png_ptr,
+    png_uint_32 mng_features_permitted));
+#endif
+
+/* For use in png_set_keep_unknown, added to version 1.2.6 */
+#define PNG_HANDLE_CHUNK_AS_DEFAULT   0
+#define PNG_HANDLE_CHUNK_NEVER        1
+#define PNG_HANDLE_CHUNK_IF_SAFE      2
+#define PNG_HANDLE_CHUNK_ALWAYS       3
+#define PNG_HANDLE_CHUNK_LAST         4
+
+/* Strip the prepended error numbers ("#nnn ") from error and warning
+ * messages before passing them to the error or warning handler.
+ */
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+PNG_EXPORT(185, void, png_set_strip_error_numbers, (png_structrp png_ptr,
+    png_uint_32 strip_mode));
+#endif
+
+/* Added in libpng-1.2.6 */
+#ifdef PNG_SET_USER_LIMITS_SUPPORTED
+PNG_EXPORT(186, void, png_set_user_limits, (png_structrp png_ptr,
+    png_uint_32 user_width_max, png_uint_32 user_height_max));
+PNG_EXPORT(187, png_uint_32, png_get_user_width_max,
+    (png_const_structrp png_ptr));
+PNG_EXPORT(188, png_uint_32, png_get_user_height_max,
+    (png_const_structrp png_ptr));
+/* Added in libpng-1.4.0 */
+PNG_EXPORT(189, void, png_set_chunk_cache_max, (png_structrp png_ptr,
+    png_uint_32 user_chunk_cache_max));
+PNG_EXPORT(190, png_uint_32, png_get_chunk_cache_max,
+    (png_const_structrp png_ptr));
+/* Added in libpng-1.4.1 */
+PNG_EXPORT(191, void, png_set_chunk_malloc_max, (png_structrp png_ptr,
+    png_alloc_size_t user_chunk_cache_max));
+PNG_EXPORT(192, png_alloc_size_t, png_get_chunk_malloc_max,
+    (png_const_structrp png_ptr));
+#endif
+
+#if defined(PNG_INCH_CONVERSIONS_SUPPORTED)
+PNG_EXPORT(193, png_uint_32, png_get_pixels_per_inch,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+
+PNG_EXPORT(194, png_uint_32, png_get_x_pixels_per_inch,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+
+PNG_EXPORT(195, png_uint_32, png_get_y_pixels_per_inch,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr));
+
+PNG_FP_EXPORT(196, float, png_get_x_offset_inches,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr))
+#ifdef PNG_FIXED_POINT_SUPPORTED /* otherwise not implemented. */
+PNG_FIXED_EXPORT(211, png_fixed_point, png_get_x_offset_inches_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr))
+#endif
+
+PNG_FP_EXPORT(197, float, png_get_y_offset_inches, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr))
+#ifdef PNG_FIXED_POINT_SUPPORTED /* otherwise not implemented. */
+PNG_FIXED_EXPORT(212, png_fixed_point, png_get_y_offset_inches_fixed,
+    (png_const_structrp png_ptr, png_const_inforp info_ptr))
+#endif
+
+#  ifdef PNG_pHYs_SUPPORTED
+PNG_EXPORT(198, png_uint_32, png_get_pHYs_dpi, (png_const_structrp png_ptr,
+    png_const_inforp info_ptr, png_uint_32 *res_x, png_uint_32 *res_y,
+    int *unit_type));
+#  endif /* PNG_pHYs_SUPPORTED */
+#endif  /* PNG_INCH_CONVERSIONS_SUPPORTED */
+
+/* Added in libpng-1.4.0 */
+#ifdef PNG_IO_STATE_SUPPORTED
+PNG_EXPORT(199, png_uint_32, png_get_io_state, (png_const_structrp png_ptr));
+
+/* Removed from libpng 1.6; use png_get_io_chunk_type. */
+PNG_REMOVED(200, png_const_bytep, png_get_io_chunk_name, (png_structrp png_ptr),
+    PNG_DEPRECATED)
+
+PNG_EXPORT(216, png_uint_32, png_get_io_chunk_type,
+    (png_const_structrp png_ptr));
+
+/* The flags returned by png_get_io_state() are the following: */
+#  define PNG_IO_NONE        0x0000   /* no I/O at this moment */
+#  define PNG_IO_READING     0x0001   /* currently reading */
+#  define PNG_IO_WRITING     0x0002   /* currently writing */
+#  define PNG_IO_SIGNATURE   0x0010   /* currently at the file signature */
+#  define PNG_IO_CHUNK_HDR   0x0020   /* currently at the chunk header */
+#  define PNG_IO_CHUNK_DATA  0x0040   /* currently at the chunk data */
+#  define PNG_IO_CHUNK_CRC   0x0080   /* currently at the chunk crc */
+#  define PNG_IO_MASK_OP     0x000f   /* current operation: reading/writing */
+#  define PNG_IO_MASK_LOC    0x00f0   /* current location: sig/hdr/data/crc */
+#endif /* ?PNG_IO_STATE_SUPPORTED */
+
+/* Interlace support.  The following macros are always defined so that if
+ * libpng interlace handling is turned off the macros may be used to handle
+ * interlaced images within the application.
+ */
+#define PNG_INTERLACE_ADAM7_PASSES 7
+
+/* Two macros to return the first row and first column of the original,
+ * full, image which appears in a given pass.  'pass' is in the range 0
+ * to 6 and the result is in the range 0 to 7.
+ */
+#define PNG_PASS_START_ROW(pass) (((1&~(pass))<<(3-((pass)>>1)))&7)
+#define PNG_PASS_START_COL(pass) (((1& (pass))<<(3-(((pass)+1)>>1)))&7)
+
+/* A macro to return the offset between pixels in the output row for a pair of
+ * pixels in the input - effectively the inverse of the 'COL_SHIFT' macro that
+ * follows.  Note that ROW_OFFSET is the offset from one row to the next whereas
+ * COL_OFFSET is from one column to the next, within a row.
+ */
+#define PNG_PASS_ROW_OFFSET(pass) ((pass)>2?(8>>(((pass)-1)>>1)):8)
+#define PNG_PASS_COL_OFFSET(pass) (1<<((7-(pass))>>1))
+
+/* Two macros to help evaluate the number of rows or columns in each
+ * pass.  This is expressed as a shift - effectively log2 of the number or
+ * rows or columns in each 8x8 tile of the original image.
+ */
+#define PNG_PASS_ROW_SHIFT(pass) ((pass)>2?(8-(pass))>>1:3)
+#define PNG_PASS_COL_SHIFT(pass) ((pass)>1?(7-(pass))>>1:3)
+
+/* Hence two macros to determine the number of rows or columns in a given
+ * pass of an image given its height or width.  In fact these macros may
+ * return non-zero even though the sub-image is empty, because the other
+ * dimension may be empty for a small image.
+ */
+#define PNG_PASS_ROWS(height, pass) (((height)+(((1<<PNG_PASS_ROW_SHIFT(pass))\
+   -1)-PNG_PASS_START_ROW(pass)))>>PNG_PASS_ROW_SHIFT(pass))
+#define PNG_PASS_COLS(width, pass) (((width)+(((1<<PNG_PASS_COL_SHIFT(pass))\
+   -1)-PNG_PASS_START_COL(pass)))>>PNG_PASS_COL_SHIFT(pass))
+
+/* For the reader row callbacks (both progressive and sequential) it is
+ * necessary to find the row in the output image given a row in an interlaced
+ * image, so two more macros:
+ */
+#define PNG_ROW_FROM_PASS_ROW(y_in, pass) \
+   (((y_in)<<PNG_PASS_ROW_SHIFT(pass))+PNG_PASS_START_ROW(pass))
+#define PNG_COL_FROM_PASS_COL(x_in, pass) \
+   (((x_in)<<PNG_PASS_COL_SHIFT(pass))+PNG_PASS_START_COL(pass))
+
+/* Two macros which return a boolean (0 or 1) saying whether the given row
+ * or column is in a particular pass.  These use a common utility macro that
+ * returns a mask for a given pass - the offset 'off' selects the row or
+ * column version.  The mask has the appropriate bit set for each column in
+ * the tile.
+ */
+#define PNG_PASS_MASK(pass,off) ( \
+   ((0x110145AF>>(((7-(off))-(pass))<<2)) & 0xF) | \
+   ((0x01145AF0>>(((7-(off))-(pass))<<2)) & 0xF0))
+
+#define PNG_ROW_IN_INTERLACE_PASS(y, pass) \
+   ((PNG_PASS_MASK(pass,0) >> ((y)&7)) & 1)
+#define PNG_COL_IN_INTERLACE_PASS(x, pass) \
+   ((PNG_PASS_MASK(pass,1) >> ((x)&7)) & 1)
+
+#ifdef PNG_READ_COMPOSITE_NODIV_SUPPORTED
+/* With these routines we avoid an integer divide, which will be slower on
+ * most machines.  However, it does take more operations than the corresponding
+ * divide method, so it may be slower on a few RISC systems.  There are two
+ * shifts (by 8 or 16 bits) and an addition, versus a single integer divide.
+ *
+ * Note that the rounding factors are NOT supposed to be the same!  128 and
+ * 32768 are correct for the NODIV code; 127 and 32767 are correct for the
+ * standard method.
+ *
+ * [Optimized code by Greg Roelofs and Mark Adler...blame us for bugs. :-) ]
+ */
+
+ /* fg and bg should be in `gamma 1.0' space; alpha is the opacity */
+
+#  define png_composite(composite, fg, alpha, bg)         \
+     { png_uint_16 temp = (png_uint_16)((png_uint_16)(fg) \
+           * (png_uint_16)(alpha)                         \
+           + (png_uint_16)(bg)*(png_uint_16)(255          \
+           - (png_uint_16)(alpha)) + 128);                \
+       (composite) = (png_byte)((temp + (temp >> 8)) >> 8); }
+
+#  define png_composite_16(composite, fg, alpha, bg)       \
+     { png_uint_32 temp = (png_uint_32)((png_uint_32)(fg)  \
+           * (png_uint_32)(alpha)                          \
+           + (png_uint_32)(bg)*(65535                      \
+           - (png_uint_32)(alpha)) + 32768);               \
+       (composite) = (png_uint_16)((temp + (temp >> 16)) >> 16); }
+
+#else  /* Standard method using integer division */
+
+#  define png_composite(composite, fg, alpha, bg)                          \
+     (composite) = (png_byte)(((png_uint_16)(fg) * (png_uint_16)(alpha) +  \
+     (png_uint_16)(bg) * (png_uint_16)(255 - (png_uint_16)(alpha)) +       \
+     127) / 255)
+
+#  define png_composite_16(composite, fg, alpha, bg)                         \
+     (composite) = (png_uint_16)(((png_uint_32)(fg) * (png_uint_32)(alpha) + \
+     (png_uint_32)(bg)*(png_uint_32)(65535 - (png_uint_32)(alpha)) +         \
+     32767) / 65535)
+#endif /* PNG_READ_COMPOSITE_NODIV_SUPPORTED */
+
+#ifdef PNG_READ_INT_FUNCTIONS_SUPPORTED
+PNG_EXPORT(201, png_uint_32, png_get_uint_32, (png_const_bytep buf));
+PNG_EXPORT(202, png_uint_16, png_get_uint_16, (png_const_bytep buf));
+PNG_EXPORT(203, png_int_32, png_get_int_32, (png_const_bytep buf));
+#endif
+
+PNG_EXPORT(204, png_uint_32, png_get_uint_31, (png_const_structrp png_ptr,
+    png_const_bytep buf));
+/* No png_get_int_16 -- may be added if there's a real need for it. */
+
+/* Place a 32-bit number into a buffer in PNG byte order (big-endian). */
+#ifdef PNG_WRITE_INT_FUNCTIONS_SUPPORTED
+PNG_EXPORT(205, void, png_save_uint_32, (png_bytep buf, png_uint_32 i));
+#endif
+#ifdef PNG_SAVE_INT_32_SUPPORTED
+PNG_EXPORT(206, void, png_save_int_32, (png_bytep buf, png_int_32 i));
+#endif
+
+/* Place a 16-bit number into a buffer in PNG byte order.
+ * The parameter is declared unsigned int, not png_uint_16,
+ * just to avoid potential problems on pre-ANSI C compilers.
+ */
+#ifdef PNG_WRITE_INT_FUNCTIONS_SUPPORTED
+PNG_EXPORT(207, void, png_save_uint_16, (png_bytep buf, unsigned int i));
+/* No png_save_int_16 -- may be added if there's a real need for it. */
+#endif
+
+#ifdef PNG_USE_READ_MACROS
+/* Inline macros to do direct reads of bytes from the input buffer.
+ * The png_get_int_32() routine assumes we are using two's complement
+ * format for negative values, which is almost certainly true.
+ */
+#  define PNG_get_uint_32(buf) \
+     (((png_uint_32)(*(buf)) << 24) + \
+      ((png_uint_32)(*((buf) + 1)) << 16) + \
+      ((png_uint_32)(*((buf) + 2)) << 8) + \
+      ((png_uint_32)(*((buf) + 3))))
+
+   /* From libpng-1.4.0 until 1.4.4, the png_get_uint_16 macro (but not the
+    * function) incorrectly returned a value of type png_uint_32.
+    */
+#  define PNG_get_uint_16(buf) \
+     ((png_uint_16) \
+      (((unsigned int)(*(buf)) << 8) + \
+       ((unsigned int)(*((buf) + 1)))))
+
+#  define PNG_get_int_32(buf) \
+     ((png_int_32)((*(buf) & 0x80) \
+      ? -((png_int_32)((png_get_uint_32(buf) ^ 0xffffffffL) + 1)) \
+      : (png_int_32)png_get_uint_32(buf)))
+
+   /* If PNG_PREFIX is defined the same thing as below happens in pnglibconf.h,
+    * but defining a macro name prefixed with PNG_PREFIX.
+    */
+#  ifndef PNG_PREFIX
+#     define png_get_uint_32(buf) PNG_get_uint_32(buf)
+#     define png_get_uint_16(buf) PNG_get_uint_16(buf)
+#     define png_get_int_32(buf)  PNG_get_int_32(buf)
+#  endif
+#else
+#  ifdef PNG_PREFIX
+      /* No macros; revert to the (redefined) function */
+#     define PNG_get_uint_32 (png_get_uint_32)
+#     define PNG_get_uint_16 (png_get_uint_16)
+#     define PNG_get_int_32  (png_get_int_32)
+#  endif
+#endif
+
+/*******************************************************************************
+ *  SIMPLIFIED API
+ *******************************************************************************
+ *
+ * Please read the documentation in libpng-manual.txt (TODO: write said
+ * documentation) if you don't understand what follows.
+ *
+ * The simplified API hides the details of both libpng and the PNG file format
+ * itself.  It allows PNG files to be read into a very limited number of
+ * in-memory bitmap formats or to be written from the same formats.  If these
+ * formats do not accomodate your needs then you can, and should, use the more
+ * sophisticated APIs above - these support a wide variety of in-memory formats
+ * and a wide variety of sophisticated transformations to those formats as well
+ * as a wide variety of APIs to manipulate ancillary information.
+ *
+ * To read a PNG file using the simplified API:
+ *
+ * 1) Declare a 'png_image' structure (see below) on the stack and set the
+ *    version field to PNG_IMAGE_VERSION.
+ * 2) Call the appropriate png_image_begin_read... function.
+ * 3) Set the png_image 'format' member to the required sample format.
+ * 4) Allocate a buffer for the image and, if required, the color-map.
+ * 5) Call png_image_finish_read to read the image and, if required, the
+ *    color-map into your buffers.
+ *
+ * There are no restrictions on the format of the PNG input itself; all valid
+ * color types, bit depths, and interlace methods are acceptable, and the
+ * input image is transformed as necessary to the requested in-memory format
+ * during the png_image_finish_read() step.  The only caveat is that if you
+ * request a color-mapped image from a PNG that is full-color or makes
+ * complex use of an alpha channel the transformation is extremely lossy and the
+ * result may look terrible.
+ *
+ * To write a PNG file using the simplified API:
+ *
+ * 1) Declare a 'png_image' structure on the stack and memset() it to all zero.
+ * 2) Initialize the members of the structure that describe the image, setting
+ *    the 'format' member to the format of the image samples.
+ * 3) Call the appropriate png_image_write... function with a pointer to the
+ *    image and, if necessary, the color-map to write the PNG data.
+ *
+ * png_image is a structure that describes the in-memory format of an image
+ * when it is being read or defines the in-memory format of an image that you
+ * need to write:
+ */
+#define PNG_IMAGE_VERSION 1
+
+typedef struct png_control *png_controlp;
+typedef struct
+{
+   png_controlp opaque;    /* Initialize to NULL, free with png_image_free */
+   png_uint_32  version;   /* Set to PNG_IMAGE_VERSION */
+   png_uint_32  width;     /* Image width in pixels (columns) */
+   png_uint_32  height;    /* Image height in pixels (rows) */
+   png_uint_32  format;    /* Image format as defined below */
+   png_uint_32  flags;     /* A bit mask containing informational flags */
+   png_uint_32  colormap_entries;
+                           /* Number of entries in the color-map */
+
+   /* In the event of an error or warning the following field will be set to a
+    * non-zero value and the 'message' field will contain a '\0' terminated
+    * string with the libpng error or warning message.  If both warnings and
+    * an error were encountered, only the error is recorded.  If there
+    * are multiple warnings, only the first one is recorded.
+    *
+    * The upper 30 bits of this value are reserved, the low two bits contain
+    * a value as follows:
+    */
+#  define PNG_IMAGE_WARNING 1
+#  define PNG_IMAGE_ERROR 2
+   /*
+    * The result is a two bit code such that a value more than 1 indicates
+    * a failure in the API just called:
+    *
+    *    0 - no warning or error
+    *    1 - warning
+    *    2 - error
+    *    3 - error preceded by warning
+    */
+#  define PNG_IMAGE_FAILED(png_cntrl) ((((png_cntrl).warning_or_error)&0x03)>1)
+
+   png_uint_32  warning_or_error;
+
+   char         message[64];
+} png_image, *png_imagep;
+
+/* The samples of the image have one to four channels whose components have
+ * original values in the range 0 to 1.0:
+ *
+ * 1: A single gray or luminance channel (G).
+ * 2: A gray/luminance channel and an alpha channel (GA).
+ * 3: Three red, green, blue color channels (RGB).
+ * 4: Three color channels and an alpha channel (RGBA).
+ *
+ * The components are encoded in one of two ways:
+ *
+ * a) As a small integer, value 0..255, contained in a single byte.  For the
+ * alpha channel the original value is simply value/255.  For the color or
+ * luminance channels the value is encoded according to the sRGB specification
+ * and matches the 8-bit format expected by typical display devices.
+ *
+ * The color/gray channels are not scaled (pre-multiplied) by the alpha
+ * channel and are suitable for passing to color management software.
+ *
+ * b) As a value in the range 0..65535, contained in a 2-byte integer.  All
+ * channels can be converted to the original value by dividing by 65535; all
+ * channels are linear.  Color channels use the RGB encoding (RGB end-points) of
+ * the sRGB specification.  This encoding is identified by the
+ * PNG_FORMAT_FLAG_LINEAR flag below.
+ *
+ * When the simplified API needs to convert between sRGB and linear colorspaces,
+ * the actual sRGB transfer curve defined in the sRGB specification (see the
+ * article at http://en.wikipedia.org/wiki/SRGB) is used, not the gamma=1/2.2
+ * approximation used elsewhere in libpng.
+ *
+ * When an alpha channel is present it is expected to denote pixel coverage
+ * of the color or luminance channels and is returned as an associated alpha
+ * channel: the color/gray channels are scaled (pre-multiplied) by the alpha
+ * value.
+ *
+ * The samples are either contained directly in the image data, between 1 and 8
+ * bytes per pixel according to the encoding, or are held in a color-map indexed
+ * by bytes in the image data.  In the case of a color-map the color-map entries
+ * are individual samples, encoded as above, and the image data has one byte per
+ * pixel to select the relevant sample from the color-map.
+ */
+
+/* PNG_FORMAT_*
+ *
+ * #defines to be used in png_image::format.  Each #define identifies a
+ * particular layout of sample data and, if present, alpha values.  There are
+ * separate defines for each of the two component encodings.
+ *
+ * A format is built up using single bit flag values.  All combinations are
+ * valid.  Formats can be built up from the flag values or you can use one of
+ * the predefined values below.  When testing formats always use the FORMAT_FLAG
+ * macros to test for individual features - future versions of the library may
+ * add new flags.
+ *
+ * When reading or writing color-mapped images the format should be set to the
+ * format of the entries in the color-map then png_image_{read,write}_colormap
+ * called to read or write the color-map and set the format correctly for the
+ * image data.  Do not set the PNG_FORMAT_FLAG_COLORMAP bit directly!
+ *
+ * NOTE: libpng can be built with particular features disabled, if you see
+ * compiler errors because the definition of one of the following flags has been
+ * compiled out it is because libpng does not have the required support.  It is
+ * possible, however, for the libpng configuration to enable the format on just
+ * read or just write; in that case you may see an error at run time.  You can
+ * guard against this by checking for the definition of the appropriate
+ * "_SUPPORTED" macro, one of:
+ *
+ *    PNG_SIMPLIFIED_{READ,WRITE}_{BGR,AFIRST}_SUPPORTED
+ */
+#define PNG_FORMAT_FLAG_ALPHA    0x01U /* format with an alpha channel */
+#define PNG_FORMAT_FLAG_COLOR    0x02U /* color format: otherwise grayscale */
+#define PNG_FORMAT_FLAG_LINEAR   0x04U /* 2 byte channels else 1 byte */
+#define PNG_FORMAT_FLAG_COLORMAP 0x08U /* image data is color-mapped */
+
+#ifdef PNG_FORMAT_BGR_SUPPORTED
+#  define PNG_FORMAT_FLAG_BGR    0x10U /* BGR colors, else order is RGB */
+#endif
+
+#ifdef PNG_FORMAT_AFIRST_SUPPORTED
+#  define PNG_FORMAT_FLAG_AFIRST 0x20U /* alpha channel comes first */
+#endif
+
+/* Commonly used formats have predefined macros.
+ *
+ * First the single byte (sRGB) formats:
+ */
+#define PNG_FORMAT_GRAY 0
+#define PNG_FORMAT_GA   PNG_FORMAT_FLAG_ALPHA
+#define PNG_FORMAT_AG   (PNG_FORMAT_GA|PNG_FORMAT_FLAG_AFIRST)
+#define PNG_FORMAT_RGB  PNG_FORMAT_FLAG_COLOR
+#define PNG_FORMAT_BGR  (PNG_FORMAT_FLAG_COLOR|PNG_FORMAT_FLAG_BGR)
+#define PNG_FORMAT_RGBA (PNG_FORMAT_RGB|PNG_FORMAT_FLAG_ALPHA)
+#define PNG_FORMAT_ARGB (PNG_FORMAT_RGBA|PNG_FORMAT_FLAG_AFIRST)
+#define PNG_FORMAT_BGRA (PNG_FORMAT_BGR|PNG_FORMAT_FLAG_ALPHA)
+#define PNG_FORMAT_ABGR (PNG_FORMAT_BGRA|PNG_FORMAT_FLAG_AFIRST)
+
+/* Then the linear 2-byte formats.  When naming these "Y" is used to
+ * indicate a luminance (gray) channel.
+ */
+#define PNG_FORMAT_LINEAR_Y PNG_FORMAT_FLAG_LINEAR
+#define PNG_FORMAT_LINEAR_Y_ALPHA (PNG_FORMAT_FLAG_LINEAR|PNG_FORMAT_FLAG_ALPHA)
+#define PNG_FORMAT_LINEAR_RGB (PNG_FORMAT_FLAG_LINEAR|PNG_FORMAT_FLAG_COLOR)
+#define PNG_FORMAT_LINEAR_RGB_ALPHA \
+   (PNG_FORMAT_FLAG_LINEAR|PNG_FORMAT_FLAG_COLOR|PNG_FORMAT_FLAG_ALPHA)
+
+/* With color-mapped formats the image data is one byte for each pixel, the byte
+ * is an index into the color-map which is formatted as above.  To obtain a
+ * color-mapped format it is sufficient just to add the PNG_FOMAT_FLAG_COLORMAP
+ * to one of the above definitions, or you can use one of the definitions below.
+ */
+#define PNG_FORMAT_RGB_COLORMAP  (PNG_FORMAT_RGB|PNG_FORMAT_FLAG_COLORMAP)
+#define PNG_FORMAT_BGR_COLORMAP  (PNG_FORMAT_BGR|PNG_FORMAT_FLAG_COLORMAP)
+#define PNG_FORMAT_RGBA_COLORMAP (PNG_FORMAT_RGBA|PNG_FORMAT_FLAG_COLORMAP)
+#define PNG_FORMAT_ARGB_COLORMAP (PNG_FORMAT_ARGB|PNG_FORMAT_FLAG_COLORMAP)
+#define PNG_FORMAT_BGRA_COLORMAP (PNG_FORMAT_BGRA|PNG_FORMAT_FLAG_COLORMAP)
+#define PNG_FORMAT_ABGR_COLORMAP (PNG_FORMAT_ABGR|PNG_FORMAT_FLAG_COLORMAP)
+
+/* PNG_IMAGE macros
+ *
+ * These are convenience macros to derive information from a png_image
+ * structure.  The PNG_IMAGE_SAMPLE_ macros return values appropriate to the
+ * actual image sample values - either the entries in the color-map or the
+ * pixels in the image.  The PNG_IMAGE_PIXEL_ macros return corresponding values
+ * for the pixels and will always return 1 for color-mapped formats.  The
+ * remaining macros return information about the rows in the image and the
+ * complete image.
+ *
+ * NOTE: All the macros that take a png_image::format parameter are compile time
+ * constants if the format parameter is, itself, a constant.  Therefore these
+ * macros can be used in array declarations and case labels where required.
+ * Similarly the macros are also pre-processor constants (sizeof is not used) so
+ * they can be used in #if tests.
+ *
+ * First the information about the samples.
+ */
+#define PNG_IMAGE_SAMPLE_CHANNELS(fmt)\
+   (((fmt)&(PNG_FORMAT_FLAG_COLOR|PNG_FORMAT_FLAG_ALPHA))+1)
+   /* Return the total number of channels in a given format: 1..4 */
+
+#define PNG_IMAGE_SAMPLE_COMPONENT_SIZE(fmt)\
+   ((((fmt) & PNG_FORMAT_FLAG_LINEAR) >> 2)+1)
+   /* Return the size in bytes of a single component of a pixel or color-map
+    * entry (as appropriate) in the image: 1 or 2.
+    */
+
+#define PNG_IMAGE_SAMPLE_SIZE(fmt)\
+   (PNG_IMAGE_SAMPLE_CHANNELS(fmt) * PNG_IMAGE_SAMPLE_COMPONENT_SIZE(fmt))
+   /* This is the size of the sample data for one sample.  If the image is
+    * color-mapped it is the size of one color-map entry (and image pixels are
+    * one byte in size), otherwise it is the size of one image pixel.
+    */
+
+#define PNG_IMAGE_MAXIMUM_COLORMAP_COMPONENTS(fmt)\
+   (PNG_IMAGE_SAMPLE_CHANNELS(fmt) * 256)
+   /* The maximum size of the color-map required by the format expressed in a
+    * count of components.  This can be used to compile-time allocate a
+    * color-map:
+    *
+    * png_uint_16 colormap[PNG_IMAGE_MAXIMUM_COLORMAP_COMPONENTS(linear_fmt)];
+    *
+    * png_byte colormap[PNG_IMAGE_MAXIMUM_COLORMAP_COMPONENTS(sRGB_fmt)];
+    *
+    * Alternatively use the PNG_IMAGE_COLORMAP_SIZE macro below to use the
+    * information from one of the png_image_begin_read_ APIs and dynamically
+    * allocate the required memory.
+    */
+
+/* Corresponding information about the pixels */
+#define PNG_IMAGE_PIXEL_(test,fmt)\
+   (((fmt)&PNG_FORMAT_FLAG_COLORMAP)?1:test(fmt))
+
+#define PNG_IMAGE_PIXEL_CHANNELS(fmt)\
+   PNG_IMAGE_PIXEL_(PNG_IMAGE_SAMPLE_CHANNELS,fmt)
+   /* The number of separate channels (components) in a pixel; 1 for a
+    * color-mapped image.
+    */
+
+#define PNG_IMAGE_PIXEL_COMPONENT_SIZE(fmt)\
+   PNG_IMAGE_PIXEL_(PNG_IMAGE_SAMPLE_COMPONENT_SIZE,fmt)
+   /* The size, in bytes, of each component in a pixel; 1 for a color-mapped
+    * image.
+    */
+
+#define PNG_IMAGE_PIXEL_SIZE(fmt) PNG_IMAGE_PIXEL_(PNG_IMAGE_SAMPLE_SIZE,fmt)
+   /* The size, in bytes, of a complete pixel; 1 for a color-mapped image. */
+
+/* Information about the whole row, or whole image */
+#define PNG_IMAGE_ROW_STRIDE(image)\
+   (PNG_IMAGE_PIXEL_CHANNELS((image).format) * (image).width)
+   /* Return the total number of components in a single row of the image; this
+    * is the minimum 'row stride', the minimum count of components between each
+    * row.  For a color-mapped image this is the minimum number of bytes in a
+    * row.
+    */
+
+#define PNG_IMAGE_BUFFER_SIZE(image, row_stride)\
+   (PNG_IMAGE_PIXEL_COMPONENT_SIZE((image).format)*(image).height*(row_stride))
+   /* Return the size, in bytes, of an image buffer given a png_image and a row
+    * stride - the number of components to leave space for in each row.
+    */
+
+#define PNG_IMAGE_SIZE(image)\
+   PNG_IMAGE_BUFFER_SIZE(image, PNG_IMAGE_ROW_STRIDE(image))
+   /* Return the size, in bytes, of the image in memory given just a png_image;
+    * the row stride is the minimum stride required for the image.
+    */
+
+#define PNG_IMAGE_COLORMAP_SIZE(image)\
+   (PNG_IMAGE_SAMPLE_SIZE((image).format) * (image).colormap_entries)
+   /* Return the size, in bytes, of the color-map of this image.  If the image
+    * format is not a color-map format this will return a size sufficient for
+    * 256 entries in the given format; check PNG_FORMAT_FLAG_COLORMAP if
+    * you don't want to allocate a color-map in this case.
+    */
+
+/* PNG_IMAGE_FLAG_*
+ *
+ * Flags containing additional information about the image are held in the
+ * 'flags' field of png_image.
+ */
+#define PNG_IMAGE_FLAG_COLORSPACE_NOT_sRGB 0x01
+   /* This indicates the the RGB values of the in-memory bitmap do not
+    * correspond to the red, green and blue end-points defined by sRGB.
+    */
+
+#define PNG_IMAGE_FLAG_FAST 0x02
+   /* On write emphasise speed over compression; the resultant PNG file will be
+    * larger but will be produced significantly faster, particular for large
+    * images.  Do not use this option for images which will be distributed, only
+    * used it when producing intermediate files that will be read back in
+    * repeatedly.  For a typical 24-bit image the option will double the read
+    * speed at the cost of increasing the image size by 25%, however for many
+    * more compressible images the PNG file can be 10 times larger with only a
+    * slight speed gain.
+    */
+
+#define PNG_IMAGE_FLAG_16BIT_sRGB 0x04
+   /* On read if the image is a 16-bit per component image and there is no gAMA
+    * or sRGB chunk assume that the components are sRGB encoded.  Notice that
+    * images output by the simplified API always have gamma information; setting
+    * this flag only affects the interpretation of 16-bit images from an
+    * external source.  It is recommended that the application expose this flag
+    * to the user; the user can normally easily recognize the difference between
+    * linear and sRGB encoding.  This flag has no effect on write - the data
+    * passed to the write APIs must have the correct encoding (as defined
+    * above.)
+    *
+    * If the flag is not set (the default) input 16-bit per component data is
+    * assumed to be linear.
+    *
+    * NOTE: the flag can only be set after the png_image_begin_read_ call,
+    * because that call initializes the 'flags' field.
+    */
+
+#ifdef PNG_SIMPLIFIED_READ_SUPPORTED
+/* READ APIs
+ * ---------
+ *
+ * The png_image passed to the read APIs must have been initialized by setting
+ * the png_controlp field 'opaque' to NULL (or, safer, memset the whole thing.)
+ */
+#ifdef PNG_STDIO_SUPPORTED
+PNG_EXPORT(234, int, png_image_begin_read_from_file, (png_imagep image,
+   const char *file_name));
+   /* The named file is opened for read and the image header is filled in
+    * from the PNG header in the file.
+    */
+
+PNG_EXPORT(235, int, png_image_begin_read_from_stdio, (png_imagep image,
+   FILE* file));
+   /* The PNG header is read from the stdio FILE object. */
+#endif /* PNG_STDIO_SUPPORTED */
+
+PNG_EXPORT(236, int, png_image_begin_read_from_memory, (png_imagep image,
+   png_const_voidp memory, png_size_t size));
+   /* The PNG header is read from the given memory buffer. */
+
+PNG_EXPORT(237, int, png_image_finish_read, (png_imagep image,
+   png_const_colorp background, void *buffer, png_int_32 row_stride,
+   void *colormap));
+   /* Finish reading the image into the supplied buffer and clean up the
+    * png_image structure.
+    *
+    * row_stride is the step, in byte or 2-byte units as appropriate,
+    * between adjacent rows.  A positive stride indicates that the top-most row
+    * is first in the buffer - the normal top-down arrangement.  A negative
+    * stride indicates that the bottom-most row is first in the buffer.
+    *
+    * background need only be supplied if an alpha channel must be removed from
+    * a png_byte format and the removal is to be done by compositing on a solid
+    * color; otherwise it may be NULL and any composition will be done directly
+    * onto the buffer.  The value is an sRGB color to use for the background,
+    * for grayscale output the green channel is used.
+    *
+    * background must be supplied when an alpha channel must be removed from a
+    * single byte color-mapped output format, in other words if:
+    *
+    * 1) The original format from png_image_begin_read_from_* had
+    *    PNG_FORMAT_FLAG_ALPHA set.
+    * 2) The format set by the application does not.
+    * 3) The format set by the application has PNG_FORMAT_FLAG_COLORMAP set and
+    *    PNG_FORMAT_FLAG_LINEAR *not* set.
+    *
+    * For linear output removing the alpha channel is always done by compositing
+    * on black and background is ignored.
+    *
+    * colormap must be supplied when PNG_FORMAT_FLAG_COLORMAP is set.  It must
+    * be at least the size (in bytes) returned by PNG_IMAGE_COLORMAP_SIZE.
+    * image->colormap_entries will be updated to the actual number of entries
+    * written to the colormap; this may be less than the original value.
+    */
+
+PNG_EXPORT(238, void, png_image_free, (png_imagep image));
+   /* Free any data allocated by libpng in image->opaque, setting the pointer to
+    * NULL.  May be called at any time after the structure is initialized.
+    */
+#endif /* PNG_SIMPLIFIED_READ_SUPPORTED */
+
+#ifdef PNG_SIMPLIFIED_WRITE_SUPPORTED
+#ifdef PNG_STDIO_SUPPORTED
+/* WRITE APIS
+ * ----------
+ * For write you must initialize a png_image structure to describe the image to
+ * be written.  To do this use memset to set the whole structure to 0 then
+ * initialize fields describing your image.
+ *
+ * version: must be set to PNG_IMAGE_VERSION
+ * opaque: must be initialized to NULL
+ * width: image width in pixels
+ * height: image height in rows
+ * format: the format of the data (image and color-map) you wish to write
+ * flags: set to 0 unless one of the defined flags applies; set
+ *    PNG_IMAGE_FLAG_COLORSPACE_NOT_sRGB for color format images where the RGB
+ *    values do not correspond to the colors in sRGB.
+ * colormap_entries: set to the number of entries in the color-map (0 to 256)
+ */
+PNG_EXPORT(239, int, png_image_write_to_file, (png_imagep image,
+   const char *file, int convert_to_8bit, const void *buffer,
+   png_int_32 row_stride, const void *colormap));
+   /* Write the image to the named file. */
+
+PNG_EXPORT(240, int, png_image_write_to_stdio, (png_imagep image, FILE *file,
+   int convert_to_8_bit, const void *buffer, png_int_32 row_stride,
+   const void *colormap));
+   /* Write the image to the given (FILE*). */
+
+/* With both write APIs if image is in one of the linear formats with 16-bit
+ * data then setting convert_to_8_bit will cause the output to be an 8-bit PNG
+ * gamma encoded according to the sRGB specification, otherwise a 16-bit linear
+ * encoded PNG file is written.
+ *
+ * With color-mapped data formats the colormap parameter point to a color-map
+ * with at least image->colormap_entries encoded in the specified format.  If
+ * the format is linear the written PNG color-map will be converted to sRGB
+ * regardless of the convert_to_8_bit flag.
+ *
+ * With all APIs row_stride is handled as in the read APIs - it is the spacing
+ * from one row to the next in component sized units (1 or 2 bytes) and if
+ * negative indicates a bottom-up row layout in the buffer.
+ *
+ * Note that the write API does not support interlacing or sub-8-bit pixels.
+ */
+#endif /* PNG_STDIO_SUPPORTED */
+#endif /* PNG_SIMPLIFIED_WRITE_SUPPORTED */
+/*******************************************************************************
+ *  END OF SIMPLIFIED API
+ ******************************************************************************/
+
+#ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+PNG_EXPORT(242, void, png_set_check_for_invalid_index,
+    (png_structrp png_ptr, int allowed));
+#  ifdef PNG_GET_PALETTE_MAX_SUPPORTED
+PNG_EXPORT(243, int, png_get_palette_max, (png_const_structp png_ptr,
+    png_const_infop info_ptr));
+#  endif
+#endif /* CHECK_FOR_INVALID_INDEX */
+
+/*******************************************************************************
+ *  IMPLEMENTATION OPTIONS
+ *******************************************************************************
+ *
+ * Support for arbitrary implementation-specific optimizations.  The API allows
+ * particular options to be turned on or off.  'Option' is the number of the
+ * option and 'onoff' is 0 (off) or non-0 (on).  The value returned is given
+ * by the PNG_OPTION_ defines below.
+ *
+ * HARDWARE: normally hardware capabilites, such as the Intel SSE instructions,
+ *           are detected at run time, however sometimes it may be impossible
+ *           to do this in user mode, in which case it is necessary to discover
+ *           the capabilities in an OS specific way.  Such capabilities are
+ *           listed here when libpng has support for them and must be turned
+ *           ON by the application if present.
+ *
+ * SOFTWARE: sometimes software optimizations actually result in performance
+ *           decrease on some architectures or systems, or with some sets of
+ *           PNG images.  'Software' options allow such optimizations to be
+ *           selected at run time.
+ */
+#ifdef PNG_SET_OPTION_SUPPORTED
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+#  define PNG_ARM_NEON   0 /* HARDWARE: ARM Neon SIMD instructions supported */
+#endif
+#define PNG_MAXIMUM_INFLATE_WINDOW 2 /* SOFTWARE: force maximum window */
+#define PNG_OPTION_NEXT  4 /* Next option - numbers must be even */
+
+/* Return values: NOTE: there are four values and 'off' is *not* zero */
+#define PNG_OPTION_UNSET   0 /* Unset - defaults to off */
+#define PNG_OPTION_INVALID 1 /* Option number out of range */
+#define PNG_OPTION_OFF     2
+#define PNG_OPTION_ON      3
+
+PNG_EXPORT(244, int, png_set_option, (png_structrp png_ptr, int option,
+   int onoff));
+#endif
+
+/*******************************************************************************
+ *  END OF HARDWARE OPTIONS
+ ******************************************************************************/
+
+/* Maintainer: Put new public prototypes here ^, in libpng.3, and project
+ * defs, scripts/pnglibconf.h, and scripts/pnglibconf.h.prebuilt
+ */
+
+/* The last ordinal number (this is the *last* one already used; the next
+ * one to use is one more than this.)  Maintainer, remember to add an entry to
+ * scripts/symbols.def as well.
+ */
+#ifdef PNG_EXPORT_LAST_ORDINAL
+  PNG_EXPORT_LAST_ORDINAL(244);
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* PNG_VERSION_INFO_ONLY */
+/* Do not put anything past this line */
+#endif /* PNG_H */
diff --git a/ml/dlib/dlib/external/libpng/pngconf.h b/ml/dlib/dlib/external/libpng/pngconf.h
new file mode 100644
index 000000000..8c5347224
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngconf.h
@@ -0,0 +1,626 @@
+
+/* pngconf.h - machine configurable file for libpng
+ *
+ * libpng version 1.6.7 - November 14, 2013
+ *
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ */
+
+/* Any machine specific code is near the front of this file, so if you
+ * are configuring libpng for a machine, you may want to read the section
+ * starting here down to where it starts to typedef png_color, png_text,
+ * and png_info.
+ */
+
+#ifdef _MSC_VER
+// Disable the following warnings for Visual Studio
+// This is a warning you get from visual studio 2005 about things in the standard C++
+// library being "deprecated."  I checked the C++ standard and it doesn't say jack 
+// about any of them (I checked the searchable PDF).   So this warning is total Bunk.
+#pragma warning(disable : 4996)
+#endif
+
+
+#ifndef PNGCONF_H
+#define PNGCONF_H
+
+/* To do: Do all of this in scripts/pnglibconf.dfa */
+#ifdef PNG_SAFE_LIMITS_SUPPORTED
+#  ifdef PNG_USER_WIDTH_MAX
+#    undef PNG_USER_WIDTH_MAX
+#    define PNG_USER_WIDTH_MAX 1000000L
+#  endif
+#  ifdef PNG_USER_HEIGHT_MAX
+#    undef PNG_USER_HEIGHT_MAX
+#    define PNG_USER_HEIGHT_MAX 1000000L
+#  endif
+#  ifdef PNG_USER_CHUNK_MALLOC_MAX
+#    undef PNG_USER_CHUNK_MALLOC_MAX
+#    define PNG_USER_CHUNK_MALLOC_MAX 4000000L
+#  endif
+#  ifdef PNG_USER_CHUNK_CACHE_MAX
+#    undef PNG_USER_CHUNK_CACHE_MAX
+#    define PNG_USER_CHUNK_CACHE_MAX 128
+#  endif
+#endif
+
+#ifndef PNG_BUILDING_SYMBOL_TABLE /* else includes may cause problems */
+
+/* From libpng 1.6.0 libpng requires an ANSI X3.159-1989 ("ISOC90") compliant C
+ * compiler for correct compilation.  The following header files are required by
+ * the standard.  If your compiler doesn't provide these header files, or they
+ * do not match the standard, you will need to provide/improve them.
+ */
+#include <limits.h>
+#include <stddef.h>
+
+/* Library header files.  These header files are all defined by ISOC90; libpng
+ * expects conformant implementations, however, an ISOC90 conformant system need
+ * not provide these header files if the functionality cannot be implemented.
+ * In this case it will be necessary to disable the relevant parts of libpng in
+ * the build of pnglibconf.h.
+ *
+ * Prior to 1.6.0 string.h was included here; the API changes in 1.6.0 to not
+ * include this unnecessary header file.
+ */
+
+#ifdef PNG_STDIO_SUPPORTED
+   /* Required for the definition of FILE: */
+#  include <stdio.h>
+#endif
+
+#ifdef PNG_SETJMP_SUPPORTED
+   /* Required for the definition of jmp_buf and the declaration of longjmp: */
+#  include <setjmp.h>
+#endif
+
+#ifdef PNG_CONVERT_tIME_SUPPORTED
+   /* Required for struct tm: */
+#  include <time.h>
+#endif
+
+#endif /* PNG_BUILDING_SYMBOL_TABLE */
+
+/* Prior to 1.6.0 it was possible to turn off 'const' in declarations using
+ * PNG_NO_CONST; this is no longer supported except for data declarations which
+ * apparently still cause problems in 2011 on some compilers.
+ */
+#define PNG_CONST const /* backward compatibility only */
+
+/* This controls optimization of the reading of 16 and 32 bit values
+ * from PNG files.  It can be set on a per-app-file basis - it
+ * just changes whether a macro is used when the function is called.
+ * The library builder sets the default; if read functions are not
+ * built into the library the macro implementation is forced on.
+ */
+#ifndef PNG_READ_INT_FUNCTIONS_SUPPORTED
+#  define PNG_USE_READ_MACROS
+#endif
+#if !defined(PNG_NO_USE_READ_MACROS) && !defined(PNG_USE_READ_MACROS)
+#  if PNG_DEFAULT_READ_MACROS
+#    define PNG_USE_READ_MACROS
+#  endif
+#endif
+
+/* COMPILER SPECIFIC OPTIONS.
+ *
+ * These options are provided so that a variety of difficult compilers
+ * can be used.  Some are fixed at build time (e.g. PNG_API_RULE
+ * below) but still have compiler specific implementations, others
+ * may be changed on a per-file basis when compiling against libpng.
+ */
+
+/* The PNGARG macro was used in versions of libpng prior to 1.6.0 to protect
+ * against legacy (pre ISOC90) compilers that did not understand function
+ * prototypes.  It is not required for modern C compilers.
+ */
+#ifndef PNGARG
+#  define PNGARG(arglist) arglist
+#endif
+
+/* Function calling conventions.
+ * =============================
+ * Normally it is not necessary to specify to the compiler how to call
+ * a function - it just does it - however on x86 systems derived from
+ * Microsoft and Borland C compilers ('IBM PC', 'DOS', 'Windows' systems
+ * and some others) there are multiple ways to call a function and the
+ * default can be changed on the compiler command line.  For this reason
+ * libpng specifies the calling convention of every exported function and
+ * every function called via a user supplied function pointer.  This is
+ * done in this file by defining the following macros:
+ *
+ * PNGAPI    Calling convention for exported functions.
+ * PNGCBAPI  Calling convention for user provided (callback) functions.
+ * PNGCAPI   Calling convention used by the ANSI-C library (required
+ *           for longjmp callbacks and sometimes used internally to
+ *           specify the calling convention for zlib).
+ *
+ * These macros should never be overridden.  If it is necessary to
+ * change calling convention in a private build this can be done
+ * by setting PNG_API_RULE (which defaults to 0) to one of the values
+ * below to select the correct 'API' variants.
+ *
+ * PNG_API_RULE=0 Use PNGCAPI - the 'C' calling convention - throughout.
+ *                This is correct in every known environment.
+ * PNG_API_RULE=1 Use the operating system convention for PNGAPI and
+ *                the 'C' calling convention (from PNGCAPI) for
+ *                callbacks (PNGCBAPI).  This is no longer required
+ *                in any known environment - if it has to be used
+ *                please post an explanation of the problem to the
+ *                libpng mailing list.
+ *
+ * These cases only differ if the operating system does not use the C
+ * calling convention, at present this just means the above cases
+ * (x86 DOS/Windows sytems) and, even then, this does not apply to
+ * Cygwin running on those systems.
+ *
+ * Note that the value must be defined in pnglibconf.h so that what
+ * the application uses to call the library matches the conventions
+ * set when building the library.
+ */
+
+/* Symbol export
+ * =============
+ * When building a shared library it is almost always necessary to tell
+ * the compiler which symbols to export.  The png.h macro 'PNG_EXPORT'
+ * is used to mark the symbols.  On some systems these symbols can be
+ * extracted at link time and need no special processing by the compiler,
+ * on other systems the symbols are flagged by the compiler and just
+ * the declaration requires a special tag applied (unfortunately) in a
+ * compiler dependent way.  Some systems can do either.
+ *
+ * A small number of older systems also require a symbol from a DLL to
+ * be flagged to the program that calls it.  This is a problem because
+ * we do not know in the header file included by application code that
+ * the symbol will come from a shared library, as opposed to a statically
+ * linked one.  For this reason the application must tell us by setting
+ * the magic flag PNG_USE_DLL to turn on the special processing before
+ * it includes png.h.
+ *
+ * Four additional macros are used to make this happen:
+ *
+ * PNG_IMPEXP The magic (if any) to cause a symbol to be exported from
+ *            the build or imported if PNG_USE_DLL is set - compiler
+ *            and system specific.
+ *
+ * PNG_EXPORT_TYPE(type) A macro that pre or appends PNG_IMPEXP to
+ *                       'type', compiler specific.
+ *
+ * PNG_DLL_EXPORT Set to the magic to use during a libpng build to
+ *                make a symbol exported from the DLL.  Not used in the
+ *                public header files; see pngpriv.h for how it is used
+ *                in the libpng build.
+ *
+ * PNG_DLL_IMPORT Set to the magic to force the libpng symbols to come
+ *                from a DLL - used to define PNG_IMPEXP when
+ *                PNG_USE_DLL is set.
+ */
+
+/* System specific discovery.
+ * ==========================
+ * This code is used at build time to find PNG_IMPEXP, the API settings
+ * and PNG_EXPORT_TYPE(), it may also set a macro to indicate the DLL
+ * import processing is possible.  On Windows systems it also sets
+ * compiler-specific macros to the values required to change the calling
+ * conventions of the various functions.
+ */
+#if defined(_Windows) || defined(_WINDOWS) || defined(WIN32) ||\
+    defined(_WIN32) || defined(__WIN32__) || defined(__CYGWIN__)
+  /* Windows system (DOS doesn't support DLLs).  Includes builds under Cygwin or
+   * MinGW on any architecture currently supported by Windows.  Also includes
+   * Watcom builds but these need special treatment because they are not
+   * compatible with GCC or Visual C because of different calling conventions.
+   */
+#  if PNG_API_RULE == 2
+    /* If this line results in an error, either because __watcall is not
+     * understood or because of a redefine just below you cannot use *this*
+     * build of the library with the compiler you are using.  *This* build was
+     * build using Watcom and applications must also be built using Watcom!
+     */
+#    define PNGCAPI __watcall
+#  endif
+
+#  if defined(__GNUC__) || (defined(_MSC_VER) && (_MSC_VER >= 800))
+#    define PNGCAPI __cdecl
+#    if PNG_API_RULE == 1
+       /* If this line results in an error __stdcall is not understood and
+        * PNG_API_RULE should not have been set to '1'.
+        */
+#      define PNGAPI __stdcall
+#    endif
+#  else
+    /* An older compiler, or one not detected (erroneously) above,
+     * if necessary override on the command line to get the correct
+     * variants for the compiler.
+     */
+#    ifndef PNGCAPI
+#      define PNGCAPI _cdecl
+#    endif
+#    if PNG_API_RULE == 1 && !defined(PNGAPI)
+#      define PNGAPI _stdcall
+#    endif
+#  endif /* compiler/api */
+
+  /* NOTE: PNGCBAPI always defaults to PNGCAPI. */
+
+#  if defined(PNGAPI) && !defined(PNG_USER_PRIVATEBUILD)
+#     error "PNG_USER_PRIVATEBUILD must be defined if PNGAPI is changed"
+#  endif
+
+#  if (defined(_MSC_VER) && _MSC_VER < 800) ||\
+      (defined(__BORLANDC__) && __BORLANDC__ < 0x500)
+    /* older Borland and MSC
+     * compilers used '__export' and required this to be after
+     * the type.
+     */
+#    ifndef PNG_EXPORT_TYPE
+#      define PNG_EXPORT_TYPE(type) type PNG_IMPEXP
+#    endif
+#    define PNG_DLL_EXPORT __export
+#  else /* newer compiler */
+#    define PNG_DLL_EXPORT __declspec(dllexport)
+#    ifndef PNG_DLL_IMPORT
+#      define PNG_DLL_IMPORT __declspec(dllimport)
+#    endif
+#  endif /* compiler */
+
+#else /* !Windows */
+#  if (defined(__IBMC__) || defined(__IBMCPP__)) && defined(__OS2__)
+#    define PNGAPI _System
+#  else /* !Windows/x86 && !OS/2 */
+    /* Use the defaults, or define PNG*API on the command line (but
+     * this will have to be done for every compile!)
+     */
+#  endif /* other system, !OS/2 */
+#endif /* !Windows/x86 */
+
+/* Now do all the defaulting . */
+#ifndef PNGCAPI
+#  define PNGCAPI
+#endif
+#ifndef PNGCBAPI
+#  define PNGCBAPI PNGCAPI
+#endif
+#ifndef PNGAPI
+#  define PNGAPI PNGCAPI
+#endif
+
+/* PNG_IMPEXP may be set on the compilation system command line or (if not set)
+ * then in an internal header file when building the library, otherwise (when
+ * using the library) it is set here.
+ */
+#ifndef PNG_IMPEXP
+#  if defined(PNG_USE_DLL) && defined(PNG_DLL_IMPORT)
+     /* This forces use of a DLL, disallowing static linking */
+#    define PNG_IMPEXP PNG_DLL_IMPORT
+#  endif
+
+#  ifndef PNG_IMPEXP
+#    define PNG_IMPEXP
+#  endif
+#endif
+
+/* In 1.5.2 the definition of PNG_FUNCTION has been changed to always treat
+ * 'attributes' as a storage class - the attributes go at the start of the
+ * function definition, and attributes are always appended regardless of the
+ * compiler.  This considerably simplifies these macros but may cause problems
+ * if any compilers both need function attributes and fail to handle them as
+ * a storage class (this is unlikely.)
+ */
+#ifndef PNG_FUNCTION
+#  define PNG_FUNCTION(type, name, args, attributes) attributes type name args
+#endif
+
+#ifndef PNG_EXPORT_TYPE
+#  define PNG_EXPORT_TYPE(type) PNG_IMPEXP type
+#endif
+
+   /* The ordinal value is only relevant when preprocessing png.h for symbol
+    * table entries, so we discard it here.  See the .dfn files in the
+    * scripts directory.
+    */
+#ifndef PNG_EXPORTA
+
+#  define PNG_EXPORTA(ordinal, type, name, args, attributes)\
+      PNG_FUNCTION(PNG_EXPORT_TYPE(type),(PNGAPI name),PNGARG(args), \
+        extern attributes)
+#endif
+
+/* ANSI-C (C90) does not permit a macro to be invoked with an empty argument,
+ * so make something non-empty to satisfy the requirement:
+ */
+#define PNG_EMPTY /*empty list*/
+
+#define PNG_EXPORT(ordinal, type, name, args)\
+   PNG_EXPORTA(ordinal, type, name, args, PNG_EMPTY)
+
+/* Use PNG_REMOVED to comment out a removed interface. */
+#ifndef PNG_REMOVED
+#  define PNG_REMOVED(ordinal, type, name, args, attributes)
+#endif
+
+#ifndef PNG_CALLBACK
+#  define PNG_CALLBACK(type, name, args) type (PNGCBAPI name) PNGARG(args)
+#endif
+
+/* Support for compiler specific function attributes.  These are used
+ * so that where compiler support is available incorrect use of API
+ * functions in png.h will generate compiler warnings.
+ *
+ * Added at libpng-1.2.41.
+ */
+
+#ifndef PNG_NO_PEDANTIC_WARNINGS
+#  ifndef PNG_PEDANTIC_WARNINGS_SUPPORTED
+#    define PNG_PEDANTIC_WARNINGS_SUPPORTED
+#  endif
+#endif
+
+#ifdef PNG_PEDANTIC_WARNINGS_SUPPORTED
+  /* Support for compiler specific function attributes.  These are used
+   * so that where compiler support is available, incorrect use of API
+   * functions in png.h will generate compiler warnings.  Added at libpng
+   * version 1.2.41.  Disabling these removes the warnings but may also produce
+   * less efficient code.
+   */
+#  if defined(__GNUC__)
+#    ifndef PNG_USE_RESULT
+#      define PNG_USE_RESULT __attribute__((__warn_unused_result__))
+#    endif
+#    ifndef PNG_NORETURN
+#      define PNG_NORETURN   __attribute__((__noreturn__))
+#    endif
+#    if __GNUC__ >= 3
+#      ifndef PNG_ALLOCATED
+#        define PNG_ALLOCATED  __attribute__((__malloc__))
+#      endif
+#      ifndef PNG_DEPRECATED
+#        define PNG_DEPRECATED __attribute__((__deprecated__))
+#      endif
+#      ifndef PNG_PRIVATE
+#        if 0 /* Doesn't work so we use deprecated instead*/
+#          define PNG_PRIVATE \
+            __attribute__((warning("This function is not exported by libpng.")))
+#        else
+#          define PNG_PRIVATE \
+            __attribute__((__deprecated__))
+#        endif
+#      endif
+#      if ((__GNUC__ != 3) || !defined(__GNUC_MINOR__) || (__GNUC_MINOR__ >= 1))
+#        ifndef PNG_RESTRICT
+#          define PNG_RESTRICT __restrict
+#        endif
+#      endif /*  __GNUC__ == 3.0 */
+#    endif /*  __GNUC__ >= 3 */
+
+#  elif defined(_MSC_VER)  && (_MSC_VER >= 1300)
+#    ifndef PNG_USE_RESULT
+#      define PNG_USE_RESULT /* not supported */
+#    endif
+#    ifndef PNG_NORETURN
+#      define PNG_NORETURN   __declspec(noreturn)
+#    endif
+#    ifndef PNG_ALLOCATED
+#      if (_MSC_VER >= 1400)
+#        define PNG_ALLOCATED __declspec(restrict)
+#      endif
+#    endif
+#    ifndef PNG_DEPRECATED
+#      define PNG_DEPRECATED __declspec(deprecated)
+#    endif
+#    ifndef PNG_PRIVATE
+#      define PNG_PRIVATE __declspec(deprecated)
+#    endif
+#    ifndef PNG_RESTRICT
+#      if (_MSC_VER >= 1400)
+#        define PNG_RESTRICT __restrict
+#      endif
+#    endif
+
+#  elif defined(__WATCOMC__)
+#    ifndef PNG_RESTRICT
+#      define PNG_RESTRICT __restrict
+#    endif
+#  endif /* _MSC_VER */
+#endif /* PNG_PEDANTIC_WARNINGS */
+
+#ifndef PNG_DEPRECATED
+#  define PNG_DEPRECATED  /* Use of this function is deprecated */
+#endif
+#ifndef PNG_USE_RESULT
+#  define PNG_USE_RESULT  /* The result of this function must be checked */
+#endif
+#ifndef PNG_NORETURN
+#  define PNG_NORETURN    /* This function does not return */
+#endif
+#ifndef PNG_ALLOCATED
+#  define PNG_ALLOCATED   /* The result of the function is new memory */
+#endif
+#ifndef PNG_PRIVATE
+#  define PNG_PRIVATE     /* This is a private libpng function */
+#endif
+#ifndef PNG_RESTRICT
+#  define PNG_RESTRICT    /* The C99 "restrict" feature */
+#endif
+#ifndef PNG_FP_EXPORT     /* A floating point API. */
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+#     define PNG_FP_EXPORT(ordinal, type, name, args)\
+         PNG_EXPORT(ordinal, type, name, args);
+#  else                   /* No floating point APIs */
+#     define PNG_FP_EXPORT(ordinal, type, name, args)
+#  endif
+#endif
+#ifndef PNG_FIXED_EXPORT  /* A fixed point API. */
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+#     define PNG_FIXED_EXPORT(ordinal, type, name, args)\
+         PNG_EXPORT(ordinal, type, name, args);
+#  else                   /* No fixed point APIs */
+#     define PNG_FIXED_EXPORT(ordinal, type, name, args)
+#  endif
+#endif
+
+#ifndef PNG_BUILDING_SYMBOL_TABLE
+/* Some typedefs to get us started.  These should be safe on most of the common
+ * platforms.
+ *
+ * png_uint_32 and png_int_32 may, currently, be larger than required to hold a
+ * 32-bit value however this is not normally advisable.
+ *
+ * png_uint_16 and png_int_16 should always be two bytes in size - this is
+ * verified at library build time.
+ *
+ * png_byte must always be one byte in size.
+ *
+ * The checks below use constants from limits.h, as defined by the ISOC90
+ * standard.
+ */
+#if CHAR_BIT == 8 && UCHAR_MAX == 255
+   typedef unsigned char png_byte;
+#else
+#  error "libpng requires 8 bit bytes"
+#endif
+
+#if INT_MIN == -32768 && INT_MAX == 32767
+   typedef int png_int_16;
+#elif SHRT_MIN == -32768 && SHRT_MAX == 32767
+   typedef short png_int_16;
+#else
+#  error "libpng requires a signed 16 bit type"
+#endif
+
+#if UINT_MAX == 65535
+   typedef unsigned int png_uint_16;
+#elif USHRT_MAX == 65535
+   typedef unsigned short png_uint_16;
+#else
+#  error "libpng requires an unsigned 16 bit type"
+#endif
+
+#if INT_MIN < -2147483646 && INT_MAX > 2147483646
+   typedef int png_int_32;
+#elif LONG_MIN < -2147483646 && LONG_MAX > 2147483646
+   typedef long int png_int_32;
+#else
+#  error "libpng requires a signed 32 bit (or more) type"
+#endif
+
+#if UINT_MAX > 4294967294
+   typedef unsigned int png_uint_32;
+#elif ULONG_MAX > 4294967294
+   typedef unsigned long int png_uint_32;
+#else
+#  error "libpng requires an unsigned 32 bit (or more) type"
+#endif
+
+/* Prior to 1.6.0 it was possible to disable the use of size_t, 1.6.0, however,
+ * requires an ISOC90 compiler and relies on consistent behavior of sizeof.
+ */
+typedef size_t png_size_t;
+typedef ptrdiff_t png_ptrdiff_t;
+
+/* libpng needs to know the maximum value of 'size_t' and this controls the
+ * definition of png_alloc_size_t, below.  This maximum value of size_t limits
+ * but does not control the maximum allocations the library makes - there is
+ * direct application control of this through png_set_user_limits().
+ */
+#ifndef PNG_SMALL_SIZE_T
+   /* Compiler specific tests for systems where size_t is known to be less than
+    * 32 bits (some of these systems may no longer work because of the lack of
+    * 'far' support; see above.)
+    */
+#  if (defined(__TURBOC__) && !defined(__FLAT__)) ||\
+   (defined(_MSC_VER) && defined(MAXSEG_64K))
+#     define PNG_SMALL_SIZE_T
+#  endif
+#endif
+
+/* png_alloc_size_t is guaranteed to be no smaller than png_size_t, and no
+ * smaller than png_uint_32.  Casts from png_size_t or png_uint_32 to
+ * png_alloc_size_t are not necessary; in fact, it is recommended not to use
+ * them at all so that the compiler can complain when something turns out to be
+ * problematic.
+ *
+ * Casts in the other direction (from png_alloc_size_t to png_size_t or
+ * png_uint_32) should be explicitly applied; however, we do not expect to
+ * encounter practical situations that require such conversions.
+ *
+ * PNG_SMALL_SIZE_T must be defined if the maximum value of size_t is less than
+ * 4294967295 - i.e. less than the maximum value of png_uint_32.
+ */
+#ifdef PNG_SMALL_SIZE_T
+   typedef png_uint_32 png_alloc_size_t;
+#else
+   typedef png_size_t png_alloc_size_t;
+#endif
+
+/* Prior to 1.6.0 libpng offered limited support for Microsoft C compiler
+ * implementations of Intel CPU specific support of user-mode segmented address
+ * spaces, where 16-bit pointers address more than 65536 bytes of memory using
+ * separate 'segment' registers.  The implementation requires two different
+ * types of pointer (only one of which includes the segment value.)
+ *
+ * If required this support is available in version 1.2 of libpng and may be
+ * available in versions through 1.5, although the correctness of the code has
+ * not been verified recently.
+ */
+
+/* Typedef for floating-point numbers that are converted to fixed-point with a
+ * multiple of 100,000, e.g., gamma
+ */
+typedef png_int_32 png_fixed_point;
+
+/* Add typedefs for pointers */
+typedef void                  * png_voidp;
+typedef const void            * png_const_voidp;
+typedef png_byte              * png_bytep;
+typedef const png_byte        * png_const_bytep;
+typedef png_uint_32           * png_uint_32p;
+typedef const png_uint_32     * png_const_uint_32p;
+typedef png_int_32            * png_int_32p;
+typedef const png_int_32      * png_const_int_32p;
+typedef png_uint_16           * png_uint_16p;
+typedef const png_uint_16     * png_const_uint_16p;
+typedef png_int_16            * png_int_16p;
+typedef const png_int_16      * png_const_int_16p;
+typedef char                  * png_charp;
+typedef const char            * png_const_charp;
+typedef png_fixed_point       * png_fixed_point_p;
+typedef const png_fixed_point * png_const_fixed_point_p;
+typedef png_size_t            * png_size_tp;
+typedef const png_size_t      * png_const_size_tp;
+
+#ifdef PNG_STDIO_SUPPORTED
+typedef FILE            * png_FILE_p;
+#endif
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+typedef double       * png_doublep;
+typedef const double * png_const_doublep;
+#endif
+
+/* Pointers to pointers; i.e. arrays */
+typedef png_byte        * * png_bytepp;
+typedef png_uint_32     * * png_uint_32pp;
+typedef png_int_32      * * png_int_32pp;
+typedef png_uint_16     * * png_uint_16pp;
+typedef png_int_16      * * png_int_16pp;
+typedef const char      * * png_const_charpp;
+typedef char            * * png_charpp;
+typedef png_fixed_point * * png_fixed_point_pp;
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+typedef double          * * png_doublepp;
+#endif
+
+/* Pointers to pointers to pointers; i.e., pointer to array */
+typedef char            * * * png_charppp;
+
+#endif /* PNG_BUILDING_SYMBOL_TABLE */
+
+#endif /* PNGCONF_H */
diff --git a/ml/dlib/dlib/external/libpng/pngdebug.h b/ml/dlib/dlib/external/libpng/pngdebug.h
new file mode 100644
index 000000000..16f81fdd1
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngdebug.h
@@ -0,0 +1,157 @@
+
+/* pngdebug.h - Debugging macros for libpng, also used in pngtest.c
+ *
+ * Copyright (c) 1998-2011 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * Last changed in libpng 1.5.0 [January 6, 2011]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* Define PNG_DEBUG at compile time for debugging information.  Higher
+ * numbers for PNG_DEBUG mean more debugging information.  This has
+ * only been added since version 0.95 so it is not implemented throughout
+ * libpng yet, but more support will be added as needed.
+ *
+ * png_debug[1-2]?(level, message ,arg{0-2})
+ *   Expands to a statement (either a simple expression or a compound
+ *   do..while(0) statement) that outputs a message with parameter
+ *   substitution if PNG_DEBUG is defined to 2 or more.  If PNG_DEBUG
+ *   is undefined, 0 or 1 every png_debug expands to a simple expression
+ *   (actually ((void)0)).
+ *
+ *   level: level of detail of message, starting at 0.  A level 'n'
+ *          message is preceded by 'n' tab characters (not implemented
+ *          on Microsoft compilers unless PNG_DEBUG_FILE is also
+ *          defined, to allow debug DLL compilation with no standard IO).
+ *   message: a printf(3) style text string.  A trailing '\n' is added
+ *            to the message.
+ *   arg: 0 to 2 arguments for printf(3) style substitution in message.
+ */
+#ifndef PNGDEBUG_H
+#define PNGDEBUG_H
+/* These settings control the formatting of messages in png.c and pngerror.c */
+/* Moved to pngdebug.h at 1.5.0 */
+#  ifndef PNG_LITERAL_SHARP
+#    define PNG_LITERAL_SHARP 0x23
+#  endif
+#  ifndef PNG_LITERAL_LEFT_SQUARE_BRACKET
+#    define PNG_LITERAL_LEFT_SQUARE_BRACKET 0x5b
+#  endif
+#  ifndef PNG_LITERAL_RIGHT_SQUARE_BRACKET
+#    define PNG_LITERAL_RIGHT_SQUARE_BRACKET 0x5d
+#  endif
+#  ifndef PNG_STRING_NEWLINE
+#    define PNG_STRING_NEWLINE "\n"
+#  endif
+
+#ifdef PNG_DEBUG
+#  if (PNG_DEBUG > 0)
+#    if !defined(PNG_DEBUG_FILE) && defined(_MSC_VER)
+#      include <crtdbg.h>
+#      if (PNG_DEBUG > 1)
+#        ifndef _DEBUG
+#          define _DEBUG
+#        endif
+#        ifndef png_debug
+#          define png_debug(l,m)  _RPT0(_CRT_WARN,m PNG_STRING_NEWLINE)
+#        endif
+#        ifndef png_debug1
+#          define png_debug1(l,m,p1)  _RPT1(_CRT_WARN,m PNG_STRING_NEWLINE,p1)
+#        endif
+#        ifndef png_debug2
+#          define png_debug2(l,m,p1,p2) \
+             _RPT2(_CRT_WARN,m PNG_STRING_NEWLINE,p1,p2)
+#        endif
+#      endif
+#    else /* PNG_DEBUG_FILE || !_MSC_VER */
+#      ifndef PNG_STDIO_SUPPORTED
+#        include <stdio.h> /* not included yet */
+#      endif
+#      ifndef PNG_DEBUG_FILE
+#        define PNG_DEBUG_FILE stderr
+#      endif /* PNG_DEBUG_FILE */
+
+#      if (PNG_DEBUG > 1)
+/* Note: ["%s"m PNG_STRING_NEWLINE] probably does not work on
+ * non-ISO compilers
+ */
+#        ifdef __STDC__
+#          ifndef png_debug
+#            define png_debug(l,m) \
+       do { \
+       int num_tabs=l; \
+       fprintf(PNG_DEBUG_FILE,"%s"m PNG_STRING_NEWLINE,(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":"")))); \
+       } while (0)
+#          endif
+#          ifndef png_debug1
+#            define png_debug1(l,m,p1) \
+       do { \
+       int num_tabs=l; \
+       fprintf(PNG_DEBUG_FILE,"%s"m PNG_STRING_NEWLINE,(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":""))),p1); \
+       } while (0)
+#          endif
+#          ifndef png_debug2
+#            define png_debug2(l,m,p1,p2) \
+       do { \
+       int num_tabs=l; \
+       fprintf(PNG_DEBUG_FILE,"%s"m PNG_STRING_NEWLINE,(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":""))),p1,p2); \
+       } while (0)
+#          endif
+#        else /* __STDC __ */
+#          ifndef png_debug
+#            define png_debug(l,m) \
+       do { \
+       int num_tabs=l; \
+       char format[256]; \
+       snprintf(format,256,"%s%s%s",(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":""))), \
+         m,PNG_STRING_NEWLINE); \
+       fprintf(PNG_DEBUG_FILE,format); \
+       } while (0)
+#          endif
+#          ifndef png_debug1
+#            define png_debug1(l,m,p1) \
+       do { \
+       int num_tabs=l; \
+       char format[256]; \
+       snprintf(format,256,"%s%s%s",(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":""))), \
+         m,PNG_STRING_NEWLINE); \
+       fprintf(PNG_DEBUG_FILE,format,p1); \
+       } while (0)
+#          endif
+#          ifndef png_debug2
+#            define png_debug2(l,m,p1,p2) \
+       do { \
+       int num_tabs=l; \
+       char format[256]; \
+       snprintf(format,256,"%s%s%s",(num_tabs==1 ? "\t" : \
+         (num_tabs==2 ? "\t\t":(num_tabs>2 ? "\t\t\t":""))), \
+         m,PNG_STRING_NEWLINE); \
+       fprintf(PNG_DEBUG_FILE,format,p1,p2); \
+       } while (0)
+#          endif
+#        endif /* __STDC __ */
+#      endif /* (PNG_DEBUG > 1) */
+
+#    endif /* _MSC_VER */
+#  endif /* (PNG_DEBUG > 0) */
+#endif /* PNG_DEBUG */
+#ifndef png_debug
+#  define png_debug(l, m) ((void)0)
+#endif
+#ifndef png_debug1
+#  define png_debug1(l, m, p1) ((void)0)
+#endif
+#ifndef png_debug2
+#  define png_debug2(l, m, p1, p2) ((void)0)
+#endif
+#endif /* PNGDEBUG_H */
diff --git a/ml/dlib/dlib/external/libpng/pngerror.c b/ml/dlib/dlib/external/libpng/pngerror.c
new file mode 100644
index 000000000..f469206ee
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngerror.c
@@ -0,0 +1,932 @@
+
+/* pngerror.c - stub functions for i/o and memory allocation
+ *
+ * Last changed in libpng 1.6.1 [March 28, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file provides a location for all error handling.  Users who
+ * need special error handling are expected to write replacement functions
+ * and use png_set_error_fn() to use those functions.  See the instructions
+ * at each function.
+ */
+
+#include "pngpriv.h"
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+
+static PNG_FUNCTION(void, png_default_error,PNGARG((png_const_structrp png_ptr,
+    png_const_charp error_message)),PNG_NORETURN);
+
+#ifdef PNG_WARNINGS_SUPPORTED
+static void /* PRIVATE */
+png_default_warning PNGARG((png_const_structrp png_ptr,
+   png_const_charp warning_message));
+#endif /* PNG_WARNINGS_SUPPORTED */
+
+/* This function is called whenever there is a fatal error.  This function
+ * should not be changed.  If there is a need to handle errors differently,
+ * you should supply a replacement error function and use png_set_error_fn()
+ * to replace the error function at run-time.
+ */
+#ifdef PNG_ERROR_TEXT_SUPPORTED
+PNG_FUNCTION(void,PNGAPI
+png_error,(png_const_structrp png_ptr, png_const_charp error_message),
+   PNG_NORETURN)
+{
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+   char msg[16];
+   if (png_ptr != NULL)
+   {
+      if (png_ptr->flags&
+         (PNG_FLAG_STRIP_ERROR_NUMBERS|PNG_FLAG_STRIP_ERROR_TEXT))
+      {
+         if (*error_message == PNG_LITERAL_SHARP)
+         {
+            /* Strip "#nnnn " from beginning of error message. */
+            int offset;
+            for (offset = 1; offset<15; offset++)
+               if (error_message[offset] == ' ')
+                  break;
+
+            if (png_ptr->flags&PNG_FLAG_STRIP_ERROR_TEXT)
+            {
+               int i;
+               for (i = 0; i < offset - 1; i++)
+                  msg[i] = error_message[i + 1];
+               msg[i - 1] = '\0';
+               error_message = msg;
+            }
+
+            else
+               error_message += offset;
+      }
+
+      else
+      {
+         if (png_ptr->flags&PNG_FLAG_STRIP_ERROR_TEXT)
+         {
+            msg[0] = '0';
+            msg[1] = '\0';
+            error_message = msg;
+         }
+       }
+     }
+   }
+#endif
+   if (png_ptr != NULL && png_ptr->error_fn != NULL)
+      (*(png_ptr->error_fn))(png_constcast(png_structrp,png_ptr),
+          error_message);
+
+   /* If the custom handler doesn't exist, or if it returns,
+      use the default handler, which will not return. */
+   png_default_error(png_ptr, error_message);
+}
+#else
+PNG_FUNCTION(void,PNGAPI
+png_err,(png_const_structrp png_ptr),PNG_NORETURN)
+{
+   /* Prior to 1.5.2 the error_fn received a NULL pointer, expressed
+    * erroneously as '\0', instead of the empty string "".  This was
+    * apparently an error, introduced in libpng-1.2.20, and png_default_error
+    * will crash in this case.
+    */
+   if (png_ptr != NULL && png_ptr->error_fn != NULL)
+      (*(png_ptr->error_fn))(png_constcast(png_structrp,png_ptr), "");
+
+   /* If the custom handler doesn't exist, or if it returns,
+      use the default handler, which will not return. */
+   png_default_error(png_ptr, "");
+}
+#endif /* PNG_ERROR_TEXT_SUPPORTED */
+
+/* Utility to safely appends strings to a buffer.  This never errors out so
+ * error checking is not required in the caller.
+ */
+size_t
+png_safecat(png_charp buffer, size_t bufsize, size_t pos,
+   png_const_charp string)
+{
+   if (buffer != NULL && pos < bufsize)
+   {
+      if (string != NULL)
+         while (*string != '\0' && pos < bufsize-1)
+           buffer[pos++] = *string++;
+
+      buffer[pos] = '\0';
+   }
+
+   return pos;
+}
+
+#if defined(PNG_WARNINGS_SUPPORTED) || defined(PNG_TIME_RFC1123_SUPPORTED)
+/* Utility to dump an unsigned value into a buffer, given a start pointer and
+ * and end pointer (which should point just *beyond* the end of the buffer!)
+ * Returns the pointer to the start of the formatted string.
+ */
+png_charp
+png_format_number(png_const_charp start, png_charp end, int format,
+   png_alloc_size_t number)
+{
+   int count = 0;    /* number of digits output */
+   int mincount = 1; /* minimum number required */
+   int output = 0;   /* digit output (for the fixed point format) */
+
+   *--end = '\0';
+
+   /* This is written so that the loop always runs at least once, even with
+    * number zero.
+    */
+   while (end > start && (number != 0 || count < mincount))
+   {
+
+      static const char digits[] = "0123456789ABCDEF";
+
+      switch (format)
+      {
+         case PNG_NUMBER_FORMAT_fixed:
+            /* Needs five digits (the fraction) */
+            mincount = 5;
+            if (output || number % 10 != 0)
+            {
+               *--end = digits[number % 10];
+               output = 1;
+            }
+            number /= 10;
+            break;
+
+         case PNG_NUMBER_FORMAT_02u:
+            /* Expects at least 2 digits. */
+            mincount = 2;
+            /* FALL THROUGH */
+
+         case PNG_NUMBER_FORMAT_u:
+            *--end = digits[number % 10];
+            number /= 10;
+            break;
+
+         case PNG_NUMBER_FORMAT_02x:
+            /* This format expects at least two digits */
+            mincount = 2;
+            /* FALL THROUGH */
+
+         case PNG_NUMBER_FORMAT_x:
+            *--end = digits[number & 0xf];
+            number >>= 4;
+            break;
+
+         default: /* an error */
+            number = 0;
+            break;
+      }
+
+      /* Keep track of the number of digits added */
+      ++count;
+
+      /* Float a fixed number here: */
+      if (format == PNG_NUMBER_FORMAT_fixed) if (count == 5) if (end > start)
+      {
+         /* End of the fraction, but maybe nothing was output?  In that case
+          * drop the decimal point.  If the number is a true zero handle that
+          * here.
+          */
+         if (output)
+            *--end = '.';
+         else if (number == 0) /* and !output */
+            *--end = '0';
+      }
+   }
+
+   return end;
+}
+#endif
+
+#ifdef PNG_WARNINGS_SUPPORTED
+/* This function is called whenever there is a non-fatal error.  This function
+ * should not be changed.  If there is a need to handle warnings differently,
+ * you should supply a replacement warning function and use
+ * png_set_error_fn() to replace the warning function at run-time.
+ */
+void PNGAPI
+png_warning(png_const_structrp png_ptr, png_const_charp warning_message)
+{
+   int offset = 0;
+   if (png_ptr != NULL)
+   {
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+   if (png_ptr->flags&
+       (PNG_FLAG_STRIP_ERROR_NUMBERS|PNG_FLAG_STRIP_ERROR_TEXT))
+#endif
+      {
+         if (*warning_message == PNG_LITERAL_SHARP)
+         {
+            for (offset = 1; offset < 15; offset++)
+               if (warning_message[offset] == ' ')
+                  break;
+         }
+      }
+   }
+   if (png_ptr != NULL && png_ptr->warning_fn != NULL)
+      (*(png_ptr->warning_fn))(png_constcast(png_structrp,png_ptr),
+         warning_message + offset);
+   else
+      png_default_warning(png_ptr, warning_message + offset);
+}
+
+/* These functions support 'formatted' warning messages with up to
+ * PNG_WARNING_PARAMETER_COUNT parameters.  In the format string the parameter
+ * is introduced by @<number>, where 'number' starts at 1.  This follows the
+ * standard established by X/Open for internationalizable error messages.
+ */
+void
+png_warning_parameter(png_warning_parameters p, int number,
+   png_const_charp string)
+{
+   if (number > 0 && number <= PNG_WARNING_PARAMETER_COUNT)
+      (void)png_safecat(p[number-1], (sizeof p[number-1]), 0, string);
+}
+
+void
+png_warning_parameter_unsigned(png_warning_parameters p, int number, int format,
+   png_alloc_size_t value)
+{
+   char buffer[PNG_NUMBER_BUFFER_SIZE];
+   png_warning_parameter(p, number, PNG_FORMAT_NUMBER(buffer, format, value));
+}
+
+void
+png_warning_parameter_signed(png_warning_parameters p, int number, int format,
+   png_int_32 value)
+{
+   png_alloc_size_t u;
+   png_charp str;
+   char buffer[PNG_NUMBER_BUFFER_SIZE];
+
+   /* Avoid overflow by doing the negate in a png_alloc_size_t: */
+   u = (png_alloc_size_t)value;
+   if (value < 0)
+      u = ~u + 1;
+
+   str = PNG_FORMAT_NUMBER(buffer, format, u);
+
+   if (value < 0 && str > buffer)
+      *--str = '-';
+
+   png_warning_parameter(p, number, str);
+}
+
+void
+png_formatted_warning(png_const_structrp png_ptr, png_warning_parameters p,
+   png_const_charp message)
+{
+   /* The internal buffer is just 192 bytes - enough for all our messages,
+    * overflow doesn't happen because this code checks!  If someone figures
+    * out how to send us a message longer than 192 bytes, all that will
+    * happen is that the message will be truncated appropriately.
+    */
+   size_t i = 0; /* Index in the msg[] buffer: */
+   char msg[192];
+
+   /* Each iteration through the following loop writes at most one character
+    * to msg[i++] then returns here to validate that there is still space for
+    * the trailing '\0'.  It may (in the case of a parameter) read more than
+    * one character from message[]; it must check for '\0' and continue to the
+    * test if it finds the end of string.
+    */
+   while (i<(sizeof msg)-1 && *message != '\0')
+   {
+      /* '@' at end of string is now just printed (previously it was skipped);
+       * it is an error in the calling code to terminate the string with @.
+       */
+      if (p != NULL && *message == '@' && message[1] != '\0')
+      {
+         int parameter_char = *++message; /* Consume the '@' */
+         static const char valid_parameters[] = "123456789";
+         int parameter = 0;
+
+         /* Search for the parameter digit, the index in the string is the
+          * parameter to use.
+          */
+         while (valid_parameters[parameter] != parameter_char &&
+            valid_parameters[parameter] != '\0')
+            ++parameter;
+
+         /* If the parameter digit is out of range it will just get printed. */
+         if (parameter < PNG_WARNING_PARAMETER_COUNT)
+         {
+            /* Append this parameter */
+            png_const_charp parm = p[parameter];
+            png_const_charp pend = p[parameter] + (sizeof p[parameter]);
+
+            /* No need to copy the trailing '\0' here, but there is no guarantee
+             * that parm[] has been initialized, so there is no guarantee of a
+             * trailing '\0':
+             */
+            while (i<(sizeof msg)-1 && *parm != '\0' && parm < pend)
+               msg[i++] = *parm++;
+
+            /* Consume the parameter digit too: */
+            ++message;
+            continue;
+         }
+
+         /* else not a parameter and there is a character after the @ sign; just
+          * copy that.  This is known not to be '\0' because of the test above.
+          */
+      }
+
+      /* At this point *message can't be '\0', even in the bad parameter case
+       * above where there is a lone '@' at the end of the message string.
+       */
+      msg[i++] = *message++;
+   }
+
+   /* i is always less than (sizeof msg), so: */
+   msg[i] = '\0';
+
+   /* And this is the formatted message. It may be larger than
+    * PNG_MAX_ERROR_TEXT, but that is only used for 'chunk' errors and these
+    * are not (currently) formatted.
+    */
+   png_warning(png_ptr, msg);
+}
+#endif /* PNG_WARNINGS_SUPPORTED */
+
+#ifdef PNG_BENIGN_ERRORS_SUPPORTED
+void PNGAPI
+png_benign_error(png_const_structrp png_ptr, png_const_charp error_message)
+{
+   if (png_ptr->flags & PNG_FLAG_BENIGN_ERRORS_WARN)
+   {
+#     ifdef PNG_READ_SUPPORTED
+         if ((png_ptr->mode & PNG_IS_READ_STRUCT) != 0 &&
+            png_ptr->chunk_name != 0)
+            png_chunk_warning(png_ptr, error_message);
+         else
+#     endif
+      png_warning(png_ptr, error_message);
+   }
+
+   else
+   {
+#     ifdef PNG_READ_SUPPORTED
+         if ((png_ptr->mode & PNG_IS_READ_STRUCT) != 0 &&
+            png_ptr->chunk_name != 0)
+            png_chunk_error(png_ptr, error_message);
+         else
+#     endif
+      png_error(png_ptr, error_message);
+   }
+}
+
+void /* PRIVATE */
+png_app_warning(png_const_structrp png_ptr, png_const_charp error_message)
+{
+  if (png_ptr->flags & PNG_FLAG_APP_WARNINGS_WARN)
+     png_warning(png_ptr, error_message);
+  else
+     png_error(png_ptr, error_message);
+}
+
+void /* PRIVATE */
+png_app_error(png_const_structrp png_ptr, png_const_charp error_message)
+{
+  if (png_ptr->flags & PNG_FLAG_APP_ERRORS_WARN)
+     png_warning(png_ptr, error_message);
+  else
+     png_error(png_ptr, error_message);
+}
+#endif /* BENIGN_ERRORS */
+
+/* These utilities are used internally to build an error message that relates
+ * to the current chunk.  The chunk name comes from png_ptr->chunk_name,
+ * this is used to prefix the message.  The message is limited in length
+ * to 63 bytes, the name characters are output as hex digits wrapped in []
+ * if the character is invalid.
+ */
+#define isnonalpha(c) ((c) < 65 || (c) > 122 || ((c) > 90 && (c) < 97))
+static PNG_CONST char png_digit[16] = {
+   '0', '1', '2', '3', '4', '5', '6', '7', '8', '9',
+   'A', 'B', 'C', 'D', 'E', 'F'
+};
+
+#define PNG_MAX_ERROR_TEXT 196 /* Currently limited be profile_error in png.c */
+#if defined(PNG_WARNINGS_SUPPORTED) || defined(PNG_ERROR_TEXT_SUPPORTED)
+static void /* PRIVATE */
+png_format_buffer(png_const_structrp png_ptr, png_charp buffer, png_const_charp
+    error_message)
+{
+   png_uint_32 chunk_name = png_ptr->chunk_name;
+   int iout = 0, ishift = 24;
+
+   while (ishift >= 0)
+   {
+      int c = (int)(chunk_name >> ishift) & 0xff;
+
+      ishift -= 8;
+      if (isnonalpha(c))
+      {
+         buffer[iout++] = PNG_LITERAL_LEFT_SQUARE_BRACKET;
+         buffer[iout++] = png_digit[(c & 0xf0) >> 4];
+         buffer[iout++] = png_digit[c & 0x0f];
+         buffer[iout++] = PNG_LITERAL_RIGHT_SQUARE_BRACKET;
+      }
+
+      else
+      {
+         buffer[iout++] = (char)c;
+      }
+   }
+
+   if (error_message == NULL)
+      buffer[iout] = '\0';
+
+   else
+   {
+      int iin = 0;
+
+      buffer[iout++] = ':';
+      buffer[iout++] = ' ';
+
+      while (iin < PNG_MAX_ERROR_TEXT-1 && error_message[iin] != '\0')
+         buffer[iout++] = error_message[iin++];
+
+      /* iin < PNG_MAX_ERROR_TEXT, so the following is safe: */
+      buffer[iout] = '\0';
+   }
+}
+#endif /* PNG_WARNINGS_SUPPORTED || PNG_ERROR_TEXT_SUPPORTED */
+
+#if defined(PNG_READ_SUPPORTED) && defined(PNG_ERROR_TEXT_SUPPORTED)
+PNG_FUNCTION(void,PNGAPI
+png_chunk_error,(png_const_structrp png_ptr, png_const_charp error_message),
+   PNG_NORETURN)
+{
+   char msg[18+PNG_MAX_ERROR_TEXT];
+   if (png_ptr == NULL)
+      png_error(png_ptr, error_message);
+
+   else
+   {
+      png_format_buffer(png_ptr, msg, error_message);
+      png_error(png_ptr, msg);
+   }
+}
+#endif /* PNG_READ_SUPPORTED && PNG_ERROR_TEXT_SUPPORTED */
+
+#ifdef PNG_WARNINGS_SUPPORTED
+void PNGAPI
+png_chunk_warning(png_const_structrp png_ptr, png_const_charp warning_message)
+{
+   char msg[18+PNG_MAX_ERROR_TEXT];
+   if (png_ptr == NULL)
+      png_warning(png_ptr, warning_message);
+
+   else
+   {
+      png_format_buffer(png_ptr, msg, warning_message);
+      png_warning(png_ptr, msg);
+   }
+}
+#endif /* PNG_WARNINGS_SUPPORTED */
+
+#ifdef PNG_READ_SUPPORTED
+#ifdef PNG_BENIGN_ERRORS_SUPPORTED
+void PNGAPI
+png_chunk_benign_error(png_const_structrp png_ptr, png_const_charp
+    error_message)
+{
+   if (png_ptr->flags & PNG_FLAG_BENIGN_ERRORS_WARN)
+      png_chunk_warning(png_ptr, error_message);
+
+   else
+      png_chunk_error(png_ptr, error_message);
+}
+#endif
+#endif /* PNG_READ_SUPPORTED */
+
+void /* PRIVATE */
+png_chunk_report(png_const_structrp png_ptr, png_const_charp message, int error)
+{
+   /* This is always supported, but for just read or just write it
+    * unconditionally does the right thing.
+    */
+#  if defined(PNG_READ_SUPPORTED) && defined(PNG_WRITE_SUPPORTED)
+      if (png_ptr->mode & PNG_IS_READ_STRUCT)
+#  endif
+
+#  ifdef PNG_READ_SUPPORTED
+      {
+         if (error < PNG_CHUNK_ERROR)
+            png_chunk_warning(png_ptr, message);
+
+         else
+            png_chunk_benign_error(png_ptr, message);
+      }
+#  endif
+
+#  if defined(PNG_READ_SUPPORTED) && defined(PNG_WRITE_SUPPORTED)
+      else if (!(png_ptr->mode & PNG_IS_READ_STRUCT))
+#  endif
+
+#  ifdef PNG_WRITE_SUPPORTED
+      {
+         if (error < PNG_CHUNK_WRITE_ERROR)
+            png_app_warning(png_ptr, message);
+
+         else
+            png_app_error(png_ptr, message);
+      }
+#  endif
+}
+
+#ifdef PNG_ERROR_TEXT_SUPPORTED
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+PNG_FUNCTION(void,
+png_fixed_error,(png_const_structrp png_ptr, png_const_charp name),PNG_NORETURN)
+{
+#  define fixed_message "fixed point overflow in "
+#  define fixed_message_ln ((sizeof fixed_message)-1)
+   int  iin;
+   char msg[fixed_message_ln+PNG_MAX_ERROR_TEXT];
+   memcpy(msg, fixed_message, fixed_message_ln);
+   iin = 0;
+   if (name != NULL) while (iin < (PNG_MAX_ERROR_TEXT-1) && name[iin] != 0)
+   {
+      msg[fixed_message_ln + iin] = name[iin];
+      ++iin;
+   }
+   msg[fixed_message_ln + iin] = 0;
+   png_error(png_ptr, msg);
+}
+#endif
+#endif
+
+#ifdef PNG_SETJMP_SUPPORTED
+/* This API only exists if ANSI-C style error handling is used,
+ * otherwise it is necessary for png_default_error to be overridden.
+ */
+jmp_buf* PNGAPI
+png_set_longjmp_fn(png_structrp png_ptr, png_longjmp_ptr longjmp_fn,
+    size_t jmp_buf_size)
+{
+   /* From libpng 1.6.0 the app gets one chance to set a 'jmpbuf_size' value
+    * and it must not change after that.  Libpng doesn't care how big the
+    * buffer is, just that it doesn't change.
+    *
+    * If the buffer size is no *larger* than the size of jmp_buf when libpng is
+    * compiled a built in jmp_buf is returned; this preserves the pre-1.6.0
+    * semantics that this call will not fail.  If the size is larger, however,
+    * the buffer is allocated and this may fail, causing the function to return
+    * NULL.
+    */
+   if (png_ptr == NULL)
+      return NULL;
+
+   if (png_ptr->jmp_buf_ptr == NULL)
+   {
+      png_ptr->jmp_buf_size = 0; /* not allocated */
+
+      if (jmp_buf_size <= (sizeof png_ptr->jmp_buf_local))
+         png_ptr->jmp_buf_ptr = &png_ptr->jmp_buf_local;
+
+      else
+      {
+         png_ptr->jmp_buf_ptr = png_voidcast(jmp_buf *,
+            png_malloc_warn(png_ptr, jmp_buf_size));
+
+         if (png_ptr->jmp_buf_ptr == NULL)
+            return NULL; /* new NULL return on OOM */
+
+         png_ptr->jmp_buf_size = jmp_buf_size;
+      }
+   }
+
+   else /* Already allocated: check the size */
+   {
+      size_t size = png_ptr->jmp_buf_size;
+
+      if (size == 0)
+      {
+         size = (sizeof png_ptr->jmp_buf_local);
+         if (png_ptr->jmp_buf_ptr != &png_ptr->jmp_buf_local)
+         {
+            /* This is an internal error in libpng: somehow we have been left
+             * with a stack allocated jmp_buf when the application regained
+             * control.  It's always possible to fix this up, but for the moment
+             * this is a png_error because that makes it easy to detect.
+             */
+            png_error(png_ptr, "Libpng jmp_buf still allocated");
+            /* png_ptr->jmp_buf_ptr = &png_ptr->jmp_buf_local; */
+         }
+      }
+
+      if (size != jmp_buf_size)
+      {
+         png_warning(png_ptr, "Application jmp_buf size changed");
+         return NULL; /* caller will probably crash: no choice here */
+      }
+   }
+
+   /* Finally fill in the function, now we have a satisfactory buffer. It is
+    * valid to change the function on every call.
+    */
+   png_ptr->longjmp_fn = longjmp_fn;
+   return png_ptr->jmp_buf_ptr;
+}
+
+void /* PRIVATE */
+png_free_jmpbuf(png_structrp png_ptr)
+{
+   if (png_ptr != NULL)
+   {
+      jmp_buf *jb = png_ptr->jmp_buf_ptr;
+
+      /* A size of 0 is used to indicate a local, stack, allocation of the
+       * pointer; used here and in png.c
+       */
+      if (jb != NULL && png_ptr->jmp_buf_size > 0)
+      {
+
+         /* This stuff is so that a failure to free the error control structure
+          * does not leave libpng in a state with no valid error handling: the
+          * free always succeeds, if there is an error it gets ignored.
+          */
+         if (jb != &png_ptr->jmp_buf_local)
+         {
+            /* Make an internal, libpng, jmp_buf to return here */
+            jmp_buf free_jmp_buf;
+
+            if (!setjmp(free_jmp_buf))
+            {
+               png_ptr->jmp_buf_ptr = &free_jmp_buf; /* come back here */
+               png_ptr->jmp_buf_size = 0; /* stack allocation */
+               png_ptr->longjmp_fn = longjmp;
+               png_free(png_ptr, jb); /* Return to setjmp on error */
+            }
+         }
+      }
+
+      /* *Always* cancel everything out: */
+      png_ptr->jmp_buf_size = 0;
+      png_ptr->jmp_buf_ptr = NULL;
+      png_ptr->longjmp_fn = 0;
+   }
+}
+#endif
+
+/* This is the default error handling function.  Note that replacements for
+ * this function MUST NOT RETURN, or the program will likely crash.  This
+ * function is used by default, or if the program supplies NULL for the
+ * error function pointer in png_set_error_fn().
+ */
+static PNG_FUNCTION(void /* PRIVATE */,
+png_default_error,(png_const_structrp png_ptr, png_const_charp error_message),
+   PNG_NORETURN)
+{
+#ifdef PNG_CONSOLE_IO_SUPPORTED
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+   /* Check on NULL only added in 1.5.4 */
+   if (error_message != NULL && *error_message == PNG_LITERAL_SHARP)
+   {
+      /* Strip "#nnnn " from beginning of error message. */
+      int offset;
+      char error_number[16];
+      for (offset = 0; offset<15; offset++)
+      {
+         error_number[offset] = error_message[offset + 1];
+         if (error_message[offset] == ' ')
+            break;
+      }
+
+      if ((offset > 1) && (offset < 15))
+      {
+         error_number[offset - 1] = '\0';
+         fprintf(stderr, "libpng error no. %s: %s",
+             error_number, error_message + offset + 1);
+         fprintf(stderr, PNG_STRING_NEWLINE);
+      }
+
+      else
+      {
+         fprintf(stderr, "libpng error: %s, offset=%d",
+             error_message, offset);
+         fprintf(stderr, PNG_STRING_NEWLINE);
+      }
+   }
+   else
+#endif
+   {
+      fprintf(stderr, "libpng error: %s", error_message ? error_message :
+         "undefined");
+      fprintf(stderr, PNG_STRING_NEWLINE);
+   }
+#else
+   PNG_UNUSED(error_message) /* Make compiler happy */
+#endif
+   png_longjmp(png_ptr, 1);
+}
+
+PNG_FUNCTION(void,PNGAPI
+png_longjmp,(png_const_structrp png_ptr, int val),PNG_NORETURN)
+{
+#ifdef PNG_SETJMP_SUPPORTED
+   if (png_ptr && png_ptr->longjmp_fn && png_ptr->jmp_buf_ptr)
+      png_ptr->longjmp_fn(*png_ptr->jmp_buf_ptr, val);
+#endif
+
+   /* Here if not setjmp support or if png_ptr is null. */
+   PNG_ABORT();
+}
+
+#ifdef PNG_WARNINGS_SUPPORTED
+/* This function is called when there is a warning, but the library thinks
+ * it can continue anyway.  Replacement functions don't have to do anything
+ * here if you don't want them to.  In the default configuration, png_ptr is
+ * not used, but it is passed in case it may be useful.
+ */
+static void /* PRIVATE */
+png_default_warning(png_const_structrp png_ptr, png_const_charp warning_message)
+{
+#ifdef PNG_CONSOLE_IO_SUPPORTED
+#  ifdef PNG_ERROR_NUMBERS_SUPPORTED
+   if (*warning_message == PNG_LITERAL_SHARP)
+   {
+      int offset;
+      char warning_number[16];
+      for (offset = 0; offset < 15; offset++)
+      {
+         warning_number[offset] = warning_message[offset + 1];
+         if (warning_message[offset] == ' ')
+            break;
+      }
+
+      if ((offset > 1) && (offset < 15))
+      {
+         warning_number[offset + 1] = '\0';
+         fprintf(stderr, "libpng warning no. %s: %s",
+             warning_number, warning_message + offset);
+         fprintf(stderr, PNG_STRING_NEWLINE);
+      }
+
+      else
+      {
+         fprintf(stderr, "libpng warning: %s",
+             warning_message);
+         fprintf(stderr, PNG_STRING_NEWLINE);
+      }
+   }
+   else
+#  endif
+
+   {
+      fprintf(stderr, "libpng warning: %s", warning_message);
+      fprintf(stderr, PNG_STRING_NEWLINE);
+   }
+#else
+   PNG_UNUSED(warning_message) /* Make compiler happy */
+#endif
+   PNG_UNUSED(png_ptr) /* Make compiler happy */
+}
+#endif /* PNG_WARNINGS_SUPPORTED */
+
+/* This function is called when the application wants to use another method
+ * of handling errors and warnings.  Note that the error function MUST NOT
+ * return to the calling routine or serious problems will occur.  The return
+ * method used in the default routine calls longjmp(png_ptr->jmp_buf_ptr, 1)
+ */
+void PNGAPI
+png_set_error_fn(png_structrp png_ptr, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warning_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->error_ptr = error_ptr;
+   png_ptr->error_fn = error_fn;
+#ifdef PNG_WARNINGS_SUPPORTED
+   png_ptr->warning_fn = warning_fn;
+#else
+   PNG_UNUSED(warning_fn)
+#endif
+}
+
+
+/* This function returns a pointer to the error_ptr associated with the user
+ * functions.  The application should free any memory associated with this
+ * pointer before png_write_destroy and png_read_destroy are called.
+ */
+png_voidp PNGAPI
+png_get_error_ptr(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return NULL;
+
+   return ((png_voidp)png_ptr->error_ptr);
+}
+
+
+#ifdef PNG_ERROR_NUMBERS_SUPPORTED
+void PNGAPI
+png_set_strip_error_numbers(png_structrp png_ptr, png_uint_32 strip_mode)
+{
+   if (png_ptr != NULL)
+   {
+      png_ptr->flags &=
+         ((~(PNG_FLAG_STRIP_ERROR_NUMBERS |
+         PNG_FLAG_STRIP_ERROR_TEXT))&strip_mode);
+   }
+}
+#endif
+
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) ||\
+   defined(PNG_SIMPLIFIED_WRITE_SUPPORTED)
+   /* Currently the above both depend on SETJMP_SUPPORTED, however it would be
+    * possible to implement without setjmp support just so long as there is some
+    * way to handle the error return here:
+    */
+PNG_FUNCTION(void /* PRIVATE */,
+png_safe_error,(png_structp png_nonconst_ptr, png_const_charp error_message),
+   PNG_NORETURN)
+{
+   const png_const_structrp png_ptr = png_nonconst_ptr;
+   png_imagep image = png_voidcast(png_imagep, png_ptr->error_ptr);
+
+   /* An error is always logged here, overwriting anything (typically a warning)
+    * that is already there:
+    */
+   if (image != NULL)
+   {
+      png_safecat(image->message, (sizeof image->message), 0, error_message);
+      image->warning_or_error |= PNG_IMAGE_ERROR;
+
+      /* Retrieve the jmp_buf from within the png_control, making this work for
+       * C++ compilation too is pretty tricky: C++ wants a pointer to the first
+       * element of a jmp_buf, but C doesn't tell us the type of that.
+       */
+      if (image->opaque != NULL && image->opaque->error_buf != NULL)
+         longjmp(png_control_jmp_buf(image->opaque), 1);
+
+      /* Missing longjmp buffer, the following is to help debugging: */
+      {
+         size_t pos = png_safecat(image->message, (sizeof image->message), 0,
+            "bad longjmp: ");
+         png_safecat(image->message, (sizeof image->message), pos,
+             error_message);
+      }
+   }
+
+   /* Here on an internal programming error. */
+   abort();
+}
+
+#ifdef PNG_WARNINGS_SUPPORTED
+void /* PRIVATE */
+png_safe_warning(png_structp png_nonconst_ptr, png_const_charp warning_message)
+{
+   const png_const_structrp png_ptr = png_nonconst_ptr;
+   png_imagep image = png_voidcast(png_imagep, png_ptr->error_ptr);
+
+   /* A warning is only logged if there is no prior warning or error. */
+   if (image->warning_or_error == 0)
+   {
+      png_safecat(image->message, (sizeof image->message), 0, warning_message);
+      image->warning_or_error |= PNG_IMAGE_WARNING;
+   }
+}
+#endif
+
+int /* PRIVATE */
+png_safe_execute(png_imagep image_in, int (*function)(png_voidp), png_voidp arg)
+{
+   volatile png_imagep image = image_in;
+   volatile int result;
+   volatile png_voidp saved_error_buf;
+   jmp_buf safe_jmpbuf;
+
+   /* Safely execute function(arg) with png_error returning to this function. */
+   saved_error_buf = image->opaque->error_buf;
+   result = setjmp(safe_jmpbuf) == 0;
+
+   if (result)
+   {
+
+      image->opaque->error_buf = safe_jmpbuf;
+      result = function(arg);
+   }
+
+   image->opaque->error_buf = saved_error_buf;
+
+   /* And do the cleanup prior to any failure return. */
+   if (!result)
+      png_image_free(image);
+
+   return result;
+}
+#endif /* SIMPLIFIED READ/WRITE */
+#endif /* PNG_READ_SUPPORTED || PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngget.c b/ml/dlib/dlib/external/libpng/pngget.c
new file mode 100644
index 000000000..aca63a958
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngget.c
@@ -0,0 +1,1177 @@
+
+/* pngget.c - retrieval of values from info struct
+ *
+ * Last changed in libpng 1.6.1 [March 28, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ */
+
+#include "pngpriv.h"
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+
+png_uint_32 PNGAPI
+png_get_valid(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_uint_32 flag)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return(info_ptr->valid & flag);
+
+   return(0);
+}
+
+png_size_t PNGAPI
+png_get_rowbytes(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return(info_ptr->rowbytes);
+
+   return(0);
+}
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+png_bytepp PNGAPI
+png_get_rows(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return(info_ptr->row_pointers);
+
+   return(0);
+}
+#endif
+
+#ifdef PNG_EASY_ACCESS_SUPPORTED
+/* Easy access to info, added in libpng-0.99 */
+png_uint_32 PNGAPI
+png_get_image_width(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->width;
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_image_height(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->height;
+
+   return (0);
+}
+
+png_byte PNGAPI
+png_get_bit_depth(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->bit_depth;
+
+   return (0);
+}
+
+png_byte PNGAPI
+png_get_color_type(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->color_type;
+
+   return (0);
+}
+
+png_byte PNGAPI
+png_get_filter_type(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->filter_type;
+
+   return (0);
+}
+
+png_byte PNGAPI
+png_get_interlace_type(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->interlace_type;
+
+   return (0);
+}
+
+png_byte PNGAPI
+png_get_compression_type(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return info_ptr->compression_type;
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_x_pixels_per_meter(png_const_structrp png_ptr, png_const_inforp
+   info_ptr)
+{
+#ifdef PNG_pHYs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+      {
+         png_debug1(1, "in %s retrieval function",
+             "png_get_x_pixels_per_meter");
+
+         if (info_ptr->phys_unit_type == PNG_RESOLUTION_METER)
+            return (info_ptr->x_pixels_per_unit);
+      }
+#endif
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_y_pixels_per_meter(png_const_structrp png_ptr, png_const_inforp
+    info_ptr)
+{
+#ifdef PNG_pHYs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      png_debug1(1, "in %s retrieval function",
+          "png_get_y_pixels_per_meter");
+
+      if (info_ptr->phys_unit_type == PNG_RESOLUTION_METER)
+         return (info_ptr->y_pixels_per_unit);
+   }
+#endif
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_pixels_per_meter(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+#ifdef PNG_pHYs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_pixels_per_meter");
+
+      if (info_ptr->phys_unit_type == PNG_RESOLUTION_METER &&
+          info_ptr->x_pixels_per_unit == info_ptr->y_pixels_per_unit)
+         return (info_ptr->x_pixels_per_unit);
+   }
+#endif
+
+   return (0);
+}
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+float PNGAPI
+png_get_pixel_aspect_ratio(png_const_structrp png_ptr, png_const_inforp
+   info_ptr)
+{
+#ifdef PNG_READ_pHYs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_aspect_ratio");
+
+      if (info_ptr->x_pixels_per_unit != 0)
+         return ((float)((float)info_ptr->y_pixels_per_unit
+             /(float)info_ptr->x_pixels_per_unit));
+   }
+#else
+   PNG_UNUSED(png_ptr)
+   PNG_UNUSED(info_ptr)
+#endif
+
+   return ((float)0.0);
+}
+#endif
+
+#ifdef PNG_FIXED_POINT_SUPPORTED
+png_fixed_point PNGAPI
+png_get_pixel_aspect_ratio_fixed(png_const_structrp png_ptr,
+    png_const_inforp info_ptr)
+{
+#ifdef PNG_READ_pHYs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs)
+       && info_ptr->x_pixels_per_unit > 0 && info_ptr->y_pixels_per_unit > 0
+       && info_ptr->x_pixels_per_unit <= PNG_UINT_31_MAX
+       && info_ptr->y_pixels_per_unit <= PNG_UINT_31_MAX)
+   {
+      png_fixed_point res;
+
+      png_debug1(1, "in %s retrieval function", "png_get_aspect_ratio_fixed");
+
+      /* The following casts work because a PNG 4 byte integer only has a valid
+       * range of 0..2^31-1; otherwise the cast might overflow.
+       */
+      if (png_muldiv(&res, (png_int_32)info_ptr->y_pixels_per_unit, PNG_FP_1,
+          (png_int_32)info_ptr->x_pixels_per_unit))
+         return res;
+   }
+#else
+   PNG_UNUSED(png_ptr)
+   PNG_UNUSED(info_ptr)
+#endif
+
+   return 0;
+}
+#endif
+
+png_int_32 PNGAPI
+png_get_x_offset_microns(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+#ifdef PNG_oFFs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_x_offset_microns");
+
+      if (info_ptr->offset_unit_type == PNG_OFFSET_MICROMETER)
+         return (info_ptr->x_offset);
+   }
+#endif
+
+   return (0);
+}
+
+png_int_32 PNGAPI
+png_get_y_offset_microns(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+#ifdef PNG_oFFs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_y_offset_microns");
+
+      if (info_ptr->offset_unit_type == PNG_OFFSET_MICROMETER)
+         return (info_ptr->y_offset);
+   }
+#endif
+
+   return (0);
+}
+
+png_int_32 PNGAPI
+png_get_x_offset_pixels(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+#ifdef PNG_oFFs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_x_offset_pixels");
+
+      if (info_ptr->offset_unit_type == PNG_OFFSET_PIXEL)
+         return (info_ptr->x_offset);
+   }
+#endif
+
+   return (0);
+}
+
+png_int_32 PNGAPI
+png_get_y_offset_pixels(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+#ifdef PNG_oFFs_SUPPORTED
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs))
+   {
+      png_debug1(1, "in %s retrieval function", "png_get_y_offset_pixels");
+
+      if (info_ptr->offset_unit_type == PNG_OFFSET_PIXEL)
+         return (info_ptr->y_offset);
+   }
+#endif
+
+   return (0);
+}
+
+#ifdef PNG_INCH_CONVERSIONS_SUPPORTED
+static png_uint_32
+ppi_from_ppm(png_uint_32 ppm)
+{
+#if 0
+   /* The conversion is *(2.54/100), in binary (32 digits):
+    * .00000110100000001001110101001001
+    */
+   png_uint_32 t1001, t1101;
+   ppm >>= 1;                  /* .1 */
+   t1001 = ppm + (ppm >> 3);   /* .1001 */
+   t1101 = t1001 + (ppm >> 1); /* .1101 */
+   ppm >>= 20;                 /* .000000000000000000001 */
+   t1101 += t1101 >> 15;       /* .1101000000000001101 */
+   t1001 >>= 11;               /* .000000000001001 */
+   t1001 += t1001 >> 12;       /* .000000000001001000000001001 */
+   ppm += t1001;               /* .000000000001001000001001001 */
+   ppm += t1101;               /* .110100000001001110101001001 */
+   return (ppm + 16) >> 5;/* .00000110100000001001110101001001 */
+#else
+   /* The argument is a PNG unsigned integer, so it is not permitted
+    * to be bigger than 2^31.
+    */
+   png_fixed_point result;
+   if (ppm <= PNG_UINT_31_MAX && png_muldiv(&result, (png_int_32)ppm, 127,
+       5000))
+      return result;
+
+   /* Overflow. */
+   return 0;
+#endif
+}
+
+png_uint_32 PNGAPI
+png_get_pixels_per_inch(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   return ppi_from_ppm(png_get_pixels_per_meter(png_ptr, info_ptr));
+}
+
+png_uint_32 PNGAPI
+png_get_x_pixels_per_inch(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   return ppi_from_ppm(png_get_x_pixels_per_meter(png_ptr, info_ptr));
+}
+
+png_uint_32 PNGAPI
+png_get_y_pixels_per_inch(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   return ppi_from_ppm(png_get_y_pixels_per_meter(png_ptr, info_ptr));
+}
+
+#ifdef PNG_FIXED_POINT_SUPPORTED
+static png_fixed_point
+png_fixed_inches_from_microns(png_const_structrp png_ptr, png_int_32 microns)
+{
+   /* Convert from metres * 1,000,000 to inches * 100,000, meters to
+    * inches is simply *(100/2.54), so we want *(10/2.54) == 500/127.
+    * Notice that this can overflow - a warning is output and 0 is
+    * returned.
+    */
+   return png_muldiv_warn(png_ptr, microns, 500, 127);
+}
+
+png_fixed_point PNGAPI
+png_get_x_offset_inches_fixed(png_const_structrp png_ptr,
+    png_const_inforp info_ptr)
+{
+   return png_fixed_inches_from_microns(png_ptr,
+       png_get_x_offset_microns(png_ptr, info_ptr));
+}
+#endif
+
+#ifdef PNG_FIXED_POINT_SUPPORTED
+png_fixed_point PNGAPI
+png_get_y_offset_inches_fixed(png_const_structrp png_ptr,
+    png_const_inforp info_ptr)
+{
+   return png_fixed_inches_from_microns(png_ptr,
+       png_get_y_offset_microns(png_ptr, info_ptr));
+}
+#endif
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+float PNGAPI
+png_get_x_offset_inches(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   /* To avoid the overflow do the conversion directly in floating
+    * point.
+    */
+   return (float)(png_get_x_offset_microns(png_ptr, info_ptr) * .00003937);
+}
+#endif
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+float PNGAPI
+png_get_y_offset_inches(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   /* To avoid the overflow do the conversion directly in floating
+    * point.
+    */
+   return (float)(png_get_y_offset_microns(png_ptr, info_ptr) * .00003937);
+}
+#endif
+
+#ifdef PNG_pHYs_SUPPORTED
+png_uint_32 PNGAPI
+png_get_pHYs_dpi(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_uint_32 *res_x, png_uint_32 *res_y, int *unit_type)
+{
+   png_uint_32 retval = 0;
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      png_debug1(1, "in %s retrieval function", "pHYs");
+
+      if (res_x != NULL)
+      {
+         *res_x = info_ptr->x_pixels_per_unit;
+         retval |= PNG_INFO_pHYs;
+      }
+
+      if (res_y != NULL)
+      {
+         *res_y = info_ptr->y_pixels_per_unit;
+         retval |= PNG_INFO_pHYs;
+      }
+
+      if (unit_type != NULL)
+      {
+         *unit_type = (int)info_ptr->phys_unit_type;
+         retval |= PNG_INFO_pHYs;
+
+         if (*unit_type == 1)
+         {
+            if (res_x != NULL) *res_x = (png_uint_32)(*res_x * .0254 + .50);
+            if (res_y != NULL) *res_y = (png_uint_32)(*res_y * .0254 + .50);
+         }
+      }
+   }
+
+   return (retval);
+}
+#endif /* PNG_pHYs_SUPPORTED */
+#endif  /* PNG_INCH_CONVERSIONS_SUPPORTED */
+
+/* png_get_channels really belongs in here, too, but it's been around longer */
+
+#endif  /* PNG_EASY_ACCESS_SUPPORTED */
+
+
+png_byte PNGAPI
+png_get_channels(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return(info_ptr->channels);
+
+   return (0);
+}
+
+#ifdef PNG_READ_SUPPORTED
+png_const_bytep PNGAPI
+png_get_signature(png_const_structrp png_ptr, png_const_inforp info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return(info_ptr->signature);
+
+   return (NULL);
+}
+#endif
+
+#ifdef PNG_bKGD_SUPPORTED
+png_uint_32 PNGAPI
+png_get_bKGD(png_const_structrp png_ptr, png_inforp info_ptr,
+   png_color_16p *background)
+{
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_bKGD)
+       && background != NULL)
+   {
+      png_debug1(1, "in %s retrieval function", "bKGD");
+
+      *background = &(info_ptr->background);
+      return (PNG_INFO_bKGD);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_cHRM_SUPPORTED
+/* The XYZ APIs were added in 1.5.5 to take advantage of the code added at the
+ * same time to correct the rgb grayscale coefficient defaults obtained from the
+ * cHRM chunk in 1.5.4
+ */
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_cHRM(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    double *white_x, double *white_y, double *red_x, double *red_y,
+    double *green_x, double *green_y, double *blue_x, double *blue_y)
+{
+   /* Quiet API change: this code used to only return the end points if a cHRM
+    * chunk was present, but the end points can also come from iCCP or sRGB
+    * chunks, so in 1.6.0 the png_get_ APIs return the end points regardless and
+    * the png_set_ APIs merely check that set end points are mutually
+    * consistent.
+    */
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS))
+   {
+      png_debug1(1, "in %s retrieval function", "cHRM");
+
+      if (white_x != NULL)
+         *white_x = png_float(png_ptr,
+            info_ptr->colorspace.end_points_xy.whitex, "cHRM white X");
+      if (white_y != NULL)
+         *white_y = png_float(png_ptr,
+            info_ptr->colorspace.end_points_xy.whitey, "cHRM white Y");
+      if (red_x != NULL)
+         *red_x = png_float(png_ptr, info_ptr->colorspace.end_points_xy.redx,
+            "cHRM red X");
+      if (red_y != NULL)
+         *red_y = png_float(png_ptr, info_ptr->colorspace.end_points_xy.redy,
+            "cHRM red Y");
+      if (green_x != NULL)
+         *green_x = png_float(png_ptr,
+            info_ptr->colorspace.end_points_xy.greenx, "cHRM green X");
+      if (green_y != NULL)
+         *green_y = png_float(png_ptr,
+            info_ptr->colorspace.end_points_xy.greeny, "cHRM green Y");
+      if (blue_x != NULL)
+         *blue_x = png_float(png_ptr, info_ptr->colorspace.end_points_xy.bluex,
+            "cHRM blue X");
+      if (blue_y != NULL)
+         *blue_y = png_float(png_ptr, info_ptr->colorspace.end_points_xy.bluey,
+            "cHRM blue Y");
+      return (PNG_INFO_cHRM);
+   }
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_cHRM_XYZ(png_const_structrp png_ptr, png_const_inforp info_ptr,
+   double *red_X, double *red_Y, double *red_Z, double *green_X,
+   double *green_Y, double *green_Z, double *blue_X, double *blue_Y,
+   double *blue_Z)
+{
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS))
+   {
+      png_debug1(1, "in %s retrieval function", "cHRM_XYZ(float)");
+
+      if (red_X != NULL)
+         *red_X = png_float(png_ptr, info_ptr->colorspace.end_points_XYZ.red_X,
+            "cHRM red X");
+      if (red_Y != NULL)
+         *red_Y = png_float(png_ptr, info_ptr->colorspace.end_points_XYZ.red_Y,
+            "cHRM red Y");
+      if (red_Z != NULL)
+         *red_Z = png_float(png_ptr, info_ptr->colorspace.end_points_XYZ.red_Z,
+            "cHRM red Z");
+      if (green_X != NULL)
+         *green_X = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.green_X, "cHRM green X");
+      if (green_Y != NULL)
+         *green_Y = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.green_Y, "cHRM green Y");
+      if (green_Z != NULL)
+         *green_Z = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.green_Z, "cHRM green Z");
+      if (blue_X != NULL)
+         *blue_X = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.blue_X, "cHRM blue X");
+      if (blue_Y != NULL)
+         *blue_Y = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.blue_Y, "cHRM blue Y");
+      if (blue_Z != NULL)
+         *blue_Z = png_float(png_ptr,
+            info_ptr->colorspace.end_points_XYZ.blue_Z, "cHRM blue Z");
+      return (PNG_INFO_cHRM);
+   }
+
+   return (0);
+}
+#  endif
+
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_cHRM_XYZ_fixed(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *int_red_X, png_fixed_point *int_red_Y,
+    png_fixed_point *int_red_Z, png_fixed_point *int_green_X,
+    png_fixed_point *int_green_Y, png_fixed_point *int_green_Z,
+    png_fixed_point *int_blue_X, png_fixed_point *int_blue_Y,
+    png_fixed_point *int_blue_Z)
+{
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS))
+   {
+      png_debug1(1, "in %s retrieval function", "cHRM_XYZ");
+
+      if (int_red_X != NULL)
+         *int_red_X = info_ptr->colorspace.end_points_XYZ.red_X;
+      if (int_red_Y != NULL)
+         *int_red_Y = info_ptr->colorspace.end_points_XYZ.red_Y;
+      if (int_red_Z != NULL)
+         *int_red_Z = info_ptr->colorspace.end_points_XYZ.red_Z;
+      if (int_green_X != NULL)
+         *int_green_X = info_ptr->colorspace.end_points_XYZ.green_X;
+      if (int_green_Y != NULL)
+         *int_green_Y = info_ptr->colorspace.end_points_XYZ.green_Y;
+      if (int_green_Z != NULL)
+         *int_green_Z = info_ptr->colorspace.end_points_XYZ.green_Z;
+      if (int_blue_X != NULL)
+         *int_blue_X = info_ptr->colorspace.end_points_XYZ.blue_X;
+      if (int_blue_Y != NULL)
+         *int_blue_Y = info_ptr->colorspace.end_points_XYZ.blue_Y;
+      if (int_blue_Z != NULL)
+         *int_blue_Z = info_ptr->colorspace.end_points_XYZ.blue_Z;
+      return (PNG_INFO_cHRM);
+   }
+
+   return (0);
+}
+
+png_uint_32 PNGAPI
+png_get_cHRM_fixed(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *white_x, png_fixed_point *white_y, png_fixed_point *red_x,
+    png_fixed_point *red_y, png_fixed_point *green_x, png_fixed_point *green_y,
+    png_fixed_point *blue_x, png_fixed_point *blue_y)
+{
+   png_debug1(1, "in %s retrieval function", "cHRM");
+
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_ENDPOINTS))
+   {
+      if (white_x != NULL)
+         *white_x = info_ptr->colorspace.end_points_xy.whitex;
+      if (white_y != NULL)
+         *white_y = info_ptr->colorspace.end_points_xy.whitey;
+      if (red_x != NULL)
+         *red_x = info_ptr->colorspace.end_points_xy.redx;
+      if (red_y != NULL)
+         *red_y = info_ptr->colorspace.end_points_xy.redy;
+      if (green_x != NULL)
+         *green_x = info_ptr->colorspace.end_points_xy.greenx;
+      if (green_y != NULL)
+         *green_y = info_ptr->colorspace.end_points_xy.greeny;
+      if (blue_x != NULL)
+         *blue_x = info_ptr->colorspace.end_points_xy.bluex;
+      if (blue_y != NULL)
+         *blue_y = info_ptr->colorspace.end_points_xy.bluey;
+      return (PNG_INFO_cHRM);
+   }
+
+   return (0);
+}
+#  endif
+#endif
+
+#ifdef PNG_gAMA_SUPPORTED
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_gAMA_fixed(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_fixed_point *file_gamma)
+{
+   png_debug1(1, "in %s retrieval function", "gAMA");
+
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_GAMMA) &&
+      file_gamma != NULL)
+   {
+      *file_gamma = info_ptr->colorspace.gamma;
+      return (PNG_INFO_gAMA);
+   }
+
+   return (0);
+}
+#  endif
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_gAMA(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    double *file_gamma)
+{
+   png_debug1(1, "in %s retrieval function", "gAMA(float)");
+
+   if (png_ptr != NULL && info_ptr != NULL &&
+      (info_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_GAMMA) &&
+      file_gamma != NULL)
+   {
+      *file_gamma = png_float(png_ptr, info_ptr->colorspace.gamma,
+         "png_get_gAMA");
+      return (PNG_INFO_gAMA);
+   }
+
+   return (0);
+}
+#  endif
+#endif
+
+#ifdef PNG_sRGB_SUPPORTED
+png_uint_32 PNGAPI
+png_get_sRGB(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    int *file_srgb_intent)
+{
+   png_debug1(1, "in %s retrieval function", "sRGB");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_sRGB)
+       && file_srgb_intent != NULL)
+   {
+      *file_srgb_intent = info_ptr->colorspace.rendering_intent;
+      return (PNG_INFO_sRGB);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_iCCP_SUPPORTED
+png_uint_32 PNGAPI
+png_get_iCCP(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_charpp name, int *compression_type,
+    png_bytepp profile, png_uint_32 *proflen)
+{
+   png_debug1(1, "in %s retrieval function", "iCCP");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_iCCP)
+       && name != NULL && compression_type != NULL && profile != NULL &&
+		 proflen != NULL)
+   {
+      *name = info_ptr->iccp_name;
+      *profile = info_ptr->iccp_profile;
+      *proflen = png_get_uint_32(info_ptr->iccp_profile);
+      /* This is somewhat irrelevant since the profile data returned has
+       * actually been uncompressed.
+       */
+      *compression_type = PNG_COMPRESSION_TYPE_BASE;
+      return (PNG_INFO_iCCP);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+int PNGAPI
+png_get_sPLT(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_sPLT_tpp spalettes)
+{
+   if (png_ptr != NULL && info_ptr != NULL && spalettes != NULL)
+   {
+      *spalettes = info_ptr->splt_palettes;
+      return info_ptr->splt_palettes_num;
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+png_uint_32 PNGAPI
+png_get_hIST(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_uint_16p *hist)
+{
+   png_debug1(1, "in %s retrieval function", "hIST");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_hIST)
+       && hist != NULL)
+   {
+      *hist = info_ptr->hist;
+      return (PNG_INFO_hIST);
+   }
+
+   return (0);
+}
+#endif
+
+png_uint_32 PNGAPI
+png_get_IHDR(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_uint_32 *width, png_uint_32 *height, int *bit_depth,
+    int *color_type, int *interlace_type, int *compression_type,
+    int *filter_type)
+{
+   png_debug1(1, "in %s retrieval function", "IHDR");
+
+   if (png_ptr == NULL || info_ptr == NULL || width == NULL ||
+       height == NULL || bit_depth == NULL || color_type == NULL)
+      return (0);
+
+   *width = info_ptr->width;
+   *height = info_ptr->height;
+   *bit_depth = info_ptr->bit_depth;
+   *color_type = info_ptr->color_type;
+
+   if (compression_type != NULL)
+      *compression_type = info_ptr->compression_type;
+
+   if (filter_type != NULL)
+      *filter_type = info_ptr->filter_type;
+
+   if (interlace_type != NULL)
+      *interlace_type = info_ptr->interlace_type;
+
+   /* This is redundant if we can be sure that the info_ptr values were all
+    * assigned in png_set_IHDR().  We do the check anyhow in case an
+    * application has ignored our advice not to mess with the members
+    * of info_ptr directly.
+    */
+   png_check_IHDR(png_ptr, info_ptr->width, info_ptr->height,
+       info_ptr->bit_depth, info_ptr->color_type, info_ptr->interlace_type,
+       info_ptr->compression_type, info_ptr->filter_type);
+
+   return (1);
+}
+
+#ifdef PNG_oFFs_SUPPORTED
+png_uint_32 PNGAPI
+png_get_oFFs(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_int_32 *offset_x, png_int_32 *offset_y, int *unit_type)
+{
+   png_debug1(1, "in %s retrieval function", "oFFs");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs)
+       && offset_x != NULL && offset_y != NULL && unit_type != NULL)
+   {
+      *offset_x = info_ptr->x_offset;
+      *offset_y = info_ptr->y_offset;
+      *unit_type = (int)info_ptr->offset_unit_type;
+      return (PNG_INFO_oFFs);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+png_uint_32 PNGAPI
+png_get_pCAL(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_charp *purpose, png_int_32 *X0, png_int_32 *X1, int *type, int *nparams,
+    png_charp *units, png_charpp *params)
+{
+   png_debug1(1, "in %s retrieval function", "pCAL");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_pCAL)
+       && purpose != NULL && X0 != NULL && X1 != NULL && type != NULL &&
+       nparams != NULL && units != NULL && params != NULL)
+   {
+      *purpose = info_ptr->pcal_purpose;
+      *X0 = info_ptr->pcal_X0;
+      *X1 = info_ptr->pcal_X1;
+      *type = (int)info_ptr->pcal_type;
+      *nparams = (int)info_ptr->pcal_nparams;
+      *units = info_ptr->pcal_units;
+      *params = info_ptr->pcal_params;
+      return (PNG_INFO_pCAL);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_sCAL_SUPPORTED
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+#    if defined(PNG_FLOATING_ARITHMETIC_SUPPORTED) || \
+         defined(PNG_FLOATING_POINT_SUPPORTED)
+png_uint_32 PNGAPI
+png_get_sCAL_fixed(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    int *unit, png_fixed_point *width, png_fixed_point *height)
+{
+   if (png_ptr != NULL && info_ptr != NULL &&
+       (info_ptr->valid & PNG_INFO_sCAL))
+   {
+      *unit = info_ptr->scal_unit;
+      /*TODO: make this work without FP support; the API is currently eliminated
+       * if neither floating point APIs nor internal floating point arithmetic
+       * are enabled.
+       */
+      *width = png_fixed(png_ptr, atof(info_ptr->scal_s_width), "sCAL width");
+      *height = png_fixed(png_ptr, atof(info_ptr->scal_s_height),
+         "sCAL height");
+      return (PNG_INFO_sCAL);
+   }
+
+   return(0);
+}
+#    endif /* FLOATING_ARITHMETIC */
+#  endif /* FIXED_POINT */
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_sCAL(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    int *unit, double *width, double *height)
+{
+   if (png_ptr != NULL && info_ptr != NULL &&
+       (info_ptr->valid & PNG_INFO_sCAL))
+   {
+      *unit = info_ptr->scal_unit;
+      *width = atof(info_ptr->scal_s_width);
+      *height = atof(info_ptr->scal_s_height);
+      return (PNG_INFO_sCAL);
+   }
+
+   return(0);
+}
+#  endif /* FLOATING POINT */
+png_uint_32 PNGAPI
+png_get_sCAL_s(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    int *unit, png_charpp width, png_charpp height)
+{
+   if (png_ptr != NULL && info_ptr != NULL &&
+       (info_ptr->valid & PNG_INFO_sCAL))
+   {
+      *unit = info_ptr->scal_unit;
+      *width = info_ptr->scal_s_width;
+      *height = info_ptr->scal_s_height;
+      return (PNG_INFO_sCAL);
+   }
+
+   return(0);
+}
+#endif /* sCAL */
+
+#ifdef PNG_pHYs_SUPPORTED
+png_uint_32 PNGAPI
+png_get_pHYs(png_const_structrp png_ptr, png_const_inforp info_ptr,
+    png_uint_32 *res_x, png_uint_32 *res_y, int *unit_type)
+{
+   png_uint_32 retval = 0;
+
+   png_debug1(1, "in %s retrieval function", "pHYs");
+
+   if (png_ptr != NULL && info_ptr != NULL &&
+       (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      if (res_x != NULL)
+      {
+         *res_x = info_ptr->x_pixels_per_unit;
+         retval |= PNG_INFO_pHYs;
+      }
+
+      if (res_y != NULL)
+      {
+         *res_y = info_ptr->y_pixels_per_unit;
+         retval |= PNG_INFO_pHYs;
+      }
+
+      if (unit_type != NULL)
+      {
+         *unit_type = (int)info_ptr->phys_unit_type;
+         retval |= PNG_INFO_pHYs;
+      }
+   }
+
+   return (retval);
+}
+#endif /* pHYs */
+
+png_uint_32 PNGAPI
+png_get_PLTE(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_colorp *palette, int *num_palette)
+{
+   png_debug1(1, "in %s retrieval function", "PLTE");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_PLTE)
+       && palette != NULL)
+   {
+      *palette = info_ptr->palette;
+      *num_palette = info_ptr->num_palette;
+      png_debug1(3, "num_palette = %d", *num_palette);
+      return (PNG_INFO_PLTE);
+   }
+
+   return (0);
+}
+
+#ifdef PNG_sBIT_SUPPORTED
+png_uint_32 PNGAPI
+png_get_sBIT(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_color_8p *sig_bit)
+{
+   png_debug1(1, "in %s retrieval function", "sBIT");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_sBIT)
+       && sig_bit != NULL)
+   {
+      *sig_bit = &(info_ptr->sig_bit);
+      return (PNG_INFO_sBIT);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_TEXT_SUPPORTED
+int PNGAPI
+png_get_text(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_textp *text_ptr, int *num_text)
+{
+   if (png_ptr != NULL && info_ptr != NULL && info_ptr->num_text > 0)
+   {
+      png_debug1(1, "in 0x%lx retrieval function",
+         (unsigned long)png_ptr->chunk_name);
+
+      if (text_ptr != NULL)
+         *text_ptr = info_ptr->text;
+
+      if (num_text != NULL)
+         *num_text = info_ptr->num_text;
+
+      return info_ptr->num_text;
+   }
+
+   if (num_text != NULL)
+      *num_text = 0;
+
+   return(0);
+}
+#endif
+
+#ifdef PNG_tIME_SUPPORTED
+png_uint_32 PNGAPI
+png_get_tIME(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_timep *mod_time)
+{
+   png_debug1(1, "in %s retrieval function", "tIME");
+
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_tIME)
+       && mod_time != NULL)
+   {
+      *mod_time = &(info_ptr->mod_time);
+      return (PNG_INFO_tIME);
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_tRNS_SUPPORTED
+png_uint_32 PNGAPI
+png_get_tRNS(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_bytep *trans_alpha, int *num_trans, png_color_16p *trans_color)
+{
+   png_uint_32 retval = 0;
+   if (png_ptr != NULL && info_ptr != NULL && (info_ptr->valid & PNG_INFO_tRNS))
+   {
+      png_debug1(1, "in %s retrieval function", "tRNS");
+
+      if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         if (trans_alpha != NULL)
+         {
+            *trans_alpha = info_ptr->trans_alpha;
+            retval |= PNG_INFO_tRNS;
+         }
+
+         if (trans_color != NULL)
+            *trans_color = &(info_ptr->trans_color);
+      }
+
+      else /* if (info_ptr->color_type != PNG_COLOR_TYPE_PALETTE) */
+      {
+         if (trans_color != NULL)
+         {
+            *trans_color = &(info_ptr->trans_color);
+            retval |= PNG_INFO_tRNS;
+         }
+
+         if (trans_alpha != NULL)
+            *trans_alpha = NULL;
+      }
+
+      if (num_trans != NULL)
+      {
+         *num_trans = info_ptr->num_trans;
+         retval |= PNG_INFO_tRNS;
+      }
+   }
+
+   return (retval);
+}
+#endif
+
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+int PNGAPI
+png_get_unknown_chunks(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_unknown_chunkpp unknowns)
+{
+   if (png_ptr != NULL && info_ptr != NULL && unknowns != NULL)
+   {
+      *unknowns = info_ptr->unknown_chunks;
+      return info_ptr->unknown_chunks_num;
+   }
+
+   return (0);
+}
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+png_byte PNGAPI
+png_get_rgb_to_gray_status (png_const_structrp png_ptr)
+{
+   return (png_byte)(png_ptr ? png_ptr->rgb_to_gray_status : 0);
+}
+#endif
+
+#ifdef PNG_USER_CHUNKS_SUPPORTED
+png_voidp PNGAPI
+png_get_user_chunk_ptr(png_const_structrp png_ptr)
+{
+   return (png_ptr ? png_ptr->user_chunk_ptr : NULL);
+}
+#endif
+
+png_size_t PNGAPI
+png_get_compression_buffer_size(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return 0;
+
+#  ifdef PNG_WRITE_SUPPORTED
+      if (png_ptr->mode & PNG_IS_READ_STRUCT)
+#  endif
+   {
+#     ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+         return png_ptr->IDAT_read_size;
+#     else
+         return PNG_IDAT_READ_SIZE;
+#     endif
+   }
+
+#  ifdef PNG_WRITE_SUPPORTED
+      else
+         return png_ptr->zbuffer_size;
+#  endif
+}
+
+#ifdef PNG_SET_USER_LIMITS_SUPPORTED
+/* These functions were added to libpng 1.2.6 and were enabled
+ * by default in libpng-1.4.0 */
+png_uint_32 PNGAPI
+png_get_user_width_max (png_const_structrp png_ptr)
+{
+   return (png_ptr ? png_ptr->user_width_max : 0);
+}
+
+png_uint_32 PNGAPI
+png_get_user_height_max (png_const_structrp png_ptr)
+{
+   return (png_ptr ? png_ptr->user_height_max : 0);
+}
+
+/* This function was added to libpng 1.4.0 */
+png_uint_32 PNGAPI
+png_get_chunk_cache_max (png_const_structrp png_ptr)
+{
+   return (png_ptr ? png_ptr->user_chunk_cache_max : 0);
+}
+
+/* This function was added to libpng 1.4.1 */
+png_alloc_size_t PNGAPI
+png_get_chunk_malloc_max (png_const_structrp png_ptr)
+{
+   return (png_ptr ? png_ptr->user_chunk_malloc_max : 0);
+}
+#endif /* ?PNG_SET_USER_LIMITS_SUPPORTED */
+
+/* These functions were added to libpng 1.4.0 */
+#ifdef PNG_IO_STATE_SUPPORTED
+png_uint_32 PNGAPI
+png_get_io_state (png_const_structrp png_ptr)
+{
+   return png_ptr->io_state;
+}
+
+png_uint_32 PNGAPI
+png_get_io_chunk_type (png_const_structrp png_ptr)
+{
+   return png_ptr->chunk_name;
+}
+#endif /* ?PNG_IO_STATE_SUPPORTED */
+
+#ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+#  ifdef PNG_GET_PALETTE_MAX_SUPPORTED
+int PNGAPI
+png_get_palette_max(png_const_structp png_ptr, png_const_infop info_ptr)
+{
+   if (png_ptr != NULL && info_ptr != NULL)
+      return png_ptr->num_palette_max;
+
+   return (-1);
+}
+#  endif
+#endif
+
+#endif /* PNG_READ_SUPPORTED || PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pnginfo.h b/ml/dlib/dlib/external/libpng/pnginfo.h
new file mode 100644
index 000000000..26bf26502
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pnginfo.h
@@ -0,0 +1,260 @@
+
+/* pnginfo.h - header file for PNG reference library
+ *
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * Last changed in libpng 1.6.1 [March 28, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+ /* png_info is a structure that holds the information in a PNG file so
+ * that the application can find out the characteristics of the image.
+ * If you are reading the file, this structure will tell you what is
+ * in the PNG file.  If you are writing the file, fill in the information
+ * you want to put into the PNG file, using png_set_*() functions, then
+ * call png_write_info().
+ *
+ * The names chosen should be very close to the PNG specification, so
+ * consult that document for information about the meaning of each field.
+ *
+ * With libpng < 0.95, it was only possible to directly set and read the
+ * the values in the png_info_struct, which meant that the contents and
+ * order of the values had to remain fixed.  With libpng 0.95 and later,
+ * however, there are now functions that abstract the contents of
+ * png_info_struct from the application, so this makes it easier to use
+ * libpng with dynamic libraries, and even makes it possible to use
+ * libraries that don't have all of the libpng ancillary chunk-handing
+ * functionality.  In libpng-1.5.0 this was moved into a separate private
+ * file that is not visible to applications.
+ *
+ * The following members may have allocated storage attached that should be
+ * cleaned up before the structure is discarded: palette, trans, text,
+ * pcal_purpose, pcal_units, pcal_params, hist, iccp_name, iccp_profile,
+ * splt_palettes, scal_unit, row_pointers, and unknowns.   By default, these
+ * are automatically freed when the info structure is deallocated, if they were
+ * allocated internally by libpng.  This behavior can be changed by means
+ * of the png_data_freer() function.
+ *
+ * More allocation details: all the chunk-reading functions that
+ * change these members go through the corresponding png_set_*
+ * functions.  A function to clear these members is available: see
+ * png_free_data().  The png_set_* functions do not depend on being
+ * able to point info structure members to any of the storage they are
+ * passed (they make their own copies), EXCEPT that the png_set_text
+ * functions use the same storage passed to them in the text_ptr or
+ * itxt_ptr structure argument, and the png_set_rows and png_set_unknowns
+ * functions do not make their own copies.
+ */
+#ifndef PNGINFO_H
+#define PNGINFO_H
+
+struct png_info_def
+{
+   /* The following are necessary for every PNG file */
+   png_uint_32 width;  /* width of image in pixels (from IHDR) */
+   png_uint_32 height; /* height of image in pixels (from IHDR) */
+   png_uint_32 valid;  /* valid chunk data (see PNG_INFO_ below) */
+   png_size_t rowbytes; /* bytes needed to hold an untransformed row */
+   png_colorp palette;      /* array of color values (valid & PNG_INFO_PLTE) */
+   png_uint_16 num_palette; /* number of color entries in "palette" (PLTE) */
+   png_uint_16 num_trans;   /* number of transparent palette color (tRNS) */
+   png_byte bit_depth;      /* 1, 2, 4, 8, or 16 bits/channel (from IHDR) */
+   png_byte color_type;     /* see PNG_COLOR_TYPE_ below (from IHDR) */
+   /* The following three should have been named *_method not *_type */
+   png_byte compression_type; /* must be PNG_COMPRESSION_TYPE_BASE (IHDR) */
+   png_byte filter_type;    /* must be PNG_FILTER_TYPE_BASE (from IHDR) */
+   png_byte interlace_type; /* One of PNG_INTERLACE_NONE, PNG_INTERLACE_ADAM7 */
+
+   /* The following are set by png_set_IHDR, called from the application on
+    * write, but the are never actually used by the write code.
+    */
+   png_byte channels;       /* number of data channels per pixel (1, 2, 3, 4) */
+   png_byte pixel_depth;    /* number of bits per pixel */
+   png_byte spare_byte;     /* to align the data, and for future use */
+
+#ifdef PNG_READ_SUPPORTED
+   /* This is never set during write */
+   png_byte signature[8];   /* magic bytes read by libpng from start of file */
+#endif
+
+   /* The rest of the data is optional.  If you are reading, check the
+    * valid field to see if the information in these are valid.  If you
+    * are writing, set the valid field to those chunks you want written,
+    * and initialize the appropriate fields below.
+    */
+
+#if defined(PNG_COLORSPACE_SUPPORTED) || defined(PNG_GAMMA_SUPPORTED)
+   /* png_colorspace only contains 'flags' if neither GAMMA or COLORSPACE are
+    * defined.  When COLORSPACE is switched on all the colorspace-defining
+    * chunks should be enabled, when GAMMA is switched on all the gamma-defining
+    * chunks should be enabled.  If this is not done it becomes possible to read
+    * inconsistent PNG files and assign a probably incorrect interpretation to
+    * the information.  (In other words, by carefully choosing which chunks to
+    * recognize the system configuration can select an interpretation for PNG
+    * files containing ambiguous data and this will result in inconsistent
+    * behavior between different libpng builds!)
+    */
+   png_colorspace colorspace;
+#endif
+
+#ifdef PNG_iCCP_SUPPORTED
+   /* iCCP chunk data. */
+   png_charp iccp_name;     /* profile name */
+   png_bytep iccp_profile;  /* International Color Consortium profile data */
+   png_uint_32 iccp_proflen;  /* ICC profile data length */
+#endif
+
+#ifdef PNG_TEXT_SUPPORTED
+   /* The tEXt, and zTXt chunks contain human-readable textual data in
+    * uncompressed, compressed, and optionally compressed forms, respectively.
+    * The data in "text" is an array of pointers to uncompressed,
+    * null-terminated C strings. Each chunk has a keyword that describes the
+    * textual data contained in that chunk.  Keywords are not required to be
+    * unique, and the text string may be empty.  Any number of text chunks may
+    * be in an image.
+    */
+   int num_text; /* number of comments read or comments to write */
+   int max_text; /* current size of text array */
+   png_textp text; /* array of comments read or comments to write */
+#endif /* PNG_TEXT_SUPPORTED */
+
+#ifdef PNG_tIME_SUPPORTED
+   /* The tIME chunk holds the last time the displayed image data was
+    * modified.  See the png_time struct for the contents of this struct.
+    */
+   png_time mod_time;
+#endif
+
+#ifdef PNG_sBIT_SUPPORTED
+   /* The sBIT chunk specifies the number of significant high-order bits
+    * in the pixel data.  Values are in the range [1, bit_depth], and are
+    * only specified for the channels in the pixel data.  The contents of
+    * the low-order bits is not specified.  Data is valid if
+    * (valid & PNG_INFO_sBIT) is non-zero.
+    */
+   png_color_8 sig_bit; /* significant bits in color channels */
+#endif
+
+#if defined(PNG_tRNS_SUPPORTED) || defined(PNG_READ_EXPAND_SUPPORTED) || \
+defined(PNG_READ_BACKGROUND_SUPPORTED)
+   /* The tRNS chunk supplies transparency data for paletted images and
+    * other image types that don't need a full alpha channel.  There are
+    * "num_trans" transparency values for a paletted image, stored in the
+    * same order as the palette colors, starting from index 0.  Values
+    * for the data are in the range [0, 255], ranging from fully transparent
+    * to fully opaque, respectively.  For non-paletted images, there is a
+    * single color specified that should be treated as fully transparent.
+    * Data is valid if (valid & PNG_INFO_tRNS) is non-zero.
+    */
+   png_bytep trans_alpha;    /* alpha values for paletted image */
+   png_color_16 trans_color; /* transparent color for non-palette image */
+#endif
+
+#if defined(PNG_bKGD_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED)
+   /* The bKGD chunk gives the suggested image background color if the
+    * display program does not have its own background color and the image
+    * is needs to composited onto a background before display.  The colors
+    * in "background" are normally in the same color space/depth as the
+    * pixel data.  Data is valid if (valid & PNG_INFO_bKGD) is non-zero.
+    */
+   png_color_16 background;
+#endif
+
+#ifdef PNG_oFFs_SUPPORTED
+   /* The oFFs chunk gives the offset in "offset_unit_type" units rightwards
+    * and downwards from the top-left corner of the display, page, or other
+    * application-specific co-ordinate space.  See the PNG_OFFSET_ defines
+    * below for the unit types.  Valid if (valid & PNG_INFO_oFFs) non-zero.
+    */
+   png_int_32 x_offset; /* x offset on page */
+   png_int_32 y_offset; /* y offset on page */
+   png_byte offset_unit_type; /* offset units type */
+#endif
+
+#ifdef PNG_pHYs_SUPPORTED
+   /* The pHYs chunk gives the physical pixel density of the image for
+    * display or printing in "phys_unit_type" units (see PNG_RESOLUTION_
+    * defines below).  Data is valid if (valid & PNG_INFO_pHYs) is non-zero.
+    */
+   png_uint_32 x_pixels_per_unit; /* horizontal pixel density */
+   png_uint_32 y_pixels_per_unit; /* vertical pixel density */
+   png_byte phys_unit_type; /* resolution type (see PNG_RESOLUTION_ below) */
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+   /* The hIST chunk contains the relative frequency or importance of the
+    * various palette entries, so that a viewer can intelligently select a
+    * reduced-color palette, if required.  Data is an array of "num_palette"
+    * values in the range [0,65535]. Data valid if (valid & PNG_INFO_hIST)
+    * is non-zero.
+    */
+   png_uint_16p hist;
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+   /* The pCAL chunk describes a transformation between the stored pixel
+    * values and original physical data values used to create the image.
+    * The integer range [0, 2^bit_depth - 1] maps to the floating-point
+    * range given by [pcal_X0, pcal_X1], and are further transformed by a
+    * (possibly non-linear) transformation function given by "pcal_type"
+    * and "pcal_params" into "pcal_units".  Please see the PNG_EQUATION_
+    * defines below, and the PNG-Group's PNG extensions document for a
+    * complete description of the transformations and how they should be
+    * implemented, and for a description of the ASCII parameter strings.
+    * Data values are valid if (valid & PNG_INFO_pCAL) non-zero.
+    */
+   png_charp pcal_purpose;  /* pCAL chunk description string */
+   png_int_32 pcal_X0;      /* minimum value */
+   png_int_32 pcal_X1;      /* maximum value */
+   png_charp pcal_units;    /* Latin-1 string giving physical units */
+   png_charpp pcal_params;  /* ASCII strings containing parameter values */
+   png_byte pcal_type;      /* equation type (see PNG_EQUATION_ below) */
+   png_byte pcal_nparams;   /* number of parameters given in pcal_params */
+#endif
+
+/* New members added in libpng-1.0.6 */
+   png_uint_32 free_me;     /* flags items libpng is responsible for freeing */
+
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+   /* Storage for unknown chunks that the library doesn't recognize. */
+   png_unknown_chunkp unknown_chunks;
+
+   /* The type of this field is limited by the type of 
+    * png_struct::user_chunk_cache_max, else overflow can occur.
+    */
+   int                unknown_chunks_num;
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+   /* Data on sPLT chunks (there may be more than one). */
+   png_sPLT_tp splt_palettes;
+   int         splt_palettes_num; /* Match type returned by png_get API */
+#endif
+
+#ifdef PNG_sCAL_SUPPORTED
+   /* The sCAL chunk describes the actual physical dimensions of the
+    * subject matter of the graphic.  The chunk contains a unit specification
+    * a byte value, and two ASCII strings representing floating-point
+    * values.  The values are width and height corresponsing to one pixel
+    * in the image.  Data values are valid if (valid & PNG_INFO_sCAL) is
+    * non-zero.
+    */
+   png_byte scal_unit;         /* unit of physical scale */
+   png_charp scal_s_width;     /* string containing height */
+   png_charp scal_s_height;    /* string containing width */
+#endif
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+   /* Memory has been allocated if (valid & PNG_ALLOCATED_INFO_ROWS)
+      non-zero */
+   /* Data valid if (valid & PNG_INFO_IDAT) non-zero */
+   png_bytepp row_pointers;        /* the image bits */
+#endif
+
+};
+#endif /* PNGINFO_H */
diff --git a/ml/dlib/dlib/external/libpng/pnglibconf.h b/ml/dlib/dlib/external/libpng/pnglibconf.h
new file mode 100644
index 000000000..6e9a41152
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pnglibconf.h
@@ -0,0 +1,211 @@
+/* libpng 1.6.7 STANDARD API DEFINITION */
+
+/* pnglibconf.h - library build configuration */
+
+/* Libpng version 1.6.7 - November 14, 2013 */
+
+/* Copyright (c) 1998-2013 Glenn Randers-Pehrson */
+
+/* This code is released under the libpng license. */
+/* For conditions of distribution and use, see the disclaimer */
+/* and license in png.h */
+
+/* pnglibconf.h */
+/* Machine generated file: DO NOT EDIT */
+/* Derived from: scripts/pnglibconf.dfa */
+#ifndef PNGLCONF_H
+#define PNGLCONF_H
+/* options */
+#define PNG_16BIT_SUPPORTED
+#define PNG_ALIGNED_MEMORY_SUPPORTED
+/*#undef PNG_ARM_NEON_API_SUPPORTED*/
+/*#undef PNG_ARM_NEON_CHECK_SUPPORTED*/
+#define PNG_BENIGN_ERRORS_SUPPORTED
+#define PNG_BENIGN_READ_ERRORS_SUPPORTED
+/*#undef PNG_BENIGN_WRITE_ERRORS_SUPPORTED*/
+#define PNG_BUILD_GRAYSCALE_PALETTE_SUPPORTED
+#define PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+#define PNG_COLORSPACE_SUPPORTED
+#define PNG_CONSOLE_IO_SUPPORTED
+#define PNG_CONVERT_tIME_SUPPORTED
+#define PNG_EASY_ACCESS_SUPPORTED
+/*#undef PNG_ERROR_NUMBERS_SUPPORTED*/
+#define PNG_ERROR_TEXT_SUPPORTED
+#define PNG_FIXED_POINT_SUPPORTED
+#define PNG_FLOATING_ARITHMETIC_SUPPORTED
+#define PNG_FLOATING_POINT_SUPPORTED
+#define PNG_FORMAT_AFIRST_SUPPORTED
+#define PNG_FORMAT_BGR_SUPPORTED
+#define PNG_GAMMA_SUPPORTED
+#define PNG_GET_PALETTE_MAX_SUPPORTED
+#define PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+#define PNG_INCH_CONVERSIONS_SUPPORTED
+#define PNG_INFO_IMAGE_SUPPORTED
+#define PNG_IO_STATE_SUPPORTED
+#define PNG_MNG_FEATURES_SUPPORTED
+#define PNG_POINTER_INDEXING_SUPPORTED
+#define PNG_PROGRESSIVE_READ_SUPPORTED
+#define PNG_READ_16BIT_SUPPORTED
+#define PNG_READ_ALPHA_MODE_SUPPORTED
+#define PNG_READ_ANCILLARY_CHUNKS_SUPPORTED
+#define PNG_READ_BACKGROUND_SUPPORTED
+#define PNG_READ_BGR_SUPPORTED
+#define PNG_READ_CHECK_FOR_INVALID_INDEX_SUPPORTED
+#define PNG_READ_COMPOSITE_NODIV_SUPPORTED
+#define PNG_READ_COMPRESSED_TEXT_SUPPORTED
+#define PNG_READ_EXPAND_16_SUPPORTED
+#define PNG_READ_EXPAND_SUPPORTED
+#define PNG_READ_FILLER_SUPPORTED
+#define PNG_READ_GAMMA_SUPPORTED
+#define PNG_READ_GET_PALETTE_MAX_SUPPORTED
+#define PNG_READ_GRAY_TO_RGB_SUPPORTED
+#define PNG_READ_INTERLACING_SUPPORTED
+#define PNG_READ_INT_FUNCTIONS_SUPPORTED
+#define PNG_READ_INVERT_ALPHA_SUPPORTED
+#define PNG_READ_INVERT_SUPPORTED
+#define PNG_READ_OPT_PLTE_SUPPORTED
+#define PNG_READ_PACKSWAP_SUPPORTED
+#define PNG_READ_PACK_SUPPORTED
+#define PNG_READ_QUANTIZE_SUPPORTED
+#define PNG_READ_RGB_TO_GRAY_SUPPORTED
+#define PNG_READ_SCALE_16_TO_8_SUPPORTED
+#define PNG_READ_SHIFT_SUPPORTED
+#define PNG_READ_STRIP_16_TO_8_SUPPORTED
+#define PNG_READ_STRIP_ALPHA_SUPPORTED
+#define PNG_READ_SUPPORTED
+#define PNG_READ_SWAP_ALPHA_SUPPORTED
+#define PNG_READ_SWAP_SUPPORTED
+#define PNG_READ_TEXT_SUPPORTED
+#define PNG_READ_TRANSFORMS_SUPPORTED
+#define PNG_READ_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_READ_USER_CHUNKS_SUPPORTED
+#define PNG_READ_USER_TRANSFORM_SUPPORTED
+#define PNG_READ_bKGD_SUPPORTED
+#define PNG_READ_cHRM_SUPPORTED
+#define PNG_READ_gAMA_SUPPORTED
+#define PNG_READ_hIST_SUPPORTED
+#define PNG_READ_iCCP_SUPPORTED
+#define PNG_READ_iTXt_SUPPORTED
+#define PNG_READ_oFFs_SUPPORTED
+#define PNG_READ_pCAL_SUPPORTED
+#define PNG_READ_pHYs_SUPPORTED
+#define PNG_READ_sBIT_SUPPORTED
+#define PNG_READ_sCAL_SUPPORTED
+#define PNG_READ_sPLT_SUPPORTED
+#define PNG_READ_sRGB_SUPPORTED
+#define PNG_READ_tEXt_SUPPORTED
+#define PNG_READ_tIME_SUPPORTED
+#define PNG_READ_tRNS_SUPPORTED
+#define PNG_READ_zTXt_SUPPORTED
+/*#undef PNG_SAFE_LIMITS_SUPPORTED*/
+#define PNG_SAVE_INT_32_SUPPORTED
+#define PNG_SAVE_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_SEQUENTIAL_READ_SUPPORTED
+#define PNG_SETJMP_SUPPORTED
+#define PNG_SET_CHUNK_CACHE_LIMIT_SUPPORTED
+#define PNG_SET_CHUNK_MALLOC_LIMIT_SUPPORTED
+#define PNG_SET_OPTION_SUPPORTED
+#define PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_SET_USER_LIMITS_SUPPORTED
+#define PNG_SIMPLIFIED_READ_AFIRST_SUPPORTED
+#define PNG_SIMPLIFIED_READ_BGR_SUPPORTED
+#define PNG_SIMPLIFIED_READ_SUPPORTED
+#define PNG_SIMPLIFIED_WRITE_AFIRST_SUPPORTED
+#define PNG_SIMPLIFIED_WRITE_BGR_SUPPORTED
+#define PNG_SIMPLIFIED_WRITE_SUPPORTED
+#define PNG_STDIO_SUPPORTED
+#define PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_TEXT_SUPPORTED
+#define PNG_TIME_RFC1123_SUPPORTED
+#define PNG_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_USER_CHUNKS_SUPPORTED
+#define PNG_USER_LIMITS_SUPPORTED
+#define PNG_USER_MEM_SUPPORTED
+#define PNG_USER_TRANSFORM_INFO_SUPPORTED
+#define PNG_USER_TRANSFORM_PTR_SUPPORTED
+#define PNG_WARNINGS_SUPPORTED
+#define PNG_WRITE_16BIT_SUPPORTED
+#define PNG_WRITE_ANCILLARY_CHUNKS_SUPPORTED
+#define PNG_WRITE_BGR_SUPPORTED
+#define PNG_WRITE_CHECK_FOR_INVALID_INDEX_SUPPORTED
+#define PNG_WRITE_COMPRESSED_TEXT_SUPPORTED
+#define PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED
+#define PNG_WRITE_FILLER_SUPPORTED
+#define PNG_WRITE_FILTER_SUPPORTED
+#define PNG_WRITE_FLUSH_SUPPORTED
+#define PNG_WRITE_GET_PALETTE_MAX_SUPPORTED
+#define PNG_WRITE_INTERLACING_SUPPORTED
+#define PNG_WRITE_INT_FUNCTIONS_SUPPORTED
+#define PNG_WRITE_INVERT_ALPHA_SUPPORTED
+#define PNG_WRITE_INVERT_SUPPORTED
+#define PNG_WRITE_OPTIMIZE_CMF_SUPPORTED
+#define PNG_WRITE_PACKSWAP_SUPPORTED
+#define PNG_WRITE_PACK_SUPPORTED
+#define PNG_WRITE_SHIFT_SUPPORTED
+#define PNG_WRITE_SUPPORTED
+#define PNG_WRITE_SWAP_ALPHA_SUPPORTED
+#define PNG_WRITE_SWAP_SUPPORTED
+#define PNG_WRITE_TEXT_SUPPORTED
+#define PNG_WRITE_TRANSFORMS_SUPPORTED
+#define PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED
+#define PNG_WRITE_USER_TRANSFORM_SUPPORTED
+#define PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+#define PNG_WRITE_bKGD_SUPPORTED
+#define PNG_WRITE_cHRM_SUPPORTED
+#define PNG_WRITE_gAMA_SUPPORTED
+#define PNG_WRITE_hIST_SUPPORTED
+#define PNG_WRITE_iCCP_SUPPORTED
+#define PNG_WRITE_iTXt_SUPPORTED
+#define PNG_WRITE_oFFs_SUPPORTED
+#define PNG_WRITE_pCAL_SUPPORTED
+#define PNG_WRITE_pHYs_SUPPORTED
+#define PNG_WRITE_sBIT_SUPPORTED
+#define PNG_WRITE_sCAL_SUPPORTED
+#define PNG_WRITE_sPLT_SUPPORTED
+#define PNG_WRITE_sRGB_SUPPORTED
+#define PNG_WRITE_tEXt_SUPPORTED
+#define PNG_WRITE_tIME_SUPPORTED
+#define PNG_WRITE_tRNS_SUPPORTED
+#define PNG_WRITE_zTXt_SUPPORTED
+#define PNG_bKGD_SUPPORTED
+#define PNG_cHRM_SUPPORTED
+#define PNG_gAMA_SUPPORTED
+#define PNG_hIST_SUPPORTED
+#define PNG_iCCP_SUPPORTED
+#define PNG_iTXt_SUPPORTED
+#define PNG_oFFs_SUPPORTED
+#define PNG_pCAL_SUPPORTED
+#define PNG_pHYs_SUPPORTED
+#define PNG_sBIT_SUPPORTED
+#define PNG_sCAL_SUPPORTED
+#define PNG_sPLT_SUPPORTED
+#define PNG_sRGB_SUPPORTED
+#define PNG_tEXt_SUPPORTED
+#define PNG_tIME_SUPPORTED
+#define PNG_tRNS_SUPPORTED
+#define PNG_zTXt_SUPPORTED
+/* end of options */
+/* settings */
+#define PNG_API_RULE 0
+#define PNG_CALLOC_SUPPORTED
+#define PNG_COST_SHIFT 3
+#define PNG_DEFAULT_READ_MACROS 1
+#define PNG_GAMMA_THRESHOLD_FIXED 5000
+#define PNG_IDAT_READ_SIZE PNG_ZBUF_SIZE
+#define PNG_INFLATE_BUF_SIZE 1024
+#define PNG_MAX_GAMMA_8 11
+#define PNG_QUANTIZE_BLUE_BITS 5
+#define PNG_QUANTIZE_GREEN_BITS 5
+#define PNG_QUANTIZE_RED_BITS 5
+#define PNG_TEXT_Z_DEFAULT_COMPRESSION (-1)
+#define PNG_TEXT_Z_DEFAULT_STRATEGY 0
+#define PNG_WEIGHT_SHIFT 8
+#define PNG_ZBUF_SIZE 8192
+#define PNG_ZLIB_VERNUM 0 /* unknown */
+#define PNG_Z_DEFAULT_COMPRESSION (-1)
+#define PNG_Z_DEFAULT_NOFILTER_STRATEGY 0
+#define PNG_Z_DEFAULT_STRATEGY 1
+#define PNG_sCAL_PRECISION 5
+#define PNG_sRGB_PROFILE_CHECKS 2
+/* end of settings */
+#endif /* PNGLCONF_H */
diff --git a/ml/dlib/dlib/external/libpng/pngmem.c b/ml/dlib/dlib/external/libpng/pngmem.c
new file mode 100644
index 000000000..b9b3efb44
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngmem.c
@@ -0,0 +1,277 @@
+
+/* pngmem.c - stub functions for memory allocation
+ *
+ * Last changed in libpng 1.6.0 [February 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file provides a location for all memory allocation.  Users who
+ * need special memory handling are expected to supply replacement
+ * functions for png_malloc() and png_free(), and to use
+ * png_create_read_struct_2() and png_create_write_struct_2() to
+ * identify the replacement functions.
+ */
+
+#include "pngpriv.h"
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+/* Free a png_struct */
+void /* PRIVATE */
+png_destroy_png_struct(png_structrp png_ptr)
+{
+   if (png_ptr != NULL)
+   {
+      /* png_free might call png_error and may certainly call
+       * png_get_mem_ptr, so fake a temporary png_struct to support this.
+       */
+      png_struct dummy_struct = *png_ptr;
+      memset(png_ptr, 0, (sizeof *png_ptr));
+      png_free(&dummy_struct, png_ptr);
+
+#     ifdef PNG_SETJMP_SUPPORTED
+         /* We may have a jmp_buf left to deallocate. */
+         png_free_jmpbuf(&dummy_struct);
+#     endif
+   }
+}
+
+/* Allocate memory.  For reasonable files, size should never exceed
+ * 64K.  However, zlib may allocate more then 64K if you don't tell
+ * it not to.  See zconf.h and png.h for more information.  zlib does
+ * need to allocate exactly 64K, so whatever you call here must
+ * have the ability to do that.
+ */
+PNG_FUNCTION(png_voidp,PNGAPI
+png_calloc,(png_const_structrp png_ptr, png_alloc_size_t size),PNG_ALLOCATED)
+{
+   png_voidp ret;
+
+   ret = png_malloc(png_ptr, size);
+
+   if (ret != NULL)
+      memset(ret, 0, size);
+
+   return ret;
+}
+
+/* png_malloc_base, an internal function added at libpng 1.6.0, does the work of
+ * allocating memory, taking into account limits and PNG_USER_MEM_SUPPORTED.
+ * Checking and error handling must happen outside this routine; it returns NULL
+ * if the allocation cannot be done (for any reason.)
+ */
+PNG_FUNCTION(png_voidp /* PRIVATE */,
+png_malloc_base,(png_const_structrp png_ptr, png_alloc_size_t size),
+   PNG_ALLOCATED)
+{
+   /* Moved to png_malloc_base from png_malloc_default in 1.6.0; the DOS
+    * allocators have also been removed in 1.6.0, so any 16-bit system now has
+    * to implement a user memory handler.  This checks to be sure it isn't
+    * called with big numbers.
+    */
+#ifdef PNG_USER_MEM_SUPPORTED
+   PNG_UNUSED(png_ptr)
+#endif
+   if (size > 0 && size <= PNG_SIZE_MAX
+#     ifdef PNG_MAX_MALLOC_64K
+         && size <= 65536U
+#     endif
+      )
+   {
+#ifdef PNG_USER_MEM_SUPPORTED
+      if (png_ptr != NULL && png_ptr->malloc_fn != NULL)
+         return png_ptr->malloc_fn(png_constcast(png_structrp,png_ptr), size);
+
+      else
+#endif
+         return malloc((size_t)size); /* checked for truncation above */
+   }
+
+   else
+      return NULL;
+}
+
+/* This is really here only to work round a spurious warning in GCC 4.6 and 4.7
+ * that arises because of the checks in png_realloc_array that are repeated in
+ * png_malloc_array.
+ */
+static png_voidp
+png_malloc_array_checked(png_const_structrp png_ptr, int nelements,
+   size_t element_size)
+{
+   png_alloc_size_t req = nelements; /* known to be > 0 */
+
+   if (req <= PNG_SIZE_MAX/element_size)
+      return png_malloc_base(png_ptr, req * element_size);
+
+   /* The failure case when the request is too large */
+   return NULL;
+}
+
+PNG_FUNCTION(png_voidp /* PRIVATE */,
+png_malloc_array,(png_const_structrp png_ptr, int nelements,
+   size_t element_size),PNG_ALLOCATED)
+{
+   if (nelements <= 0 || element_size == 0)
+      png_error(png_ptr, "internal error: array alloc");
+
+   return png_malloc_array_checked(png_ptr, nelements, element_size);
+}
+
+PNG_FUNCTION(png_voidp /* PRIVATE */,
+png_realloc_array,(png_const_structrp png_ptr, png_const_voidp old_array,
+   int old_elements, int add_elements, size_t element_size),PNG_ALLOCATED)
+{
+   /* These are internal errors: */
+   if (add_elements <= 0 || element_size == 0 || old_elements < 0 ||
+      (old_array == NULL && old_elements > 0))
+      png_error(png_ptr, "internal error: array realloc");
+
+   /* Check for overflow on the elements count (so the caller does not have to
+    * check.)
+    */
+   if (add_elements <= INT_MAX - old_elements)
+   {
+      png_voidp new_array = png_malloc_array_checked(png_ptr,
+         old_elements+add_elements, element_size);
+
+      if (new_array != NULL)
+      {
+         /* Because png_malloc_array worked the size calculations below cannot
+          * overflow.
+          */
+         if (old_elements > 0)
+            memcpy(new_array, old_array, element_size*(unsigned)old_elements);
+
+         memset((char*)new_array + element_size*(unsigned)old_elements, 0,
+            element_size*(unsigned)add_elements);
+
+         return new_array;
+      }
+   }
+
+   return NULL; /* error */
+}
+
+/* Various functions that have different error handling are derived from this.
+ * png_malloc always exists, but if PNG_USER_MEM_SUPPORTED is defined a separate
+ * function png_malloc_default is also provided.
+ */
+PNG_FUNCTION(png_voidp,PNGAPI
+png_malloc,(png_const_structrp png_ptr, png_alloc_size_t size),PNG_ALLOCATED)
+{
+   png_voidp ret;
+
+   if (png_ptr == NULL)
+      return NULL;
+
+   ret = png_malloc_base(png_ptr, size);
+
+   if (ret == NULL)
+       png_error(png_ptr, "Out of memory"); /* 'm' means png_malloc */
+
+   return ret;
+}
+
+#ifdef PNG_USER_MEM_SUPPORTED
+PNG_FUNCTION(png_voidp,PNGAPI
+png_malloc_default,(png_const_structrp png_ptr, png_alloc_size_t size),
+   PNG_ALLOCATED PNG_DEPRECATED)
+{
+   png_voidp ret;
+
+   if (png_ptr == NULL)
+      return NULL;
+
+   /* Passing 'NULL' here bypasses the application provided memory handler. */
+   ret = png_malloc_base(NULL/*use malloc*/, size);
+
+   if (ret == NULL)
+      png_error(png_ptr, "Out of Memory"); /* 'M' means png_malloc_default */
+
+   return ret;
+}
+#endif /* PNG_USER_MEM_SUPPORTED */
+
+/* This function was added at libpng version 1.2.3.  The png_malloc_warn()
+ * function will issue a png_warning and return NULL instead of issuing a
+ * png_error, if it fails to allocate the requested memory.
+ */
+PNG_FUNCTION(png_voidp,PNGAPI
+png_malloc_warn,(png_const_structrp png_ptr, png_alloc_size_t size),
+   PNG_ALLOCATED)
+{
+   if (png_ptr != NULL)
+   {
+      png_voidp ret = png_malloc_base(png_ptr, size);
+
+      if (ret != NULL)
+         return ret;
+
+      png_warning(png_ptr, "Out of memory");
+   }
+
+   return NULL;
+}
+
+/* Free a pointer allocated by png_malloc().  If ptr is NULL, return
+ * without taking any action.
+ */
+void PNGAPI
+png_free(png_const_structrp png_ptr, png_voidp ptr)
+{
+   if (png_ptr == NULL || ptr == NULL)
+      return;
+
+#ifdef PNG_USER_MEM_SUPPORTED
+   if (png_ptr->free_fn != NULL)
+      png_ptr->free_fn(png_constcast(png_structrp,png_ptr), ptr);
+
+   else
+      png_free_default(png_ptr, ptr);
+}
+
+PNG_FUNCTION(void,PNGAPI
+png_free_default,(png_const_structrp png_ptr, png_voidp ptr),PNG_DEPRECATED)
+{
+   if (png_ptr == NULL || ptr == NULL)
+      return;
+#endif /* PNG_USER_MEM_SUPPORTED */
+
+   free(ptr);
+}
+
+#ifdef PNG_USER_MEM_SUPPORTED
+/* This function is called when the application wants to use another method
+ * of allocating and freeing memory.
+ */
+void PNGAPI
+png_set_mem_fn(png_structrp png_ptr, png_voidp mem_ptr, png_malloc_ptr
+  malloc_fn, png_free_ptr free_fn)
+{
+   if (png_ptr != NULL)
+   {
+      png_ptr->mem_ptr = mem_ptr;
+      png_ptr->malloc_fn = malloc_fn;
+      png_ptr->free_fn = free_fn;
+   }
+}
+
+/* This function returns a pointer to the mem_ptr associated with the user
+ * functions.  The application should free any memory associated with this
+ * pointer before png_write_destroy and png_read_destroy are called.
+ */
+png_voidp PNGAPI
+png_get_mem_ptr(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return NULL;
+
+   return png_ptr->mem_ptr;
+}
+#endif /* PNG_USER_MEM_SUPPORTED */
+#endif /* PNG_READ_SUPPORTED || PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngpread.c b/ml/dlib/dlib/external/libpng/pngpread.c
new file mode 100644
index 000000000..0169ecb2c
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngpread.c
@@ -0,0 +1,1291 @@
+
+/* pngpread.c - read a png file in push mode
+ *
+ * Last changed in libpng 1.6.0 [February 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+
+/* Push model modes */
+#define PNG_READ_SIG_MODE   0
+#define PNG_READ_CHUNK_MODE 1
+#define PNG_READ_IDAT_MODE  2
+#define PNG_SKIP_MODE       3
+#define PNG_READ_tEXt_MODE  4
+#define PNG_READ_zTXt_MODE  5
+#define PNG_READ_DONE_MODE  6
+#define PNG_READ_iTXt_MODE  7
+#define PNG_ERROR_MODE      8
+
+void PNGAPI
+png_process_data(png_structrp png_ptr, png_inforp info_ptr,
+    png_bytep buffer, png_size_t buffer_size)
+{
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   png_push_restore_buffer(png_ptr, buffer, buffer_size);
+
+   while (png_ptr->buffer_size)
+   {
+      png_process_some_data(png_ptr, info_ptr);
+   }
+}
+
+png_size_t PNGAPI
+png_process_data_pause(png_structrp png_ptr, int save)
+{
+   if (png_ptr != NULL)
+   {
+      /* It's easiest for the caller if we do the save, then the caller doesn't
+       * have to supply the same data again:
+       */
+      if (save)
+         png_push_save_buffer(png_ptr);
+      else
+      {
+         /* This includes any pending saved bytes: */
+         png_size_t remaining = png_ptr->buffer_size;
+         png_ptr->buffer_size = 0;
+
+         /* So subtract the saved buffer size, unless all the data
+          * is actually 'saved', in which case we just return 0
+          */
+         if (png_ptr->save_buffer_size < remaining)
+            return remaining - png_ptr->save_buffer_size;
+      }
+   }
+
+   return 0;
+}
+
+png_uint_32 PNGAPI
+png_process_data_skip(png_structrp png_ptr)
+{
+   png_uint_32 remaining = 0;
+
+   if (png_ptr != NULL && png_ptr->process_mode == PNG_SKIP_MODE &&
+      png_ptr->skip_length > 0)
+   {
+      /* At the end of png_process_data the buffer size must be 0 (see the loop
+       * above) so we can detect a broken call here:
+       */
+      if (png_ptr->buffer_size != 0)
+         png_error(png_ptr,
+            "png_process_data_skip called inside png_process_data");
+
+      /* If is impossible for there to be a saved buffer at this point -
+       * otherwise we could not be in SKIP mode.  This will also happen if
+       * png_process_skip is called inside png_process_data (but only very
+       * rarely.)
+       */
+      if (png_ptr->save_buffer_size != 0)
+         png_error(png_ptr, "png_process_data_skip called with saved data");
+
+      remaining = png_ptr->skip_length;
+      png_ptr->skip_length = 0;
+      png_ptr->process_mode = PNG_READ_CHUNK_MODE;
+   }
+
+   return remaining;
+}
+
+/* What we do with the incoming data depends on what we were previously
+ * doing before we ran out of data...
+ */
+void /* PRIVATE */
+png_process_some_data(png_structrp png_ptr, png_inforp info_ptr)
+{
+   if (png_ptr == NULL)
+      return;
+
+   switch (png_ptr->process_mode)
+   {
+      case PNG_READ_SIG_MODE:
+      {
+         png_push_read_sig(png_ptr, info_ptr);
+         break;
+      }
+
+      case PNG_READ_CHUNK_MODE:
+      {
+         png_push_read_chunk(png_ptr, info_ptr);
+         break;
+      }
+
+      case PNG_READ_IDAT_MODE:
+      {
+         png_push_read_IDAT(png_ptr);
+         break;
+      }
+
+      case PNG_SKIP_MODE:
+      {
+         png_push_crc_finish(png_ptr);
+         break;
+      }
+
+      default:
+      {
+         png_ptr->buffer_size = 0;
+         break;
+      }
+   }
+}
+
+/* Read any remaining signature bytes from the stream and compare them with
+ * the correct PNG signature.  It is possible that this routine is called
+ * with bytes already read from the signature, either because they have been
+ * checked by the calling application, or because of multiple calls to this
+ * routine.
+ */
+void /* PRIVATE */
+png_push_read_sig(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_size_t num_checked = png_ptr->sig_bytes, /* SAFE, does not exceed 8 */ 
+             num_to_check = 8 - num_checked;
+
+   if (png_ptr->buffer_size < num_to_check)
+   {
+      num_to_check = png_ptr->buffer_size;
+   }
+
+   png_push_fill_buffer(png_ptr, &(info_ptr->signature[num_checked]),
+       num_to_check);
+   png_ptr->sig_bytes = (png_byte)(png_ptr->sig_bytes + num_to_check);
+
+   if (png_sig_cmp(info_ptr->signature, num_checked, num_to_check))
+   {
+      if (num_checked < 4 &&
+          png_sig_cmp(info_ptr->signature, num_checked, num_to_check - 4))
+         png_error(png_ptr, "Not a PNG file");
+
+      else
+         png_error(png_ptr, "PNG file corrupted by ASCII conversion");
+   }
+   else
+   {
+      if (png_ptr->sig_bytes >= 8)
+      {
+         png_ptr->process_mode = PNG_READ_CHUNK_MODE;
+      }
+   }
+}
+
+void /* PRIVATE */
+png_push_read_chunk(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_uint_32 chunk_name;
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+   int keep; /* unknown handling method */
+#endif
+
+   /* First we make sure we have enough data for the 4 byte chunk name
+    * and the 4 byte chunk length before proceeding with decoding the
+    * chunk data.  To fully decode each of these chunks, we also make
+    * sure we have enough data in the buffer for the 4 byte CRC at the
+    * end of every chunk (except IDAT, which is handled separately).
+    */
+   if (!(png_ptr->mode & PNG_HAVE_CHUNK_HEADER))
+   {
+      png_byte chunk_length[4];
+      png_byte chunk_tag[4];
+
+      if (png_ptr->buffer_size < 8)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_push_fill_buffer(png_ptr, chunk_length, 4);
+      png_ptr->push_length = png_get_uint_31(png_ptr, chunk_length);
+      png_reset_crc(png_ptr);
+      png_crc_read(png_ptr, chunk_tag, 4);
+      png_ptr->chunk_name = PNG_CHUNK_FROM_STRING(chunk_tag);
+      png_check_chunk_name(png_ptr, png_ptr->chunk_name);
+      png_ptr->mode |= PNG_HAVE_CHUNK_HEADER;
+   }
+
+   chunk_name = png_ptr->chunk_name;
+
+   if (chunk_name == png_IDAT)
+   {
+      if (png_ptr->mode & PNG_AFTER_IDAT)
+         png_ptr->mode |= PNG_HAVE_CHUNK_AFTER_IDAT;
+
+      /* If we reach an IDAT chunk, this means we have read all of the
+       * header chunks, and we can start reading the image (or if this
+       * is called after the image has been read - we have an error).
+       */
+      if (!(png_ptr->mode & PNG_HAVE_IHDR))
+         png_error(png_ptr, "Missing IHDR before IDAT");
+
+      else if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE &&
+          !(png_ptr->mode & PNG_HAVE_PLTE))
+         png_error(png_ptr, "Missing PLTE before IDAT");
+
+      png_ptr->mode |= PNG_HAVE_IDAT;
+
+      if (!(png_ptr->mode & PNG_HAVE_CHUNK_AFTER_IDAT))
+         if (png_ptr->push_length == 0)
+            return;
+
+      if (png_ptr->mode & PNG_AFTER_IDAT)
+         png_benign_error(png_ptr, "Too many IDATs found");
+   }
+
+   if (chunk_name == png_IHDR)
+   {
+      if (png_ptr->push_length != 13)
+         png_error(png_ptr, "Invalid IHDR length");
+
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_IHDR(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+   else if (chunk_name == png_IEND)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_IEND(png_ptr, info_ptr, png_ptr->push_length);
+
+      png_ptr->process_mode = PNG_READ_DONE_MODE;
+      png_push_have_end(png_ptr, info_ptr);
+   }
+
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+   else if ((keep = png_chunk_unknown_handling(png_ptr, chunk_name)) != 0)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_unknown(png_ptr, info_ptr, png_ptr->push_length, keep);
+
+      if (chunk_name == png_PLTE)
+         png_ptr->mode |= PNG_HAVE_PLTE;
+   }
+
+#endif
+   else if (chunk_name == png_PLTE)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+      png_handle_PLTE(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+   else if (chunk_name == png_IDAT)
+   {
+      png_ptr->idat_size = png_ptr->push_length;
+      png_ptr->process_mode = PNG_READ_IDAT_MODE;
+      png_push_have_info(png_ptr, info_ptr);
+      png_ptr->zstream.avail_out =
+          (uInt) PNG_ROWBYTES(png_ptr->pixel_depth,
+          png_ptr->iwidth) + 1;
+      png_ptr->zstream.next_out = png_ptr->row_buf;
+      return;
+   }
+
+#ifdef PNG_READ_gAMA_SUPPORTED
+   else if (png_ptr->chunk_name == png_gAMA)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_gAMA(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_sBIT_SUPPORTED
+   else if (png_ptr->chunk_name == png_sBIT)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_sBIT(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_cHRM_SUPPORTED
+   else if (png_ptr->chunk_name == png_cHRM)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_cHRM(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_sRGB_SUPPORTED
+   else if (chunk_name == png_sRGB)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_sRGB(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_iCCP_SUPPORTED
+   else if (png_ptr->chunk_name == png_iCCP)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_iCCP(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_sPLT_SUPPORTED
+   else if (chunk_name == png_sPLT)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_sPLT(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_tRNS_SUPPORTED
+   else if (chunk_name == png_tRNS)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_tRNS(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_bKGD_SUPPORTED
+   else if (chunk_name == png_bKGD)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_bKGD(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_hIST_SUPPORTED
+   else if (chunk_name == png_hIST)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_hIST(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_pHYs_SUPPORTED
+   else if (chunk_name == png_pHYs)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_pHYs(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_oFFs_SUPPORTED
+   else if (chunk_name == png_oFFs)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_oFFs(png_ptr, info_ptr, png_ptr->push_length);
+   }
+#endif
+
+#ifdef PNG_READ_pCAL_SUPPORTED
+   else if (chunk_name == png_pCAL)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_pCAL(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_sCAL_SUPPORTED
+   else if (chunk_name == png_sCAL)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_sCAL(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_tIME_SUPPORTED
+   else if (chunk_name == png_tIME)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_tIME(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_tEXt_SUPPORTED
+   else if (chunk_name == png_tEXt)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_tEXt(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_zTXt_SUPPORTED
+   else if (chunk_name == png_zTXt)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_zTXt(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+#ifdef PNG_READ_iTXt_SUPPORTED
+   else if (chunk_name == png_iTXt)
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_handle_iTXt(png_ptr, info_ptr, png_ptr->push_length);
+   }
+
+#endif
+   else
+   {
+      if (png_ptr->push_length + 4 > png_ptr->buffer_size)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+      png_handle_unknown(png_ptr, info_ptr, png_ptr->push_length,
+         PNG_HANDLE_CHUNK_AS_DEFAULT);
+   }
+
+   png_ptr->mode &= ~PNG_HAVE_CHUNK_HEADER;
+}
+
+void /* PRIVATE */
+png_push_crc_skip(png_structrp png_ptr, png_uint_32 skip)
+{
+   png_ptr->process_mode = PNG_SKIP_MODE;
+   png_ptr->skip_length = skip;
+}
+
+void /* PRIVATE */
+png_push_crc_finish(png_structrp png_ptr)
+{
+   if (png_ptr->skip_length && png_ptr->save_buffer_size)
+   {
+      png_size_t save_size = png_ptr->save_buffer_size;
+      png_uint_32 skip_length = png_ptr->skip_length;
+
+      /* We want the smaller of 'skip_length' and 'save_buffer_size', but
+       * they are of different types and we don't know which variable has the
+       * fewest bits.  Carefully select the smaller and cast it to the type of
+       * the larger - this cannot overflow.  Do not cast in the following test
+       * - it will break on either 16 or 64 bit platforms.
+       */
+      if (skip_length < save_size)
+         save_size = (png_size_t)skip_length;
+
+      else
+         skip_length = (png_uint_32)save_size;
+
+      png_calculate_crc(png_ptr, png_ptr->save_buffer_ptr, save_size);
+
+      png_ptr->skip_length -= skip_length;
+      png_ptr->buffer_size -= save_size;
+      png_ptr->save_buffer_size -= save_size;
+      png_ptr->save_buffer_ptr += save_size;
+   }
+   if (png_ptr->skip_length && png_ptr->current_buffer_size)
+   {
+      png_size_t save_size = png_ptr->current_buffer_size;
+      png_uint_32 skip_length = png_ptr->skip_length;
+
+      /* We want the smaller of 'skip_length' and 'current_buffer_size', here,
+       * the same problem exists as above and the same solution.
+       */
+      if (skip_length < save_size)
+         save_size = (png_size_t)skip_length;
+
+      else
+         skip_length = (png_uint_32)save_size;
+
+      png_calculate_crc(png_ptr, png_ptr->current_buffer_ptr, save_size);
+
+      png_ptr->skip_length -= skip_length;
+      png_ptr->buffer_size -= save_size;
+      png_ptr->current_buffer_size -= save_size;
+      png_ptr->current_buffer_ptr += save_size;
+   }
+   if (!png_ptr->skip_length)
+   {
+      if (png_ptr->buffer_size < 4)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_crc_finish(png_ptr, 0);
+      png_ptr->process_mode = PNG_READ_CHUNK_MODE;
+   }
+}
+
+void PNGCBAPI
+png_push_fill_buffer(png_structp png_ptr, png_bytep buffer, png_size_t length)
+{
+   png_bytep ptr;
+
+   if (png_ptr == NULL)
+      return;
+
+   ptr = buffer;
+   if (png_ptr->save_buffer_size)
+   {
+      png_size_t save_size;
+
+      if (length < png_ptr->save_buffer_size)
+         save_size = length;
+
+      else
+         save_size = png_ptr->save_buffer_size;
+
+      memcpy(ptr, png_ptr->save_buffer_ptr, save_size);
+      length -= save_size;
+      ptr += save_size;
+      png_ptr->buffer_size -= save_size;
+      png_ptr->save_buffer_size -= save_size;
+      png_ptr->save_buffer_ptr += save_size;
+   }
+   if (length && png_ptr->current_buffer_size)
+   {
+      png_size_t save_size;
+
+      if (length < png_ptr->current_buffer_size)
+         save_size = length;
+
+      else
+         save_size = png_ptr->current_buffer_size;
+
+      memcpy(ptr, png_ptr->current_buffer_ptr, save_size);
+      png_ptr->buffer_size -= save_size;
+      png_ptr->current_buffer_size -= save_size;
+      png_ptr->current_buffer_ptr += save_size;
+   }
+}
+
+void /* PRIVATE */
+png_push_save_buffer(png_structrp png_ptr)
+{
+   if (png_ptr->save_buffer_size)
+   {
+      if (png_ptr->save_buffer_ptr != png_ptr->save_buffer)
+      {
+         png_size_t i, istop;
+         png_bytep sp;
+         png_bytep dp;
+
+         istop = png_ptr->save_buffer_size;
+         for (i = 0, sp = png_ptr->save_buffer_ptr, dp = png_ptr->save_buffer;
+             i < istop; i++, sp++, dp++)
+         {
+            *dp = *sp;
+         }
+      }
+   }
+   if (png_ptr->save_buffer_size + png_ptr->current_buffer_size >
+       png_ptr->save_buffer_max)
+   {
+      png_size_t new_max;
+      png_bytep old_buffer;
+
+      if (png_ptr->save_buffer_size > PNG_SIZE_MAX -
+          (png_ptr->current_buffer_size + 256))
+      {
+         png_error(png_ptr, "Potential overflow of save_buffer");
+      }
+
+      new_max = png_ptr->save_buffer_size + png_ptr->current_buffer_size + 256;
+      old_buffer = png_ptr->save_buffer;
+      png_ptr->save_buffer = (png_bytep)png_malloc_warn(png_ptr,
+          (png_size_t)new_max);
+
+      if (png_ptr->save_buffer == NULL)
+      {
+         png_free(png_ptr, old_buffer);
+         png_error(png_ptr, "Insufficient memory for save_buffer");
+      }
+
+      memcpy(png_ptr->save_buffer, old_buffer, png_ptr->save_buffer_size);
+      png_free(png_ptr, old_buffer);
+      png_ptr->save_buffer_max = new_max;
+   }
+   if (png_ptr->current_buffer_size)
+   {
+      memcpy(png_ptr->save_buffer + png_ptr->save_buffer_size,
+         png_ptr->current_buffer_ptr, png_ptr->current_buffer_size);
+      png_ptr->save_buffer_size += png_ptr->current_buffer_size;
+      png_ptr->current_buffer_size = 0;
+   }
+   png_ptr->save_buffer_ptr = png_ptr->save_buffer;
+   png_ptr->buffer_size = 0;
+}
+
+void /* PRIVATE */
+png_push_restore_buffer(png_structrp png_ptr, png_bytep buffer,
+   png_size_t buffer_length)
+{
+   png_ptr->current_buffer = buffer;
+   png_ptr->current_buffer_size = buffer_length;
+   png_ptr->buffer_size = buffer_length + png_ptr->save_buffer_size;
+   png_ptr->current_buffer_ptr = png_ptr->current_buffer;
+}
+
+void /* PRIVATE */
+png_push_read_IDAT(png_structrp png_ptr)
+{
+   if (!(png_ptr->mode & PNG_HAVE_CHUNK_HEADER))
+   {
+      png_byte chunk_length[4];
+      png_byte chunk_tag[4];
+
+      /* TODO: this code can be commoned up with the same code in push_read */
+      if (png_ptr->buffer_size < 8)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_push_fill_buffer(png_ptr, chunk_length, 4);
+      png_ptr->push_length = png_get_uint_31(png_ptr, chunk_length);
+      png_reset_crc(png_ptr);
+      png_crc_read(png_ptr, chunk_tag, 4);
+      png_ptr->chunk_name = PNG_CHUNK_FROM_STRING(chunk_tag);
+      png_ptr->mode |= PNG_HAVE_CHUNK_HEADER;
+
+      if (png_ptr->chunk_name != png_IDAT)
+      {
+         png_ptr->process_mode = PNG_READ_CHUNK_MODE;
+
+         if (!(png_ptr->flags & PNG_FLAG_ZSTREAM_ENDED))
+            png_error(png_ptr, "Not enough compressed data");
+
+         return;
+      }
+
+      png_ptr->idat_size = png_ptr->push_length;
+   }
+
+   if (png_ptr->idat_size && png_ptr->save_buffer_size)
+   {
+      png_size_t save_size = png_ptr->save_buffer_size;
+      png_uint_32 idat_size = png_ptr->idat_size;
+
+      /* We want the smaller of 'idat_size' and 'current_buffer_size', but they
+       * are of different types and we don't know which variable has the fewest
+       * bits.  Carefully select the smaller and cast it to the type of the
+       * larger - this cannot overflow.  Do not cast in the following test - it
+       * will break on either 16 or 64 bit platforms.
+       */
+      if (idat_size < save_size)
+         save_size = (png_size_t)idat_size;
+
+      else
+         idat_size = (png_uint_32)save_size;
+
+      png_calculate_crc(png_ptr, png_ptr->save_buffer_ptr, save_size);
+
+      png_process_IDAT_data(png_ptr, png_ptr->save_buffer_ptr, save_size);
+
+      png_ptr->idat_size -= idat_size;
+      png_ptr->buffer_size -= save_size;
+      png_ptr->save_buffer_size -= save_size;
+      png_ptr->save_buffer_ptr += save_size;
+   }
+
+   if (png_ptr->idat_size && png_ptr->current_buffer_size)
+   {
+      png_size_t save_size = png_ptr->current_buffer_size;
+      png_uint_32 idat_size = png_ptr->idat_size;
+
+      /* We want the smaller of 'idat_size' and 'current_buffer_size', but they
+       * are of different types and we don't know which variable has the fewest
+       * bits.  Carefully select the smaller and cast it to the type of the
+       * larger - this cannot overflow.
+       */
+      if (idat_size < save_size)
+         save_size = (png_size_t)idat_size;
+
+      else
+         idat_size = (png_uint_32)save_size;
+
+      png_calculate_crc(png_ptr, png_ptr->current_buffer_ptr, save_size);
+
+      png_process_IDAT_data(png_ptr, png_ptr->current_buffer_ptr, save_size);
+
+      png_ptr->idat_size -= idat_size;
+      png_ptr->buffer_size -= save_size;
+      png_ptr->current_buffer_size -= save_size;
+      png_ptr->current_buffer_ptr += save_size;
+   }
+   if (!png_ptr->idat_size)
+   {
+      if (png_ptr->buffer_size < 4)
+      {
+         png_push_save_buffer(png_ptr);
+         return;
+      }
+
+      png_crc_finish(png_ptr, 0);
+      png_ptr->mode &= ~PNG_HAVE_CHUNK_HEADER;
+      png_ptr->mode |= PNG_AFTER_IDAT;
+      png_ptr->zowner = 0;
+   }
+}
+
+void /* PRIVATE */
+png_process_IDAT_data(png_structrp png_ptr, png_bytep buffer,
+   png_size_t buffer_length)
+{
+   /* The caller checks for a non-zero buffer length. */
+   if (!(buffer_length > 0) || buffer == NULL)
+      png_error(png_ptr, "No IDAT data (internal error)");
+
+   /* This routine must process all the data it has been given
+    * before returning, calling the row callback as required to
+    * handle the uncompressed results.
+    */
+   png_ptr->zstream.next_in = buffer;
+   /* TODO: WARNING: TRUNCATION ERROR: DANGER WILL ROBINSON: */
+   png_ptr->zstream.avail_in = (uInt)buffer_length;
+
+   /* Keep going until the decompressed data is all processed
+    * or the stream marked as finished.
+    */
+   while (png_ptr->zstream.avail_in > 0 &&
+      !(png_ptr->flags & PNG_FLAG_ZSTREAM_ENDED))
+   {
+      int ret;
+
+      /* We have data for zlib, but we must check that zlib
+       * has someplace to put the results.  It doesn't matter
+       * if we don't expect any results -- it may be the input
+       * data is just the LZ end code.
+       */
+      if (!(png_ptr->zstream.avail_out > 0))
+      {
+         /* TODO: WARNING: TRUNCATION ERROR: DANGER WILL ROBINSON: */
+         png_ptr->zstream.avail_out = (uInt)(PNG_ROWBYTES(png_ptr->pixel_depth,
+             png_ptr->iwidth) + 1);
+
+         png_ptr->zstream.next_out = png_ptr->row_buf;
+      }
+
+      /* Using Z_SYNC_FLUSH here means that an unterminated
+       * LZ stream (a stream with a missing end code) can still
+       * be handled, otherwise (Z_NO_FLUSH) a future zlib
+       * implementation might defer output and therefore
+       * change the current behavior (see comments in inflate.c
+       * for why this doesn't happen at present with zlib 1.2.5).
+       */
+      ret = inflate(&png_ptr->zstream, Z_SYNC_FLUSH);
+
+      /* Check for any failure before proceeding. */
+      if (ret != Z_OK && ret != Z_STREAM_END)
+      {
+         /* Terminate the decompression. */
+         png_ptr->flags |= PNG_FLAG_ZSTREAM_ENDED;
+         png_ptr->zowner = 0;
+
+         /* This may be a truncated stream (missing or
+          * damaged end code).  Treat that as a warning.
+          */
+         if (png_ptr->row_number >= png_ptr->num_rows ||
+             png_ptr->pass > 6)
+            png_warning(png_ptr, "Truncated compressed data in IDAT");
+
+         else
+            png_error(png_ptr, "Decompression error in IDAT");
+
+         /* Skip the check on unprocessed input */
+         return;
+      }
+
+      /* Did inflate output any data? */
+      if (png_ptr->zstream.next_out != png_ptr->row_buf)
+      {
+         /* Is this unexpected data after the last row?
+          * If it is, artificially terminate the LZ output
+          * here.
+          */
+         if (png_ptr->row_number >= png_ptr->num_rows ||
+             png_ptr->pass > 6)
+         {
+            /* Extra data. */
+            png_warning(png_ptr, "Extra compressed data in IDAT");
+            png_ptr->flags |= PNG_FLAG_ZSTREAM_ENDED;
+            png_ptr->zowner = 0;
+
+            /* Do no more processing; skip the unprocessed
+             * input check below.
+             */
+            return;
+         }
+
+         /* Do we have a complete row? */
+         if (png_ptr->zstream.avail_out == 0)
+            png_push_process_row(png_ptr);
+      }
+
+      /* And check for the end of the stream. */
+      if (ret == Z_STREAM_END)
+         png_ptr->flags |= PNG_FLAG_ZSTREAM_ENDED;
+   }
+
+   /* All the data should have been processed, if anything
+    * is left at this point we have bytes of IDAT data
+    * after the zlib end code.
+    */
+   if (png_ptr->zstream.avail_in > 0)
+      png_warning(png_ptr, "Extra compression data in IDAT");
+}
+
+void /* PRIVATE */
+png_push_process_row(png_structrp png_ptr)
+{
+   /* 1.5.6: row_info moved out of png_struct to a local here. */
+   png_row_info row_info;
+
+   row_info.width = png_ptr->iwidth; /* NOTE: width of current interlaced row */
+   row_info.color_type = png_ptr->color_type;
+   row_info.bit_depth = png_ptr->bit_depth;
+   row_info.channels = png_ptr->channels;
+   row_info.pixel_depth = png_ptr->pixel_depth;
+   row_info.rowbytes = PNG_ROWBYTES(row_info.pixel_depth, row_info.width);
+
+   if (png_ptr->row_buf[0] > PNG_FILTER_VALUE_NONE)
+   {
+      if (png_ptr->row_buf[0] < PNG_FILTER_VALUE_LAST)
+         png_read_filter_row(png_ptr, &row_info, png_ptr->row_buf + 1,
+            png_ptr->prev_row + 1, png_ptr->row_buf[0]);
+      else
+         png_error(png_ptr, "bad adaptive filter value");
+   }
+
+   /* libpng 1.5.6: the following line was copying png_ptr->rowbytes before
+    * 1.5.6, while the buffer really is this big in current versions of libpng
+    * it may not be in the future, so this was changed just to copy the
+    * interlaced row count:
+    */
+   memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations)
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   /* The transformed pixel depth should match the depth now in row_info. */
+   if (png_ptr->transformed_pixel_depth == 0)
+   {
+      png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+      if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+         png_error(png_ptr, "progressive row overflow");
+   }
+
+   else if (png_ptr->transformed_pixel_depth != row_info.pixel_depth)
+      png_error(png_ptr, "internal progressive row size calculation error");
+
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* Blow up interlaced rows to full size */
+   if (png_ptr->interlaced && (png_ptr->transformations & PNG_INTERLACE))
+   {
+      if (png_ptr->pass < 6)
+         png_do_read_interlace(&row_info, png_ptr->row_buf + 1, png_ptr->pass,
+            png_ptr->transformations);
+
+    switch (png_ptr->pass)
+    {
+         case 0:
+         {
+            int i;
+            for (i = 0; i < 8 && png_ptr->pass == 0; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr); /* Updates png_ptr->pass */
+            }
+
+            if (png_ptr->pass == 2) /* Pass 1 might be empty */
+            {
+               for (i = 0; i < 4 && png_ptr->pass == 2; i++)
+               {
+                  png_push_have_row(png_ptr, NULL);
+                  png_read_push_finish_row(png_ptr);
+               }
+            }
+
+            if (png_ptr->pass == 4 && png_ptr->height <= 4)
+            {
+               for (i = 0; i < 2 && png_ptr->pass == 4; i++)
+               {
+                  png_push_have_row(png_ptr, NULL);
+                  png_read_push_finish_row(png_ptr);
+               }
+            }
+
+            if (png_ptr->pass == 6 && png_ptr->height <= 4)
+            {
+                png_push_have_row(png_ptr, NULL);
+                png_read_push_finish_row(png_ptr);
+            }
+
+            break;
+         }
+
+         case 1:
+         {
+            int i;
+            for (i = 0; i < 8 && png_ptr->pass == 1; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            if (png_ptr->pass == 2) /* Skip top 4 generated rows */
+            {
+               for (i = 0; i < 4 && png_ptr->pass == 2; i++)
+               {
+                  png_push_have_row(png_ptr, NULL);
+                  png_read_push_finish_row(png_ptr);
+               }
+            }
+
+            break;
+         }
+
+         case 2:
+         {
+            int i;
+
+            for (i = 0; i < 4 && png_ptr->pass == 2; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            for (i = 0; i < 4 && png_ptr->pass == 2; i++)
+            {
+               png_push_have_row(png_ptr, NULL);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            if (png_ptr->pass == 4) /* Pass 3 might be empty */
+            {
+               for (i = 0; i < 2 && png_ptr->pass == 4; i++)
+               {
+                  png_push_have_row(png_ptr, NULL);
+                  png_read_push_finish_row(png_ptr);
+               }
+            }
+
+            break;
+         }
+
+         case 3:
+         {
+            int i;
+
+            for (i = 0; i < 4 && png_ptr->pass == 3; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            if (png_ptr->pass == 4) /* Skip top two generated rows */
+            {
+               for (i = 0; i < 2 && png_ptr->pass == 4; i++)
+               {
+                  png_push_have_row(png_ptr, NULL);
+                  png_read_push_finish_row(png_ptr);
+               }
+            }
+
+            break;
+         }
+
+         case 4:
+         {
+            int i;
+
+            for (i = 0; i < 2 && png_ptr->pass == 4; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            for (i = 0; i < 2 && png_ptr->pass == 4; i++)
+            {
+               png_push_have_row(png_ptr, NULL);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            if (png_ptr->pass == 6) /* Pass 5 might be empty */
+            {
+               png_push_have_row(png_ptr, NULL);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            break;
+         }
+
+         case 5:
+         {
+            int i;
+
+            for (i = 0; i < 2 && png_ptr->pass == 5; i++)
+            {
+               png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            if (png_ptr->pass == 6) /* Skip top generated row */
+            {
+               png_push_have_row(png_ptr, NULL);
+               png_read_push_finish_row(png_ptr);
+            }
+
+            break;
+         }
+
+         default:
+         case 6:
+         {
+            png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+            png_read_push_finish_row(png_ptr);
+
+            if (png_ptr->pass != 6)
+               break;
+
+            png_push_have_row(png_ptr, NULL);
+            png_read_push_finish_row(png_ptr);
+         }
+      }
+   }
+   else
+#endif
+   {
+      png_push_have_row(png_ptr, png_ptr->row_buf + 1);
+      png_read_push_finish_row(png_ptr);
+   }
+}
+
+void /* PRIVATE */
+png_read_push_finish_row(png_structrp png_ptr)
+{
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte png_pass_inc[] = {8, 8, 4, 4, 2, 2, 1};
+
+   /* Start of interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_ystart[] = {0, 0, 4, 0, 2, 0, 1};
+
+   /* Offset to next interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_yinc[] = {8, 8, 8, 4, 4, 2, 2};
+
+   /* Height of interlace block.  This is not currently used - if you need
+    * it, uncomment it here and in png.h
+   static PNG_CONST png_byte png_pass_height[] = {8, 8, 4, 4, 2, 2, 1};
+   */
+#endif
+
+   png_ptr->row_number++;
+   if (png_ptr->row_number < png_ptr->num_rows)
+      return;
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   if (png_ptr->interlaced)
+   {
+      png_ptr->row_number = 0;
+      memset(png_ptr->prev_row, 0, png_ptr->rowbytes + 1);
+
+      do
+      {
+         png_ptr->pass++;
+         if ((png_ptr->pass == 1 && png_ptr->width < 5) ||
+             (png_ptr->pass == 3 && png_ptr->width < 3) ||
+             (png_ptr->pass == 5 && png_ptr->width < 2))
+            png_ptr->pass++;
+
+         if (png_ptr->pass > 7)
+            png_ptr->pass--;
+
+         if (png_ptr->pass >= 7)
+            break;
+
+         png_ptr->iwidth = (png_ptr->width +
+             png_pass_inc[png_ptr->pass] - 1 -
+             png_pass_start[png_ptr->pass]) /
+             png_pass_inc[png_ptr->pass];
+
+         if (png_ptr->transformations & PNG_INTERLACE)
+            break;
+
+         png_ptr->num_rows = (png_ptr->height +
+             png_pass_yinc[png_ptr->pass] - 1 -
+             png_pass_ystart[png_ptr->pass]) /
+             png_pass_yinc[png_ptr->pass];
+
+      } while (png_ptr->iwidth == 0 || png_ptr->num_rows == 0);
+   }
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+}
+
+void /* PRIVATE */
+png_push_have_info(png_structrp png_ptr, png_inforp info_ptr)
+{
+   if (png_ptr->info_fn != NULL)
+      (*(png_ptr->info_fn))(png_ptr, info_ptr);
+}
+
+void /* PRIVATE */
+png_push_have_end(png_structrp png_ptr, png_inforp info_ptr)
+{
+   if (png_ptr->end_fn != NULL)
+      (*(png_ptr->end_fn))(png_ptr, info_ptr);
+}
+
+void /* PRIVATE */
+png_push_have_row(png_structrp png_ptr, png_bytep row)
+{
+   if (png_ptr->row_fn != NULL)
+      (*(png_ptr->row_fn))(png_ptr, row, png_ptr->row_number,
+         (int)png_ptr->pass);
+}
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+void PNGAPI
+png_progressive_combine_row(png_const_structrp png_ptr, png_bytep old_row,
+    png_const_bytep new_row)
+{
+   if (png_ptr == NULL)
+      return;
+
+   /* new_row is a flag here - if it is NULL then the app callback was called
+    * from an empty row (see the calls to png_struct::row_fn below), otherwise
+    * it must be png_ptr->row_buf+1
+    */
+   if (new_row != NULL)
+      png_combine_row(png_ptr, old_row, 1/*display*/);
+}
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+
+void PNGAPI
+png_set_progressive_read_fn(png_structrp png_ptr, png_voidp progressive_ptr,
+    png_progressive_info_ptr info_fn, png_progressive_row_ptr row_fn,
+    png_progressive_end_ptr end_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->info_fn = info_fn;
+   png_ptr->row_fn = row_fn;
+   png_ptr->end_fn = end_fn;
+
+   png_set_read_fn(png_ptr, progressive_ptr, png_push_fill_buffer);
+}
+
+png_voidp PNGAPI
+png_get_progressive_ptr(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return (NULL);
+
+   return png_ptr->io_ptr;
+}
+#endif /* PNG_PROGRESSIVE_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngpriv.h b/ml/dlib/dlib/external/libpng/pngpriv.h
new file mode 100644
index 000000000..43b2ef4c5
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngpriv.h
@@ -0,0 +1,2047 @@
+
+/* pngpriv.h - private declarations for use inside libpng
+ *
+ * For conditions of distribution and use, see copyright notice in png.h
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * Last changed in libpng 1.6.7 [November 14, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* The symbols declared in this file (including the functions declared
+ * as extern) are PRIVATE.  They are not part of the libpng public
+ * interface, and are not recommended for use by regular applications.
+ * Some of them may become public in the future; others may stay private,
+ * change in an incompatible way, or even disappear.
+ * Although the libpng users are not forbidden to include this header,
+ * they should be well aware of the issues that may arise from doing so.
+ */
+
+#ifndef PNGPRIV_H
+#define PNGPRIV_H
+
+/* Feature Test Macros.  The following are defined here to ensure that correctly
+ * implemented libraries reveal the APIs libpng needs to build and hide those
+ * that are not needed and potentially damaging to the compilation.
+ *
+ * Feature Test Macros must be defined before any system header is included (see
+ * POSIX 1003.1 2.8.2 "POSIX Symbols."
+ *
+ * These macros only have an effect if the operating system supports either
+ * POSIX 1003.1 or C99, or both.  On other operating systems (particularly
+ * Windows/Visual Studio) there is no effect; the OS specific tests below are
+ * still required (as of 2011-05-02.)
+ */
+#define _POSIX_SOURCE 1 /* Just the POSIX 1003.1 and C89 APIs */
+
+#ifndef PNG_VERSION_INFO_ONLY
+/* Standard library headers not required by png.h: */
+#  include <stdlib.h>
+#  include <string.h>
+#endif
+
+#define PNGLIB_BUILD /*libpng is being built, not used*/
+
+/* If HAVE_CONFIG_H is defined during the build then the build system must
+ * provide an appropriate "config.h" file on the include path.  The header file
+ * must provide definitions as required below (search for "HAVE_CONFIG_H");
+ * see configure.ac for more details of the requirements.  The macro
+ * "PNG_NO_CONFIG_H" is provided for maintainers to test for dependencies on
+ * 'configure'; define this macro to prevent the configure build including the
+ * configure generated config.h.  Libpng is expected to compile without *any*
+ * special build system support on a reasonably ANSI-C compliant system.
+ */
+#if defined(HAVE_CONFIG_H) && !defined(PNG_NO_CONFIG_H)
+#  include <config.h>
+
+   /* Pick up the definition of 'restrict' from config.h if it was read: */
+#  define PNG_RESTRICT restrict
+#endif
+
+/* To support symbol prefixing it is necessary to know *before* including png.h
+ * whether the fixed point (and maybe other) APIs are exported, because if they
+ * are not internal definitions may be required.  This is handled below just
+ * before png.h is included, but load the configuration now if it is available.
+ */
+#ifndef PNGLCONF_H
+#  include "pnglibconf.h"
+#endif
+
+/* Local renames may change non-exported API functions from png.h */
+#if defined(PNG_PREFIX) && !defined(PNGPREFIX_H)
+#  include "pngprefix.h"
+#endif
+
+#ifdef PNG_USER_CONFIG
+#  include "pngusr.h"
+   /* These should have been defined in pngusr.h */
+#  ifndef PNG_USER_PRIVATEBUILD
+#    define PNG_USER_PRIVATEBUILD "Custom libpng build"
+#  endif
+#  ifndef PNG_USER_DLLFNAME_POSTFIX
+#    define PNG_USER_DLLFNAME_POSTFIX "Cb"
+#  endif
+#endif
+
+/* Compile time options.
+ * =====================
+ * In a multi-arch build the compiler may compile the code several times for the
+ * same object module, producing different binaries for different architectures.
+ * When this happens configure-time setting of the target host options cannot be
+ * done and this interferes with the handling of the ARM NEON optimizations, and
+ * possibly other similar optimizations.  Put additional tests here; in general
+ * this is needed when the same option can be changed at both compile time and
+ * run time depending on the target OS (i.e. iOS vs Android.)
+ *
+ * NOTE: symbol prefixing does not pass $(CFLAGS) to the preprocessor, because
+ * this is not possible with certain compilers (Oracle SUN OS CC), as a result
+ * it is necessary to ensure that all extern functions that *might* be used
+ * regardless of $(CFLAGS) get declared in this file.  The test on __ARM_NEON__
+ * below is one example of this behavior because it is controlled by the
+ * presence or not of -mfpu=neon on the GCC command line, it is possible to do
+ * this in $(CC), e.g. "CC=gcc -mfpu=neon", but people who build libpng rarely
+ * do this.
+ */
+#ifndef PNG_ARM_NEON_OPT
+   /* ARM NEON optimizations are being controlled by the compiler settings,
+    * typically the target FPU.  If the FPU has been set to NEON (-mfpu=neon
+    * with GCC) then the compiler will define __ARM_NEON__ and we can rely
+    * unconditionally on NEON instructions not crashing, otherwise we must
+    * disable use of NEON instructions:
+    */
+#  ifdef __ARM_NEON__
+#     define PNG_ARM_NEON_OPT 2
+#  else
+#     define PNG_ARM_NEON_OPT 0
+#  endif
+#endif
+
+#if PNG_ARM_NEON_OPT > 0
+   /* NEON optimizations are to be at least considered by libpng, so enable the
+    * callbacks to do this.
+    */
+#  define PNG_FILTER_OPTIMIZATIONS png_init_filter_functions_neon
+
+   /* By default the 'intrinsics' code in arm/filter_neon_intrinsics.c is used
+    * if possible - if __ARM_NEON__ is set and the compiler version is not known
+    * to be broken.  This is control by PNG_ARM_NEON_IMPLEMENTATION which can
+    * be:
+    *
+    *    1  The intrinsics code (the default with __ARM_NEON__)
+    *    2  The hand coded assembler (the default without __ARM_NEON__)
+    *
+    * It is possible to set PNG_ARM_NEON_IMPLEMENTATION in CPPFLAGS, however
+    * this is *NOT* supported and may cease to work even after a minor revision
+    * to libpng.  It *is* valid to do this for testing purposes, e.g. speed
+    * testing or a new compiler, but the results should be communicated to the
+    * libpng implementation list for incorporation in the next minor release.
+    */
+#  ifndef PNG_ARM_NEON_IMPLEMENTATION
+#     ifdef __ARM_NEON__
+#        if defined(__clang__)
+            /* At present it is unknown by the libpng developers which versions
+             * of clang support the intrinsics, however some or perhaps all
+             * versions do not work with the assembler so this may be
+             * irrelevant, so just use the default (do nothing here.)
+             */
+#        elif defined(__GNUC__)
+            /* GCC 4.5.4 NEON support is known to be broken.  4.6.3 is known to
+             * work, so if this *is* GCC, or G++, look for a version >4.5
+             */
+#           if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 6)
+#              define PNG_ARM_NEON_IMPLEMENTATION 2
+#           endif /* no GNUC support */
+#        endif /* __GNUC__ */
+#     else /* !defined __ARM_NEON__ */
+         /* The 'intrinsics' code simply won't compile without this -mfpu=neon:
+          */
+#        define PNG_ARM_NEON_IMPLEMENTATION 2
+#     endif /* __ARM_NEON__ */
+#  endif /* !defined PNG_ARM_NEON_IMPLEMENTATION */
+
+#  ifndef PNG_ARM_NEON_IMPLEMENTATION
+      /* Use the intrinsics code by default. */
+#     define PNG_ARM_NEON_IMPLEMENTATION 1
+#  endif
+#endif /* PNG_ARM_NEON_OPT > 0 */
+
+/* Is this a build of a DLL where compilation of the object modules requires
+ * different preprocessor settings to those required for a simple library?  If
+ * so PNG_BUILD_DLL must be set.
+ *
+ * If libpng is used inside a DLL but that DLL does not export the libpng APIs
+ * PNG_BUILD_DLL must not be set.  To avoid the code below kicking in build a
+ * static library of libpng then link the DLL against that.
+ */
+#ifndef PNG_BUILD_DLL
+#  ifdef DLL_EXPORT
+      /* This is set by libtool when files are compiled for a DLL; libtool
+       * always compiles twice, even on systems where it isn't necessary.  Set
+       * PNG_BUILD_DLL in case it is necessary:
+       */
+#     define PNG_BUILD_DLL
+#  else
+#     ifdef _WINDLL
+         /* This is set by the Microsoft Visual Studio IDE in projects that
+          * build a DLL.  It can't easily be removed from those projects (it
+          * isn't visible in the Visual Studio UI) so it is a fairly reliable
+          * indication that PNG_IMPEXP needs to be set to the DLL export
+          * attributes.
+          */
+#        define PNG_BUILD_DLL
+#     else
+#        ifdef __DLL__
+            /* This is set by the Borland C system when compiling for a DLL
+             * (as above.)
+             */
+#           define PNG_BUILD_DLL
+#        else
+            /* Add additional compiler cases here. */
+#        endif
+#     endif
+#  endif
+#endif /* Setting PNG_BUILD_DLL if required */
+
+/* See pngconf.h for more details: the builder of the library may set this on
+ * the command line to the right thing for the specific compilation system or it
+ * may be automagically set above (at present we know of no system where it does
+ * need to be set on the command line.)
+ *
+ * PNG_IMPEXP must be set here when building the library to prevent pngconf.h
+ * setting it to the "import" setting for a DLL build.
+ */
+#ifndef PNG_IMPEXP
+#  ifdef PNG_BUILD_DLL
+#     define PNG_IMPEXP PNG_DLL_EXPORT
+#  else
+      /* Not building a DLL, or the DLL doesn't require specific export
+       * definitions.
+       */
+#     define PNG_IMPEXP
+#  endif
+#endif
+
+/* No warnings for private or deprecated functions in the build: */
+#ifndef PNG_DEPRECATED
+#  define PNG_DEPRECATED
+#endif
+#ifndef PNG_PRIVATE
+#  define PNG_PRIVATE
+#endif
+
+/* Symbol preprocessing support.
+ *
+ * To enable listing global, but internal, symbols the following macros should
+ * always be used to declare an extern data or function object in this file.
+ */
+#ifndef PNG_INTERNAL_DATA
+#  define PNG_INTERNAL_DATA(type, name, array) extern type name array
+#endif
+
+#ifndef PNG_INTERNAL_FUNCTION
+#  define PNG_INTERNAL_FUNCTION(type, name, args, attributes)\
+      extern PNG_FUNCTION(type, name, args, PNG_EMPTY attributes)
+#endif
+
+/* If floating or fixed point APIs are disabled they may still be compiled
+ * internally.  To handle this make sure they are declared as the appropriate
+ * internal extern function (otherwise the symbol prefixing stuff won't work and
+ * the functions will be used without definitions.)
+ *
+ * NOTE: although all the API functions are declared here they are not all
+ * actually built!  Because the declarations are still made it is necessary to
+ * fake out types that they depend on.
+ */
+#ifndef PNG_FP_EXPORT
+#  ifndef PNG_FLOATING_POINT_SUPPORTED
+#     define PNG_FP_EXPORT(ordinal, type, name, args)\
+         PNG_INTERNAL_FUNCTION(type, name, args, PNG_EMPTY);
+#     ifndef PNG_VERSION_INFO_ONLY
+         typedef struct png_incomplete png_double;
+         typedef png_double*           png_doublep;
+         typedef const png_double*     png_const_doublep;
+         typedef png_double**          png_doublepp;
+#     endif
+#  endif
+#endif
+#ifndef PNG_FIXED_EXPORT
+#  ifndef PNG_FIXED_POINT_SUPPORTED
+#     define PNG_FIXED_EXPORT(ordinal, type, name, args)\
+         PNG_INTERNAL_FUNCTION(type, name, args, PNG_EMPTY);
+#  endif
+#endif
+
+#include "png.h"
+
+/* pngconf.h does not set PNG_DLL_EXPORT unless it is required, so: */
+#ifndef PNG_DLL_EXPORT
+#  define PNG_DLL_EXPORT
+#endif
+
+/* SECURITY and SAFETY:
+ *
+ * By default libpng is built without any internal limits on image size,
+ * individual heap (png_malloc) allocations or the total amount of memory used.
+ * If PNG_SAFE_LIMITS_SUPPORTED is defined, however, the limits below are used
+ * (unless individually overridden).  These limits are believed to be fairly
+ * safe, but builders of secure systems should verify the values against the
+ * real system capabilities.
+ */
+#ifdef PNG_SAFE_LIMITS_SUPPORTED
+   /* 'safe' limits */
+#  ifndef PNG_USER_WIDTH_MAX
+#     define PNG_USER_WIDTH_MAX 1000000
+#  endif
+#  ifndef PNG_USER_HEIGHT_MAX
+#     define PNG_USER_HEIGHT_MAX 1000000
+#  endif
+#  ifndef PNG_USER_CHUNK_CACHE_MAX
+#     define PNG_USER_CHUNK_CACHE_MAX 128
+#  endif
+#  ifndef PNG_USER_CHUNK_MALLOC_MAX
+#     define PNG_USER_CHUNK_MALLOC_MAX 8000000
+#  endif
+#else
+   /* values for no limits */
+#  ifndef PNG_USER_WIDTH_MAX
+#     define PNG_USER_WIDTH_MAX 0x7fffffff
+#  endif
+#  ifndef PNG_USER_HEIGHT_MAX
+#     define PNG_USER_HEIGHT_MAX 0x7fffffff
+#  endif
+#  ifndef PNG_USER_CHUNK_CACHE_MAX
+#     define PNG_USER_CHUNK_CACHE_MAX 0
+#  endif
+#  ifndef PNG_USER_CHUNK_MALLOC_MAX
+#     define PNG_USER_CHUNK_MALLOC_MAX 0
+#  endif
+#endif
+
+/* Moved to pngpriv.h at libpng-1.5.0 */
+/* NOTE: some of these may have been used in external applications as
+ * these definitions were exposed in pngconf.h prior to 1.5.
+ */
+
+/* If you are running on a machine where you cannot allocate more
+ * than 64K of memory at once, uncomment this.  While libpng will not
+ * normally need that much memory in a chunk (unless you load up a very
+ * large file), zlib needs to know how big of a chunk it can use, and
+ * libpng thus makes sure to check any memory allocation to verify it
+ * will fit into memory.
+ *
+ * zlib provides 'MAXSEG_64K' which, if defined, indicates the
+ * same limit and pngconf.h (already included) sets the limit
+ * if certain operating systems are detected.
+ */
+#if defined(MAXSEG_64K) && !defined(PNG_MAX_MALLOC_64K)
+#  define PNG_MAX_MALLOC_64K
+#endif
+
+#ifndef PNG_UNUSED
+/* Unused formal parameter warnings are silenced using the following macro
+ * which is expected to have no bad effects on performance (optimizing
+ * compilers will probably remove it entirely).  Note that if you replace
+ * it with something other than whitespace, you must include the terminating
+ * semicolon.
+ */
+#  define PNG_UNUSED(param) (void)param;
+#endif
+
+/* Just a little check that someone hasn't tried to define something
+ * contradictory.
+ */
+#if (PNG_ZBUF_SIZE > 65536L) && defined(PNG_MAX_MALLOC_64K)
+#  undef PNG_ZBUF_SIZE
+#  define PNG_ZBUF_SIZE 65536L
+#endif
+
+/* If warnings or errors are turned off the code is disabled or redirected here.
+ * From 1.5.4 functions have been added to allow very limited formatting of
+ * error and warning messages - this code will also be disabled here.
+ */
+#ifdef PNG_WARNINGS_SUPPORTED
+#  define PNG_WARNING_PARAMETERS(p) png_warning_parameters p;
+#else
+#  define png_warning(s1,s2) ((void)(s1))
+#  define png_chunk_warning(s1,s2) ((void)(s1))
+#  define png_warning_parameter(p,number,string) ((void)0)
+#  define png_warning_parameter_unsigned(p,number,format,value) ((void)0)
+#  define png_warning_parameter_signed(p,number,format,value) ((void)0)
+#  define png_formatted_warning(pp,p,message) ((void)(pp))
+#  define PNG_WARNING_PARAMETERS(p)
+#endif
+#ifndef PNG_ERROR_TEXT_SUPPORTED
+#  define png_error(s1,s2) png_err(s1)
+#  define png_chunk_error(s1,s2) png_err(s1)
+#  define png_fixed_error(s1,s2) png_err(s1)
+#endif
+
+/* C allows up-casts from (void*) to any pointer and (const void*) to any
+ * pointer to a const object.  C++ regards this as a type error and requires an
+ * explicit, static, cast and provides the static_cast<> rune to ensure that
+ * const is not cast away.
+ */
+#ifdef __cplusplus
+#  define png_voidcast(type, value) static_cast<type>(value)
+#  define png_constcast(type, value) const_cast<type>(value)
+#  define png_aligncast(type, value) \
+   static_cast<type>(static_cast<void*>(value))
+#  define png_aligncastconst(type, value) \
+   static_cast<type>(static_cast<const void*>(value))
+#else
+#  define png_voidcast(type, value) (value)
+#  define png_constcast(type, value) ((type)(value))
+#  define png_aligncast(type, value) ((void*)(value))
+#  define png_aligncastconst(type, value) ((const void*)(value))
+#endif /* __cplusplus */
+
+/* Some fixed point APIs are still required even if not exported because
+ * they get used by the corresponding floating point APIs.  This magic
+ * deals with this:
+ */
+#ifdef PNG_FIXED_POINT_SUPPORTED
+#  define PNGFAPI PNGAPI
+#else
+#  define PNGFAPI /* PRIVATE */
+#endif
+
+#ifndef PNG_VERSION_INFO_ONLY
+/* Other defines specific to compilers can go here.  Try to keep
+ * them inside an appropriate ifdef/endif pair for portability.
+ */
+#if defined(PNG_FLOATING_POINT_SUPPORTED) ||\
+    defined(PNG_FLOATING_ARITHMETIC_SUPPORTED)
+   /* png.c requires the following ANSI-C constants if the conversion of
+    * floating point to ASCII is implemented therein:
+    *
+    *  DBL_DIG  Maximum number of decimal digits (can be set to any constant)
+    *  DBL_MIN  Smallest normalized fp number (can be set to an arbitrary value)
+    *  DBL_MAX  Maximum floating point number (can be set to an arbitrary value)
+    */
+#  include <float.h>
+
+#  if (defined(__MWERKS__) && defined(macintosh)) || defined(applec) || \
+    defined(THINK_C) || defined(__SC__) || defined(TARGET_OS_MAC)
+     /* We need to check that <math.h> hasn't already been included earlier
+      * as it seems it doesn't agree with <fp.h>, yet we should really use
+      * <fp.h> if possible.
+      */
+#    if !defined(__MATH_H__) && !defined(__MATH_H) && !defined(__cmath__)
+#      include <fp.h>
+#    endif
+#  else
+#    include <math.h>
+#  endif
+#  if defined(_AMIGA) && defined(__SASC) && defined(_M68881)
+     /* Amiga SAS/C: We must include builtin FPU functions when compiling using
+      * MATH=68881
+      */
+#    include <m68881.h>
+#  endif
+#endif
+
+/* This provides the non-ANSI (far) memory allocation routines. */
+#if defined(__TURBOC__) && defined(__MSDOS__)
+#  include <mem.h>
+#  include <alloc.h>
+#endif
+
+#if defined(WIN32) || defined(_Windows) || defined(_WINDOWS) || \
+    defined(_WIN32) || defined(__WIN32__)
+#  include <windows.h>  /* defines _WINDOWS_ macro */
+#endif
+#endif /* PNG_VERSION_INFO_ONLY */
+
+/* Moved here around 1.5.0beta36 from pngconf.h */
+/* Users may want to use these so they are not private.  Any library
+ * functions that are passed far data must be model-independent.
+ */
+
+/* Memory model/platform independent fns */
+#ifndef PNG_ABORT
+#  ifdef _WINDOWS_
+#    define PNG_ABORT() ExitProcess(0)
+#  else
+#    define PNG_ABORT() abort()
+#  endif
+#endif
+
+/* These macros may need to be architecture dependent. */
+#define PNG_ALIGN_NONE   0 /* do not use data alignment */
+#define PNG_ALIGN_ALWAYS 1 /* assume unaligned accesses are OK */
+#ifdef offsetof
+#  define PNG_ALIGN_OFFSET 2 /* use offsetof to determine alignment */
+#else
+#  define PNG_ALIGN_OFFSET -1 /* prevent the use of this */
+#endif
+#define PNG_ALIGN_SIZE   3 /* use sizeof to determine alignment */
+
+#ifndef PNG_ALIGN_TYPE
+   /* Default to using aligned access optimizations and requiring alignment to a
+    * multiple of the data type size.  Override in a compiler specific fashion
+    * if necessary by inserting tests here:
+    */
+#  define PNG_ALIGN_TYPE PNG_ALIGN_SIZE
+#endif
+
+#if PNG_ALIGN_TYPE == PNG_ALIGN_SIZE
+   /* This is used because in some compiler implementations non-aligned
+    * structure members are supported, so the offsetof approach below fails.
+    * Set PNG_ALIGN_SIZE=0 for compiler combinations where unaligned access
+    * is good for performance.  Do not do this unless you have tested the result
+    * and understand it.
+    */
+#  define png_alignof(type) (sizeof (type))
+#else
+#  if PNG_ALIGN_TYPE == PNG_ALIGN_OFFSET
+#     define png_alignof(type) offsetof(struct{char c; type t;}, t)
+#  else
+#     if PNG_ALIGN_TYPE == PNG_ALIGN_ALWAYS
+#        define png_alignof(type) (1)
+#     endif
+      /* Else leave png_alignof undefined to prevent use thereof */
+#  endif
+#endif
+
+/* This implicitly assumes alignment is always to a power of 2. */
+#ifdef png_alignof
+#  define png_isaligned(ptr, type)\
+   ((((const char*)ptr-(const char*)0) & (png_alignof(type)-1)) == 0)
+#else
+#  define png_isaligned(ptr, type) 0
+#endif
+
+/* End of memory model/platform independent support */
+/* End of 1.5.0beta36 move from pngconf.h */
+
+/* CONSTANTS and UTILITY MACROS
+ * These are used internally by libpng and not exposed in the API
+ */
+
+/* Various modes of operation.  Note that after an init, mode is set to
+ * zero automatically when the structure is created.  Three of these
+ * are defined in png.h because they need to be visible to applications
+ * that call png_set_unknown_chunk().
+ */
+/* #define PNG_HAVE_IHDR            0x01 (defined in png.h) */
+/* #define PNG_HAVE_PLTE            0x02 (defined in png.h) */
+#define PNG_HAVE_IDAT               0x04
+/* #define PNG_AFTER_IDAT           0x08 (defined in png.h) */
+#define PNG_HAVE_IEND               0x10
+                   /*               0x20 (unused) */
+                   /*               0x40 (unused) */
+                   /*               0x80 (unused) */
+#define PNG_HAVE_CHUNK_HEADER      0x100
+#define PNG_WROTE_tIME             0x200
+#define PNG_WROTE_INFO_BEFORE_PLTE 0x400
+#define PNG_BACKGROUND_IS_GRAY     0x800
+#define PNG_HAVE_PNG_SIGNATURE    0x1000
+#define PNG_HAVE_CHUNK_AFTER_IDAT 0x2000 /* Have another chunk after IDAT */
+                   /*             0x4000 (unused) */
+#define PNG_IS_READ_STRUCT        0x8000 /* Else is a write struct */
+
+/* Flags for the transformations the PNG library does on the image data */
+#define PNG_BGR                 0x0001
+#define PNG_INTERLACE           0x0002
+#define PNG_PACK                0x0004
+#define PNG_SHIFT               0x0008
+#define PNG_SWAP_BYTES          0x0010
+#define PNG_INVERT_MONO         0x0020
+#define PNG_QUANTIZE            0x0040
+#define PNG_COMPOSE             0x0080     /* Was PNG_BACKGROUND */
+#define PNG_BACKGROUND_EXPAND   0x0100
+#define PNG_EXPAND_16           0x0200     /* Added to libpng 1.5.2 */
+#define PNG_16_TO_8             0x0400     /* Becomes 'chop' in 1.5.4 */
+#define PNG_RGBA                0x0800
+#define PNG_EXPAND              0x1000
+#define PNG_GAMMA               0x2000
+#define PNG_GRAY_TO_RGB         0x4000
+#define PNG_FILLER              0x8000
+#define PNG_PACKSWAP           0x10000
+#define PNG_SWAP_ALPHA         0x20000
+#define PNG_STRIP_ALPHA        0x40000
+#define PNG_INVERT_ALPHA       0x80000
+#define PNG_USER_TRANSFORM    0x100000
+#define PNG_RGB_TO_GRAY_ERR   0x200000
+#define PNG_RGB_TO_GRAY_WARN  0x400000
+#define PNG_RGB_TO_GRAY       0x600000 /* two bits, RGB_TO_GRAY_ERR|WARN */
+#define PNG_ENCODE_ALPHA      0x800000 /* Added to libpng-1.5.4 */
+#define PNG_ADD_ALPHA         0x1000000 /* Added to libpng-1.2.7 */
+#define PNG_EXPAND_tRNS       0x2000000 /* Added to libpng-1.2.9 */
+#define PNG_SCALE_16_TO_8     0x4000000 /* Added to libpng-1.5.4 */
+                       /*   0x8000000 unused */
+                       /*  0x10000000 unused */
+                       /*  0x20000000 unused */
+                       /*  0x40000000 unused */
+/* Flags for png_create_struct */
+#define PNG_STRUCT_PNG   0x0001
+#define PNG_STRUCT_INFO  0x0002
+
+/* Scaling factor for filter heuristic weighting calculations */
+#define PNG_WEIGHT_FACTOR (1<<(PNG_WEIGHT_SHIFT))
+#define PNG_COST_FACTOR (1<<(PNG_COST_SHIFT))
+
+/* Flags for the png_ptr->flags rather than declaring a byte for each one */
+#define PNG_FLAG_ZLIB_CUSTOM_STRATEGY     0x0001
+#define PNG_FLAG_ZSTREAM_INITIALIZED      0x0002 /* Added to libpng-1.6.0 */
+                                  /*      0x0004    unused */
+#define PNG_FLAG_ZSTREAM_ENDED            0x0008 /* Added to libpng-1.6.0 */
+                                  /*      0x0010    unused */
+                                  /*      0x0020    unused */
+#define PNG_FLAG_ROW_INIT                 0x0040
+#define PNG_FLAG_FILLER_AFTER             0x0080
+#define PNG_FLAG_CRC_ANCILLARY_USE        0x0100
+#define PNG_FLAG_CRC_ANCILLARY_NOWARN     0x0200
+#define PNG_FLAG_CRC_CRITICAL_USE         0x0400
+#define PNG_FLAG_CRC_CRITICAL_IGNORE      0x0800
+#define PNG_FLAG_ASSUME_sRGB              0x1000 /* Added to libpng-1.5.4 */
+#define PNG_FLAG_OPTIMIZE_ALPHA           0x2000 /* Added to libpng-1.5.4 */
+#define PNG_FLAG_DETECT_UNINITIALIZED     0x4000 /* Added to libpng-1.5.4 */
+/* #define PNG_FLAG_KEEP_UNKNOWN_CHUNKS      0x8000 */
+/* #define PNG_FLAG_KEEP_UNSAFE_CHUNKS      0x10000 */
+#define PNG_FLAG_LIBRARY_MISMATCH        0x20000
+#define PNG_FLAG_STRIP_ERROR_NUMBERS     0x40000
+#define PNG_FLAG_STRIP_ERROR_TEXT        0x80000
+#define PNG_FLAG_BENIGN_ERRORS_WARN     0x100000 /* Added to libpng-1.4.0 */
+#define PNG_FLAG_APP_WARNINGS_WARN      0x200000 /* Added to libpng-1.6.0 */
+#define PNG_FLAG_APP_ERRORS_WARN        0x400000 /* Added to libpng-1.6.0 */
+                                  /*    0x800000    unused */
+                                  /*   0x1000000    unused */
+                                  /*   0x2000000    unused */
+                                  /*   0x4000000    unused */
+                                  /*   0x8000000    unused */
+                                  /*  0x10000000    unused */
+                                  /*  0x20000000    unused */
+                                  /*  0x40000000    unused */
+
+#define PNG_FLAG_CRC_ANCILLARY_MASK (PNG_FLAG_CRC_ANCILLARY_USE | \
+                                     PNG_FLAG_CRC_ANCILLARY_NOWARN)
+
+#define PNG_FLAG_CRC_CRITICAL_MASK  (PNG_FLAG_CRC_CRITICAL_USE | \
+                                     PNG_FLAG_CRC_CRITICAL_IGNORE)
+
+#define PNG_FLAG_CRC_MASK           (PNG_FLAG_CRC_ANCILLARY_MASK | \
+                                     PNG_FLAG_CRC_CRITICAL_MASK)
+
+/* Save typing and make code easier to understand */
+
+#define PNG_COLOR_DIST(c1, c2) (abs((int)((c1).red) - (int)((c2).red)) + \
+   abs((int)((c1).green) - (int)((c2).green)) + \
+   abs((int)((c1).blue) - (int)((c2).blue)))
+
+/* Added to libpng-1.6.0: scale a 16-bit value in the range 0..65535 to 0..255
+ * by dividing by 257 *with rounding*.  This macro is exact for the given range.
+ * See the discourse in pngrtran.c png_do_scale_16_to_8.  The values in the
+ * macro were established by experiment (modifying the added value).  The macro
+ * has a second variant that takes a value already scaled by 255 and divides by
+ * 65535 - this has a maximum error of .502.  Over the range 0..65535*65535 it
+ * only gives off-by-one errors and only for 0.5% (1 in 200) of the values.
+ */
+#define PNG_DIV65535(v24) (((v24) + 32895) >> 16)
+#define PNG_DIV257(v16) PNG_DIV65535((png_uint_32)(v16) * 255)
+
+/* Added to libpng-1.2.6 JB */
+#define PNG_ROWBYTES(pixel_bits, width) \
+    ((pixel_bits) >= 8 ? \
+    ((png_size_t)(width) * (((png_size_t)(pixel_bits)) >> 3)) : \
+    (( ((png_size_t)(width) * ((png_size_t)(pixel_bits))) + 7) >> 3) )
+
+/* PNG_OUT_OF_RANGE returns true if value is outside the range
+ * ideal-delta..ideal+delta.  Each argument is evaluated twice.
+ * "ideal" and "delta" should be constants, normally simple
+ * integers, "value" a variable. Added to libpng-1.2.6 JB
+ */
+#define PNG_OUT_OF_RANGE(value, ideal, delta) \
+   ( (value) < (ideal)-(delta) || (value) > (ideal)+(delta) )
+
+/* Conversions between fixed and floating point, only defined if
+ * required (to make sure the code doesn't accidentally use float
+ * when it is supposedly disabled.)
+ */
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+/* The floating point conversion can't overflow, though it can and
+ * does lose accuracy relative to the original fixed point value.
+ * In practice this doesn't matter because png_fixed_point only
+ * stores numbers with very low precision.  The png_ptr and s
+ * arguments are unused by default but are there in case error
+ * checking becomes a requirement.
+ */
+#define png_float(png_ptr, fixed, s) (.00001 * (fixed))
+
+/* The fixed point conversion performs range checking and evaluates
+ * its argument multiple times, so must be used with care.  The
+ * range checking uses the PNG specification values for a signed
+ * 32 bit fixed point value except that the values are deliberately
+ * rounded-to-zero to an integral value - 21474 (21474.83 is roughly
+ * (2^31-1) * 100000). 's' is a string that describes the value being
+ * converted.
+ *
+ * NOTE: this macro will raise a png_error if the range check fails,
+ * therefore it is normally only appropriate to use this on values
+ * that come from API calls or other sources where an out of range
+ * error indicates a programming error, not a data error!
+ *
+ * NOTE: by default this is off - the macro is not used - because the
+ * function call saves a lot of code.
+ */
+#ifdef PNG_FIXED_POINT_MACRO_SUPPORTED
+#define png_fixed(png_ptr, fp, s) ((fp) <= 21474 && (fp) >= -21474 ?\
+    ((png_fixed_point)(100000 * (fp))) : (png_fixed_error(png_ptr, s),0))
+#endif
+/* else the corresponding function is defined below, inside the scope of the
+ * cplusplus test.
+ */
+#endif
+
+/* Constants for known chunk types.  If you need to add a chunk, define the name
+ * here.  For historical reasons these constants have the form png_<name>; i.e.
+ * the prefix is lower case.  Please use decimal values as the parameters to
+ * match the ISO PNG specification and to avoid relying on the C locale
+ * interpretation of character values.
+ *
+ * Prior to 1.5.6 these constants were strings, as of 1.5.6 png_uint_32 values
+ * are computed and a new macro (PNG_STRING_FROM_CHUNK) added to allow a string
+ * to be generated if required.
+ *
+ * PNG_32b correctly produces a value shifted by up to 24 bits, even on
+ * architectures where (int) is only 16 bits.
+ */
+#define PNG_32b(b,s) ((png_uint_32)(b) << (s))
+#define PNG_U32(b1,b2,b3,b4) \
+   (PNG_32b(b1,24) | PNG_32b(b2,16) | PNG_32b(b3,8) | PNG_32b(b4,0))
+
+/* Constants for known chunk types.
+ *
+ * MAINTAINERS: If you need to add a chunk, define the name here.
+ * For historical reasons these constants have the form png_<name>; i.e.
+ * the prefix is lower case.  Please use decimal values as the parameters to
+ * match the ISO PNG specification and to avoid relying on the C locale
+ * interpretation of character values.  Please keep the list sorted.
+ *
+ * Notice that PNG_U32 is used to define a 32-bit value for the 4 byte chunk
+ * type.  In fact the specification does not express chunk types this way,
+ * however using a 32-bit value means that the chunk type can be read from the
+ * stream using exactly the same code as used for a 32-bit unsigned value and
+ * can be examined far more efficiently (using one arithmetic compare).
+ *
+ * Prior to 1.5.6 the chunk type constants were expressed as C strings.  The
+ * libpng API still uses strings for 'unknown' chunks and a macro,
+ * PNG_STRING_FROM_CHUNK, allows a string to be generated if required.  Notice
+ * that for portable code numeric values must still be used; the string "IHDR"
+ * is not portable and neither is PNG_U32('I', 'H', 'D', 'R').
+ *
+ * In 1.7.0 the definitions will be made public in png.h to avoid having to
+ * duplicate the same definitions in application code.
+ */
+#define png_IDAT PNG_U32( 73,  68,  65,  84)
+#define png_IEND PNG_U32( 73,  69,  78,  68)
+#define png_IHDR PNG_U32( 73,  72,  68,  82)
+#define png_PLTE PNG_U32( 80,  76,  84,  69)
+#define png_bKGD PNG_U32( 98,  75,  71,  68)
+#define png_cHRM PNG_U32( 99,  72,  82,  77)
+#define png_fRAc PNG_U32(102,  82,  65,  99) /* registered, not defined */
+#define png_gAMA PNG_U32(103,  65,  77,  65)
+#define png_gIFg PNG_U32(103,  73,  70, 103)
+#define png_gIFt PNG_U32(103,  73,  70, 116) /* deprecated */
+#define png_gIFx PNG_U32(103,  73,  70, 120)
+#define png_hIST PNG_U32(104,  73,  83,  84)
+#define png_iCCP PNG_U32(105,  67,  67,  80)
+#define png_iTXt PNG_U32(105,  84,  88, 116)
+#define png_oFFs PNG_U32(111,  70,  70, 115)
+#define png_pCAL PNG_U32(112,  67,  65,  76)
+#define png_pHYs PNG_U32(112,  72,  89, 115)
+#define png_sBIT PNG_U32(115,  66,  73,  84)
+#define png_sCAL PNG_U32(115,  67,  65,  76)
+#define png_sPLT PNG_U32(115,  80,  76,  84)
+#define png_sRGB PNG_U32(115,  82,  71,  66)
+#define png_sTER PNG_U32(115,  84,  69,  82)
+#define png_tEXt PNG_U32(116,  69,  88, 116)
+#define png_tIME PNG_U32(116,  73,  77,  69)
+#define png_tRNS PNG_U32(116,  82,  78,  83)
+#define png_zTXt PNG_U32(122,  84,  88, 116)
+
+/* The following will work on (signed char*) strings, whereas the get_uint_32
+ * macro will fail on top-bit-set values because of the sign extension.
+ */
+#define PNG_CHUNK_FROM_STRING(s)\
+   PNG_U32(0xff&(s)[0], 0xff&(s)[1], 0xff&(s)[2], 0xff&(s)[3])
+
+/* This uses (char), not (png_byte) to avoid warnings on systems where (char) is
+ * signed and the argument is a (char[])  This macro will fail miserably on
+ * systems where (char) is more than 8 bits.
+ */
+#define PNG_STRING_FROM_CHUNK(s,c)\
+   (void)(((char*)(s))[0]=(char)((c)>>24), ((char*)(s))[1]=(char)((c)>>16),\
+   ((char*)(s))[2]=(char)((c)>>8), ((char*)(s))[3]=(char)((c)))
+
+/* Do the same but terminate with a null character. */
+#define PNG_CSTRING_FROM_CHUNK(s,c)\
+   (void)(PNG_STRING_FROM_CHUNK(s,c), ((char*)(s))[4] = 0)
+
+/* Test on flag values as defined in the spec (section 5.4): */
+#define PNG_CHUNK_ANCILLARY(c)   (1 & ((c) >> 29))
+#define PNG_CHUNK_CRITICAL(c)     (!PNG_CHUNK_ANCILLARY(c))
+#define PNG_CHUNK_PRIVATE(c)      (1 & ((c) >> 21))
+#define PNG_CHUNK_RESERVED(c)     (1 & ((c) >> 13))
+#define PNG_CHUNK_SAFE_TO_COPY(c) (1 & ((c) >>  5))
+
+/* Gamma values (new at libpng-1.5.4): */
+#define PNG_GAMMA_MAC_OLD 151724  /* Assume '1.8' is really 2.2/1.45! */
+#define PNG_GAMMA_MAC_INVERSE 65909
+#define PNG_GAMMA_sRGB_INVERSE 45455
+
+/* Almost everything below is C specific; the #defines above can be used in
+ * non-C code (so long as it is C-preprocessed) the rest of this stuff cannot.
+ */
+#ifndef PNG_VERSION_INFO_ONLY
+
+#include "pngstruct.h"
+#include "pnginfo.h"
+
+/* Validate the include paths - the include path used to generate pnglibconf.h
+ * must match that used in the build, or we must be using pnglibconf.h.prebuilt:
+ */
+#if PNG_ZLIB_VERNUM != 0 && PNG_ZLIB_VERNUM != ZLIB_VERNUM
+#  error ZLIB_VERNUM != PNG_ZLIB_VERNUM \
+      "-I (include path) error: see the notes in pngpriv.h"
+   /* This means that when pnglibconf.h was built the copy of zlib.h that it
+    * used is not the same as the one being used here.  Because the build of
+    * libpng makes decisions to use inflateInit2 and inflateReset2 based on the
+    * zlib version number and because this affects handling of certain broken
+    * PNG files the -I directives must match.
+    *
+    * The most likely explanation is that you passed a -I in CFLAGS, this will
+    * not work; all the preprocessor directories and in particular all the -I
+    * directives must be in CPPFLAGS.
+    */
+#endif
+
+/* This is used for 16 bit gamma tables -- only the top level pointers are
+ * const; this could be changed:
+ */
+typedef const png_uint_16p * png_const_uint_16pp;
+
+/* Added to libpng-1.5.7: sRGB conversion tables */
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) ||\
+   defined(PNG_SIMPLIFIED_WRITE_SUPPORTED)
+#ifdef PNG_SIMPLIFIED_READ_SUPPORTED
+PNG_INTERNAL_DATA(const png_uint_16, png_sRGB_table, [256]);
+   /* Convert from an sRGB encoded value 0..255 to a 16-bit linear value,
+    * 0..65535.  This table gives the closest 16-bit answers (no errors).
+    */
+#endif
+
+PNG_INTERNAL_DATA(const png_uint_16, png_sRGB_base, [512]);
+PNG_INTERNAL_DATA(const png_byte, png_sRGB_delta, [512]);
+
+#define PNG_sRGB_FROM_LINEAR(linear) ((png_byte)((png_sRGB_base[(linear)>>15] +\
+   ((((linear)&0x7fff)*png_sRGB_delta[(linear)>>15])>>12)) >> 8))
+   /* Given a value 'linear' in the range 0..255*65535 calculate the 8-bit sRGB
+    * encoded value with maximum error 0.646365.  Note that the input is not a
+    * 16-bit value; it has been multiplied by 255! */
+#endif /* PNG_SIMPLIFIED_READ/WRITE */
+
+
+/* Inhibit C++ name-mangling for libpng functions but not for system calls. */
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+/* Internal functions; these are not exported from a DLL however because they
+ * are used within several of the C source files they have to be C extern.
+ *
+ * All of these functions must be declared with PNG_INTERNAL_FUNCTION.
+ */
+
+/* Zlib support */
+#define PNG_UNEXPECTED_ZLIB_RETURN (-7)
+PNG_INTERNAL_FUNCTION(void, png_zstream_error,(png_structrp png_ptr, int ret),
+   PNG_EMPTY);
+   /* Used by the zlib handling functions to ensure that z_stream::msg is always
+    * set before they return.
+    */
+
+#ifdef PNG_WRITE_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_free_buffer_list,(png_structrp png_ptr,
+   png_compression_bufferp *list),PNG_EMPTY);
+   /* Free the buffer list used by the compressed write code. */
+#endif
+
+#if defined(PNG_FLOATING_POINT_SUPPORTED) && \
+   !defined(PNG_FIXED_POINT_MACRO_SUPPORTED) && \
+   (defined(PNG_gAMA_SUPPORTED) || defined(PNG_cHRM_SUPPORTED) || \
+   defined(PNG_sCAL_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)) || \
+   (defined(PNG_sCAL_SUPPORTED) && \
+   defined(PNG_FLOATING_ARITHMETIC_SUPPORTED))
+PNG_INTERNAL_FUNCTION(png_fixed_point,png_fixed,(png_const_structrp png_ptr,
+   double fp, png_const_charp text),PNG_EMPTY);
+#endif
+
+/* Check the user version string for compatibility, returns false if the version
+ * numbers aren't compatible.
+ */
+PNG_INTERNAL_FUNCTION(int,png_user_version_check,(png_structrp png_ptr,
+   png_const_charp user_png_ver),PNG_EMPTY);
+
+/* Internal base allocator - no messages, NULL on failure to allocate.  This
+ * does, however, call the application provided allocator and that could call
+ * png_error (although that would be a bug in the application implementation.)
+ */
+PNG_INTERNAL_FUNCTION(png_voidp,png_malloc_base,(png_const_structrp png_ptr,
+   png_alloc_size_t size),PNG_ALLOCATED);
+
+#if defined(PNG_TEXT_SUPPORTED) || defined(PNG_sPLT_SUPPORTED) ||\
+   defined(PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED)
+/* Internal array allocator, outputs no error or warning messages on failure,
+ * just returns NULL.  
+ */
+PNG_INTERNAL_FUNCTION(png_voidp,png_malloc_array,(png_const_structrp png_ptr,
+   int nelements, size_t element_size),PNG_ALLOCATED);
+
+/* The same but an existing array is extended by add_elements.  This function
+ * also memsets the new elements to 0 and copies the old elements.  The old
+ * array is not freed or altered.
+ */
+PNG_INTERNAL_FUNCTION(png_voidp,png_realloc_array,(png_const_structrp png_ptr,
+   png_const_voidp array, int old_elements, int add_elements,
+   size_t element_size),PNG_ALLOCATED);
+#endif /* text, sPLT or unknown chunks */
+
+/* Magic to create a struct when there is no struct to call the user supplied
+ * memory allocators.  Because error handling has not been set up the memory
+ * handlers can't safely call png_error, but this is an obscure and undocumented
+ * restriction so libpng has to assume that the 'free' handler, at least, might
+ * call png_error.
+ */
+PNG_INTERNAL_FUNCTION(png_structp,png_create_png_struct,
+   (png_const_charp user_png_ver, png_voidp error_ptr, png_error_ptr error_fn,
+    png_error_ptr warn_fn, png_voidp mem_ptr, png_malloc_ptr malloc_fn,
+    png_free_ptr free_fn),PNG_ALLOCATED);
+
+/* Free memory from internal libpng struct */
+PNG_INTERNAL_FUNCTION(void,png_destroy_png_struct,(png_structrp png_ptr),
+   PNG_EMPTY);
+
+/* Free an allocated jmp_buf (always succeeds) */
+PNG_INTERNAL_FUNCTION(void,png_free_jmpbuf,(png_structrp png_ptr),PNG_EMPTY);
+
+/* Function to allocate memory for zlib.  PNGAPI is disallowed. */
+PNG_INTERNAL_FUNCTION(voidpf,png_zalloc,(voidpf png_ptr, uInt items, uInt size),
+   PNG_ALLOCATED);
+
+/* Function to free memory for zlib.  PNGAPI is disallowed. */
+PNG_INTERNAL_FUNCTION(void,png_zfree,(voidpf png_ptr, voidpf ptr),PNG_EMPTY);
+
+/* Next four functions are used internally as callbacks.  PNGCBAPI is required
+ * but not PNG_EXPORT.  PNGAPI added at libpng version 1.2.3, changed to
+ * PNGCBAPI at 1.5.0
+ */
+
+PNG_INTERNAL_FUNCTION(void PNGCBAPI,png_default_read_data,(png_structp png_ptr,
+    png_bytep data, png_size_t length),PNG_EMPTY);
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+PNG_INTERNAL_FUNCTION(void PNGCBAPI,png_push_fill_buffer,(png_structp png_ptr,
+    png_bytep buffer, png_size_t length),PNG_EMPTY);
+#endif
+
+PNG_INTERNAL_FUNCTION(void PNGCBAPI,png_default_write_data,(png_structp png_ptr,
+    png_bytep data, png_size_t length),PNG_EMPTY);
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+#  ifdef PNG_STDIO_SUPPORTED
+PNG_INTERNAL_FUNCTION(void PNGCBAPI,png_default_flush,(png_structp png_ptr),
+   PNG_EMPTY);
+#  endif
+#endif
+
+/* Reset the CRC variable */
+PNG_INTERNAL_FUNCTION(void,png_reset_crc,(png_structrp png_ptr),PNG_EMPTY);
+
+/* Write the "data" buffer to whatever output you are using */
+PNG_INTERNAL_FUNCTION(void,png_write_data,(png_structrp png_ptr,
+    png_const_bytep data, png_size_t length),PNG_EMPTY);
+
+/* Read and check the PNG file signature */
+PNG_INTERNAL_FUNCTION(void,png_read_sig,(png_structrp png_ptr,
+   png_inforp info_ptr),PNG_EMPTY);
+
+/* Read the chunk header (length + type name) */
+PNG_INTERNAL_FUNCTION(png_uint_32,png_read_chunk_header,(png_structrp png_ptr),
+   PNG_EMPTY);
+
+/* Read data from whatever input you are using into the "data" buffer */
+PNG_INTERNAL_FUNCTION(void,png_read_data,(png_structrp png_ptr, png_bytep data,
+    png_size_t length),PNG_EMPTY);
+
+/* Read bytes into buf, and update png_ptr->crc */
+PNG_INTERNAL_FUNCTION(void,png_crc_read,(png_structrp png_ptr, png_bytep buf,
+    png_uint_32 length),PNG_EMPTY);
+
+/* Read "skip" bytes, read the file crc, and (optionally) verify png_ptr->crc */
+PNG_INTERNAL_FUNCTION(int,png_crc_finish,(png_structrp png_ptr,
+   png_uint_32 skip),PNG_EMPTY);
+
+/* Read the CRC from the file and compare it to the libpng calculated CRC */
+PNG_INTERNAL_FUNCTION(int,png_crc_error,(png_structrp png_ptr),PNG_EMPTY);
+
+/* Calculate the CRC over a section of data.  Note that we are only
+ * passing a maximum of 64K on systems that have this as a memory limit,
+ * since this is the maximum buffer size we can specify.
+ */
+PNG_INTERNAL_FUNCTION(void,png_calculate_crc,(png_structrp png_ptr,
+   png_const_bytep ptr, png_size_t length),PNG_EMPTY);
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_flush,(png_structrp png_ptr),PNG_EMPTY);
+#endif
+
+/* Write various chunks */
+
+/* Write the IHDR chunk, and update the png_struct with the necessary
+ * information.
+ */
+PNG_INTERNAL_FUNCTION(void,png_write_IHDR,(png_structrp png_ptr,
+   png_uint_32 width, png_uint_32 height, int bit_depth, int color_type,
+   int compression_method, int filter_method, int interlace_method),PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(void,png_write_PLTE,(png_structrp png_ptr,
+   png_const_colorp palette, png_uint_32 num_pal),PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(void,png_compress_IDAT,(png_structrp png_ptr,
+   png_const_bytep row_data, png_alloc_size_t row_data_length, int flush),
+   PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(void,png_write_IEND,(png_structrp png_ptr),PNG_EMPTY);
+
+#ifdef PNG_WRITE_gAMA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_gAMA_fixed,(png_structrp png_ptr,
+    png_fixed_point file_gamma),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_sBIT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_sBIT,(png_structrp png_ptr,
+    png_const_color_8p sbit, int color_type),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_cHRM_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_cHRM_fixed,(png_structrp png_ptr,
+    const png_xy *xy), PNG_EMPTY);
+    /* The xy value must have been previously validated */
+#endif
+
+#ifdef PNG_WRITE_sRGB_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_sRGB,(png_structrp png_ptr,
+    int intent),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_iCCP_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_iCCP,(png_structrp png_ptr,
+   png_const_charp name, png_const_bytep profile), PNG_EMPTY);
+   /* The profile must have been previously validated for correctness, the
+    * length comes from the first four bytes.  Only the base, deflate,
+    * compression is supported.
+    */
+#endif
+
+#ifdef PNG_WRITE_sPLT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_sPLT,(png_structrp png_ptr,
+    png_const_sPLT_tp palette),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_tRNS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_tRNS,(png_structrp png_ptr,
+    png_const_bytep trans, png_const_color_16p values, int number,
+    int color_type),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_bKGD_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_bKGD,(png_structrp png_ptr,
+    png_const_color_16p values, int color_type),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_hIST_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_hIST,(png_structrp png_ptr,
+    png_const_uint_16p hist, int num_hist),PNG_EMPTY);
+#endif
+
+/* Chunks that have keywords */
+#ifdef PNG_WRITE_tEXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_tEXt,(png_structrp png_ptr,
+   png_const_charp key, png_const_charp text, png_size_t text_len),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_zTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_zTXt,(png_structrp png_ptr, png_const_charp
+    key, png_const_charp text, png_size_t text_len, int compression),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_iTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_iTXt,(png_structrp png_ptr,
+    int compression, png_const_charp key, png_const_charp lang,
+    png_const_charp lang_key, png_const_charp text),PNG_EMPTY);
+#endif
+
+#ifdef PNG_TEXT_SUPPORTED  /* Added at version 1.0.14 and 1.2.4 */
+PNG_INTERNAL_FUNCTION(int,png_set_text_2,(png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_textp text_ptr, int num_text),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_oFFs_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_oFFs,(png_structrp png_ptr,
+    png_int_32 x_offset, png_int_32 y_offset, int unit_type),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_pCAL_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_pCAL,(png_structrp png_ptr,
+    png_charp purpose, png_int_32 X0, png_int_32 X1, int type, int nparams,
+    png_const_charp units, png_charpp params),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_pHYs_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_pHYs,(png_structrp png_ptr,
+    png_uint_32 x_pixels_per_unit, png_uint_32 y_pixels_per_unit,
+    int unit_type),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_tIME_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_tIME,(png_structrp png_ptr,
+    png_const_timep mod_time),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_sCAL_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_write_sCAL_s,(png_structrp png_ptr,
+    int unit, png_const_charp width, png_const_charp height),PNG_EMPTY);
+#endif
+
+/* Called when finished processing a row of data */
+PNG_INTERNAL_FUNCTION(void,png_write_finish_row,(png_structrp png_ptr),
+    PNG_EMPTY);
+
+/* Internal use only.   Called before first row of data */
+PNG_INTERNAL_FUNCTION(void,png_write_start_row,(png_structrp png_ptr),
+    PNG_EMPTY);
+
+/* Combine a row of data, dealing with alpha, etc. if requested.  'row' is an
+ * array of png_ptr->width pixels.  If the image is not interlaced or this
+ * is the final pass this just does a memcpy, otherwise the "display" flag
+ * is used to determine whether to copy pixels that are not in the current pass.
+ *
+ * Because 'png_do_read_interlace' (below) replicates pixels this allows this
+ * function to achieve the documented 'blocky' appearance during interlaced read
+ * if display is 1 and the 'sparkle' appearance, where existing pixels in 'row'
+ * are not changed if they are not in the current pass, when display is 0.
+ *
+ * 'display' must be 0 or 1, otherwise the memcpy will be done regardless.
+ *
+ * The API always reads from the png_struct row buffer and always assumes that
+ * it is full width (png_do_read_interlace has already been called.)
+ *
+ * This function is only ever used to write to row buffers provided by the
+ * caller of the relevant libpng API and the row must have already been
+ * transformed by the read transformations.
+ *
+ * The PNG_USE_COMPILE_TIME_MASKS option causes generation of pre-computed
+ * bitmasks for use within the code, otherwise runtime generated masks are used.
+ * The default is compile time masks.
+ */
+#ifndef PNG_USE_COMPILE_TIME_MASKS
+#  define PNG_USE_COMPILE_TIME_MASKS 1
+#endif
+PNG_INTERNAL_FUNCTION(void,png_combine_row,(png_const_structrp png_ptr,
+    png_bytep row, int display),PNG_EMPTY);
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+/* Expand an interlaced row: the 'row_info' describes the pass data that has
+ * been read in and must correspond to the pixels in 'row', the pixels are
+ * expanded (moved apart) in 'row' to match the final layout, when doing this
+ * the pixels are *replicated* to the intervening space.  This is essential for
+ * the correct operation of png_combine_row, above.
+ */
+PNG_INTERNAL_FUNCTION(void,png_do_read_interlace,(png_row_infop row_info,
+    png_bytep row, int pass, png_uint_32 transformations),PNG_EMPTY);
+#endif
+
+/* GRR TO DO (2.0 or whenever):  simplify other internal calling interfaces */
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+/* Grab pixels out of a row for an interlaced pass */
+PNG_INTERNAL_FUNCTION(void,png_do_write_interlace,(png_row_infop row_info,
+    png_bytep row, int pass),PNG_EMPTY);
+#endif
+
+/* Unfilter a row: check the filter value before calling this, there is no point
+ * calling it for PNG_FILTER_VALUE_NONE.
+ */
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row,(png_structrp pp, png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row, int filter),PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_up_neon,(png_row_infop row_info,
+    png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub3_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub4_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg3_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg4_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth3_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_neon,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+
+/* Choose the best filter to use and filter the row data */
+PNG_INTERNAL_FUNCTION(void,png_write_find_filter,(png_structrp png_ptr,
+    png_row_infop row_info),PNG_EMPTY);
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_read_IDAT_data,(png_structrp png_ptr,
+   png_bytep output, png_alloc_size_t avail_out),PNG_EMPTY);
+   /* Read 'avail_out' bytes of data from the IDAT stream.  If the output buffer
+    * is NULL the function checks, instead, for the end of the stream.  In this
+    * case a benign error will be issued if the stream end is not found or if
+    * extra data has to be consumed.
+    */
+PNG_INTERNAL_FUNCTION(void,png_read_finish_IDAT,(png_structrp png_ptr),
+   PNG_EMPTY);
+   /* This cleans up when the IDAT LZ stream does not end when the last image
+    * byte is read; there is still some pending input.
+    */
+
+PNG_INTERNAL_FUNCTION(void,png_read_finish_row,(png_structrp png_ptr),
+   PNG_EMPTY);
+   /* Finish a row while reading, dealing with interlacing passes, etc. */
+#endif
+
+/* Initialize the row buffers, etc. */
+PNG_INTERNAL_FUNCTION(void,png_read_start_row,(png_structrp png_ptr),PNG_EMPTY);
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+/* Optional call to update the users info structure */
+PNG_INTERNAL_FUNCTION(void,png_read_transform_info,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+#endif
+
+/* These are the functions that do the transformations */
+#ifdef PNG_READ_FILLER_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_read_filler,(png_row_infop row_info,
+    png_bytep row, png_uint_32 filler, png_uint_32 flags),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_SWAP_ALPHA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_read_swap_alpha,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_SWAP_ALPHA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_write_swap_alpha,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_INVERT_ALPHA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_read_invert_alpha,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_INVERT_ALPHA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_write_invert_alpha,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#if defined(PNG_WRITE_FILLER_SUPPORTED) || \
+    defined(PNG_READ_STRIP_ALPHA_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_strip_channel,(png_row_infop row_info,
+    png_bytep row, int at_start),PNG_EMPTY);
+#endif
+
+#ifdef PNG_16BIT_SUPPORTED
+#if defined(PNG_READ_SWAP_SUPPORTED) || defined(PNG_WRITE_SWAP_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_swap,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+#endif
+
+#if defined(PNG_READ_PACKSWAP_SUPPORTED) || \
+    defined(PNG_WRITE_PACKSWAP_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_packswap,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+PNG_INTERNAL_FUNCTION(int,png_do_rgb_to_gray,(png_structrp png_ptr,
+    png_row_infop row_info, png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_gray_to_rgb,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_PACK_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_unpack,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_SHIFT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_unshift,(png_row_infop row_info,
+    png_bytep row, png_const_color_8p sig_bits),PNG_EMPTY);
+#endif
+
+#if defined(PNG_READ_INVERT_SUPPORTED) || defined(PNG_WRITE_INVERT_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_invert,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_scale_16_to_8,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_chop,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_quantize,(png_row_infop row_info,
+    png_bytep row, png_const_bytep palette_lookup,
+    png_const_bytep quantize_lookup),PNG_EMPTY);
+
+#  ifdef PNG_CORRECT_PALETTE_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_correct_palette,(png_structrp png_ptr,
+    png_colorp palette, int num_palette),PNG_EMPTY);
+#  endif
+#endif
+
+#if defined(PNG_READ_BGR_SUPPORTED) || defined(PNG_WRITE_BGR_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_bgr,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_PACK_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_pack,(png_row_infop row_info,
+   png_bytep row, png_uint_32 bit_depth),PNG_EMPTY);
+#endif
+
+#ifdef PNG_WRITE_SHIFT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_shift,(png_row_infop row_info,
+    png_bytep row, png_const_color_8p bit_depth),PNG_EMPTY);
+#endif
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+    defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_compose,(png_row_infop row_info,
+    png_bytep row, png_structrp png_ptr),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_gamma,(png_row_infop row_info,
+    png_bytep row, png_structrp png_ptr),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_encode_alpha,(png_row_infop row_info,
+   png_bytep row, png_structrp png_ptr),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_expand_palette,(png_row_infop row_info,
+    png_bytep row, png_const_colorp palette, png_const_bytep trans,
+    int num_trans),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_do_expand,(png_row_infop row_info,
+    png_bytep row, png_const_color_16p trans_color),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_expand_16,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+/* The following decodes the appropriate chunks, and does error correction,
+ * then calls the appropriate callback for the chunk if it is valid.
+ */
+
+/* Decode the IHDR chunk */
+PNG_INTERNAL_FUNCTION(void,png_handle_IHDR,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_handle_PLTE,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_handle_IEND,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+
+#ifdef PNG_READ_bKGD_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_bKGD,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_cHRM_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_cHRM,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_gAMA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_gAMA,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_hIST_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_hIST,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_iCCP_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_iCCP,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif /* PNG_READ_iCCP_SUPPORTED */
+
+#ifdef PNG_READ_iTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_iTXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_oFFs_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_oFFs,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_pCAL_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_pCAL,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_pHYs_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_pHYs,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_sBIT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_sBIT,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_sCAL_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_sCAL,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_sPLT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_sPLT,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif /* PNG_READ_sPLT_SUPPORTED */
+
+#ifdef PNG_READ_sRGB_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_sRGB,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_tEXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_tEXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_tIME_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_tIME,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_tRNS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_tRNS,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_zTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_zTXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+#endif
+
+PNG_INTERNAL_FUNCTION(void,png_check_chunk_name,(png_structrp png_ptr,
+    png_uint_32 chunk_name),PNG_EMPTY);
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_handle_unknown,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length, int keep),PNG_EMPTY);
+   /* This is the function that gets called for unknown chunks.  The 'keep'
+    * argument is either non-zero for a known chunk that has been set to be
+    * handled as unknown or zero for an unknown chunk.  By default the function
+    * just skips the chunk or errors out if it is critical.
+    */
+
+#if defined(PNG_READ_UNKNOWN_CHUNKS_SUPPORTED) ||\
+   defined(PNG_HANDLE_AS_UNKNOWN_SUPPORTED)
+PNG_INTERNAL_FUNCTION(int,png_chunk_unknown_handling,
+    (png_const_structrp png_ptr, png_uint_32 chunk_name),PNG_EMPTY);
+   /* Exactly as the API png_handle_as_unknown() except that the argument is a
+    * 32-bit chunk name, not a string.
+    */
+#endif /* READ_UNKNOWN_CHUNKS || HANDLE_AS_UNKNOWN */
+#endif /* PNG_SET_UNKNOWN_CHUNKS_SUPPORTED */
+
+/* Handle the transformations for reading and writing */
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_read_transformations,(png_structrp png_ptr,
+   png_row_infop row_info),PNG_EMPTY);
+#endif
+#ifdef PNG_WRITE_TRANSFORMS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_write_transformations,(png_structrp png_ptr,
+   png_row_infop row_info),PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_init_read_transformations,(png_structrp png_ptr),
+    PNG_EMPTY);
+#endif
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_push_read_chunk,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_sig,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_check_crc,(png_structrp png_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_crc_skip,(png_structrp png_ptr,
+    png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_crc_finish,(png_structrp png_ptr),
+    PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_save_buffer,(png_structrp png_ptr),
+    PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_restore_buffer,(png_structrp png_ptr,
+    png_bytep buffer, png_size_t buffer_length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_IDAT,(png_structrp png_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_process_IDAT_data,(png_structrp png_ptr,
+    png_bytep buffer, png_size_t buffer_length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_process_row,(png_structrp png_ptr),
+    PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_handle_unknown,(png_structrp png_ptr,
+   png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_have_info,(png_structrp png_ptr,
+   png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_have_end,(png_structrp png_ptr,
+   png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_have_row,(png_structrp png_ptr,
+     png_bytep row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_end,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_process_some_data,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_push_finish_row,(png_structrp png_ptr),
+    PNG_EMPTY);
+#  ifdef PNG_READ_tEXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_push_handle_tEXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_tEXt,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+#  endif
+#  ifdef PNG_READ_zTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_push_handle_zTXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_zTXt,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+#  endif
+#  ifdef PNG_READ_iTXt_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_push_handle_iTXt,(png_structrp png_ptr,
+    png_inforp info_ptr, png_uint_32 length),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_push_read_iTXt,(png_structrp png_ptr,
+    png_inforp info_ptr),PNG_EMPTY);
+#  endif
+
+#endif /* PNG_PROGRESSIVE_READ_SUPPORTED */
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_do_read_intrapixel,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_do_write_intrapixel,(png_row_infop row_info,
+    png_bytep row),PNG_EMPTY);
+#endif
+
+/* Added at libpng version 1.6.0 */
+#ifdef PNG_GAMMA_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_colorspace_set_gamma,(png_const_structrp png_ptr,
+    png_colorspacerp colorspace, png_fixed_point gAMA), PNG_EMPTY);
+   /* Set the colorspace gamma with a value provided by the application or by
+    * the gAMA chunk on read.  The value will override anything set by an ICC
+    * profile.
+    */
+
+PNG_INTERNAL_FUNCTION(void,png_colorspace_sync_info,(png_const_structrp png_ptr,
+    png_inforp info_ptr), PNG_EMPTY);
+    /* Synchronize the info 'valid' flags with the colorspace */
+
+PNG_INTERNAL_FUNCTION(void,png_colorspace_sync,(png_const_structrp png_ptr,
+    png_inforp info_ptr), PNG_EMPTY);
+    /* Copy the png_struct colorspace to the info_struct and call the above to
+     * synchronize the flags.  Checks for NULL info_ptr and does nothing.
+     */
+#endif
+
+/* Added at libpng version 1.4.0 */
+#ifdef PNG_COLORSPACE_SUPPORTED
+/* These internal functions are for maintaining the colorspace structure within
+ * a png_info or png_struct (or, indeed, both).
+ */
+PNG_INTERNAL_FUNCTION(int,png_colorspace_set_chromaticities,
+   (png_const_structrp png_ptr, png_colorspacerp colorspace, const png_xy *xy,
+    int preferred), PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(int,png_colorspace_set_endpoints,
+   (png_const_structrp png_ptr, png_colorspacerp colorspace, const png_XYZ *XYZ,
+    int preferred), PNG_EMPTY);
+
+#ifdef PNG_sRGB_SUPPORTED
+PNG_INTERNAL_FUNCTION(int,png_colorspace_set_sRGB,(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, int intent), PNG_EMPTY);
+   /* This does set the colorspace gAMA and cHRM values too, but doesn't set the
+    * flags to write them, if it returns false there was a problem and an error
+    * message has already been output (but the colorspace may still need to be
+    * synced to record the invalid flag).
+    */
+#endif /* sRGB */
+
+#ifdef PNG_iCCP_SUPPORTED
+PNG_INTERNAL_FUNCTION(int,png_colorspace_set_ICC,(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_const_charp name,
+   png_uint_32 profile_length, png_const_bytep profile, int color_type),
+   PNG_EMPTY);
+   /* The 'name' is used for information only */
+
+/* Routines for checking parts of an ICC profile. */
+PNG_INTERNAL_FUNCTION(int,png_icc_check_length,(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_const_charp name,
+   png_uint_32 profile_length), PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(int,png_icc_check_header,(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_const_charp name,
+   png_uint_32 profile_length,
+   png_const_bytep profile /* first 132 bytes only */, int color_type),
+   PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(int,png_icc_check_tag_table,(png_const_structrp png_ptr,
+   png_colorspacerp colorspace, png_const_charp name,
+   png_uint_32 profile_length,
+   png_const_bytep profile /* header plus whole tag table */), PNG_EMPTY);
+#ifdef PNG_sRGB_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_icc_set_sRGB,(
+   png_const_structrp png_ptr, png_colorspacerp colorspace,
+   png_const_bytep profile, uLong adler), PNG_EMPTY);
+   /* 'adler' is the Adler32 checksum of the uncompressed profile data. It may
+    * be zero to indicate that it is not available.  It is used, if provided,
+    * as a fast check on the profile when checking to see if it is sRGB.
+    */
+#endif
+#endif /* iCCP */
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_colorspace_set_rgb_coefficients,
+   (png_structrp png_ptr), PNG_EMPTY);
+   /* Set the rgb_to_gray coefficients from the colorspace Y values */
+#endif /* READ_RGB_TO_GRAY */
+#endif /* COLORSPACE */
+
+/* Added at libpng version 1.4.0 */
+PNG_INTERNAL_FUNCTION(void,png_check_IHDR,(png_const_structrp png_ptr,
+    png_uint_32 width, png_uint_32 height, int bit_depth,
+    int color_type, int interlace_type, int compression_type,
+    int filter_type),PNG_EMPTY);
+
+/* Added at libpng version 1.5.10 */
+#if defined(PNG_READ_CHECK_FOR_INVALID_INDEX_SUPPORTED) || \
+    defined(PNG_WRITE_CHECK_FOR_INVALID_INDEX_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_do_check_palette_indexes,
+   (png_structrp png_ptr, png_row_infop row_info),PNG_EMPTY);
+#endif
+
+#if defined(PNG_FLOATING_POINT_SUPPORTED) && defined(PNG_ERROR_TEXT_SUPPORTED)
+PNG_INTERNAL_FUNCTION(void,png_fixed_error,(png_const_structrp png_ptr,
+   png_const_charp name),PNG_NORETURN);
+#endif
+
+/* Puts 'string' into 'buffer' at buffer[pos], taking care never to overwrite
+ * the end.  Always leaves the buffer nul terminated.  Never errors out (and
+ * there is no error code.)
+ */
+PNG_INTERNAL_FUNCTION(size_t,png_safecat,(png_charp buffer, size_t bufsize,
+   size_t pos, png_const_charp string),PNG_EMPTY);
+
+/* Various internal functions to handle formatted warning messages, currently
+ * only implemented for warnings.
+ */
+#if defined(PNG_WARNINGS_SUPPORTED) || defined(PNG_TIME_RFC1123_SUPPORTED)
+/* Utility to dump an unsigned value into a buffer, given a start pointer and
+ * and end pointer (which should point just *beyond* the end of the buffer!)
+ * Returns the pointer to the start of the formatted string.  This utility only
+ * does unsigned values.
+ */
+PNG_INTERNAL_FUNCTION(png_charp,png_format_number,(png_const_charp start,
+   png_charp end, int format, png_alloc_size_t number),PNG_EMPTY);
+
+/* Convenience macro that takes an array: */
+#define PNG_FORMAT_NUMBER(buffer,format,number) \
+   png_format_number(buffer, buffer + (sizeof buffer), format, number)
+
+/* Suggested size for a number buffer (enough for 64 bits and a sign!) */
+#define PNG_NUMBER_BUFFER_SIZE 24
+
+/* These are the integer formats currently supported, the name is formed from
+ * the standard printf(3) format string.
+ */
+#define PNG_NUMBER_FORMAT_u     1 /* chose unsigned API! */
+#define PNG_NUMBER_FORMAT_02u   2
+#define PNG_NUMBER_FORMAT_d     1 /* chose signed API! */
+#define PNG_NUMBER_FORMAT_02d   2
+#define PNG_NUMBER_FORMAT_x     3
+#define PNG_NUMBER_FORMAT_02x   4
+#define PNG_NUMBER_FORMAT_fixed 5 /* choose the signed API */
+#endif
+
+#ifdef PNG_WARNINGS_SUPPORTED
+/* New defines and members adding in libpng-1.5.4 */
+#  define PNG_WARNING_PARAMETER_SIZE 32
+#  define PNG_WARNING_PARAMETER_COUNT 8 /* Maximum 9; see pngerror.c */
+
+/* An l-value of this type has to be passed to the APIs below to cache the
+ * values of the parameters to a formatted warning message.
+ */
+typedef char png_warning_parameters[PNG_WARNING_PARAMETER_COUNT][
+   PNG_WARNING_PARAMETER_SIZE];
+
+PNG_INTERNAL_FUNCTION(void,png_warning_parameter,(png_warning_parameters p,
+   int number, png_const_charp string),PNG_EMPTY);
+   /* Parameters are limited in size to PNG_WARNING_PARAMETER_SIZE characters,
+    * including the trailing '\0'.
+    */
+PNG_INTERNAL_FUNCTION(void,png_warning_parameter_unsigned,
+   (png_warning_parameters p, int number, int format, png_alloc_size_t value),
+   PNG_EMPTY);
+   /* Use png_alloc_size_t because it is an unsigned type as big as any we
+    * need to output.  Use the following for a signed value.
+    */
+PNG_INTERNAL_FUNCTION(void,png_warning_parameter_signed,
+   (png_warning_parameters p, int number, int format, png_int_32 value),
+   PNG_EMPTY);
+
+PNG_INTERNAL_FUNCTION(void,png_formatted_warning,(png_const_structrp png_ptr,
+   png_warning_parameters p, png_const_charp message),PNG_EMPTY);
+   /* 'message' follows the X/Open approach of using @1, @2 to insert
+    * parameters previously supplied using the above functions.  Errors in
+    * specifying the parameters will simply result in garbage substitutions.
+    */
+#endif
+
+#ifdef PNG_BENIGN_ERRORS_SUPPORTED
+/* Application errors (new in 1.6); use these functions (declared below) for
+ * errors in the parameters or order of API function calls on read.  The
+ * 'warning' should be used for an error that can be handled completely; the
+ * 'error' for one which can be handled safely but which may lose application
+ * information or settings.
+ *
+ * By default these both result in a png_error call prior to release, while in a
+ * released version the 'warning' is just a warning.  However if the application
+ * explicitly disables benign errors (explicitly permitting the code to lose
+ * information) they both turn into warnings.
+ *
+ * If benign errors aren't supported they end up as the corresponding base call
+ * (png_warning or png_error.)
+ */
+PNG_INTERNAL_FUNCTION(void,png_app_warning,(png_const_structrp png_ptr,
+   png_const_charp message),PNG_EMPTY);
+   /* The application provided invalid parameters to an API function or called
+    * an API function at the wrong time, libpng can completely recover.
+    */
+
+PNG_INTERNAL_FUNCTION(void,png_app_error,(png_const_structrp png_ptr,
+   png_const_charp message),PNG_EMPTY);
+   /* As above but libpng will ignore the call, or attempt some other partial
+    * recovery from the error.
+    */
+#else
+#  define png_app_warning(pp,s) png_warning(pp,s)
+#  define png_app_error(pp,s) png_error(pp,s)
+#endif
+
+PNG_INTERNAL_FUNCTION(void,png_chunk_report,(png_const_structrp png_ptr,
+   png_const_charp message, int error),PNG_EMPTY);
+   /* Report a recoverable issue in chunk data.  On read this is used to report
+    * a problem found while reading a particular chunk and the
+    * png_chunk_benign_error or png_chunk_warning function is used as
+    * appropriate.  On write this is used to report an error that comes from
+    * data set via an application call to a png_set_ API and png_app_error or
+    * png_app_warning is used as appropriate.
+    *
+    * The 'error' parameter must have one of the following values:
+    */
+#define PNG_CHUNK_WARNING     0 /* never an error */
+#define PNG_CHUNK_WRITE_ERROR 1 /* an error only on write */
+#define PNG_CHUNK_ERROR       2 /* always an error */
+
+/* ASCII to FP interfaces, currently only implemented if sCAL
+ * support is required.
+ */
+#if defined(PNG_sCAL_SUPPORTED)
+/* MAX_DIGITS is actually the maximum number of characters in an sCAL
+ * width or height, derived from the precision (number of significant
+ * digits - a build time settable option) and assumptions about the
+ * maximum ridiculous exponent.
+ */
+#define PNG_sCAL_MAX_DIGITS (PNG_sCAL_PRECISION+1/*.*/+1/*E*/+10/*exponent*/)
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_ascii_from_fp,(png_const_structrp png_ptr,
+   png_charp ascii, png_size_t size, double fp, unsigned int precision),
+   PNG_EMPTY);
+#endif /* FLOATING_POINT */
+
+#ifdef PNG_FIXED_POINT_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_ascii_from_fixed,(png_const_structrp png_ptr,
+   png_charp ascii, png_size_t size, png_fixed_point fp),PNG_EMPTY);
+#endif /* FIXED_POINT */
+#endif /* sCAL */
+
+#if defined(PNG_sCAL_SUPPORTED) || defined(PNG_pCAL_SUPPORTED)
+/* An internal API to validate the format of a floating point number.
+ * The result is the index of the next character.  If the number is
+ * not valid it will be the index of a character in the supposed number.
+ *
+ * The format of a number is defined in the PNG extensions specification
+ * and this API is strictly conformant to that spec, not anyone elses!
+ *
+ * The format as a regular expression is:
+ *
+ * [+-]?[0-9]+.?([Ee][+-]?[0-9]+)?
+ *
+ * or:
+ *
+ * [+-]?.[0-9]+(.[0-9]+)?([Ee][+-]?[0-9]+)?
+ *
+ * The complexity is that either integer or fraction must be present and the
+ * fraction is permitted to have no digits only if the integer is present.
+ *
+ * NOTE: The dangling E problem.
+ *   There is a PNG valid floating point number in the following:
+ *
+ *       PNG floating point numbers are not greedy.
+ *
+ *   Working this out requires *TWO* character lookahead (because of the
+ *   sign), the parser does not do this - it will fail at the 'r' - this
+ *   doesn't matter for PNG sCAL chunk values, but it requires more care
+ *   if the value were ever to be embedded in something more complex.  Use
+ *   ANSI-C strtod if you need the lookahead.
+ */
+/* State table for the parser. */
+#define PNG_FP_INTEGER    0  /* before or in integer */
+#define PNG_FP_FRACTION   1  /* before or in fraction */
+#define PNG_FP_EXPONENT   2  /* before or in exponent */
+#define PNG_FP_STATE      3  /* mask for the above */
+#define PNG_FP_SAW_SIGN   4  /* Saw +/- in current state */
+#define PNG_FP_SAW_DIGIT  8  /* Saw a digit in current state */
+#define PNG_FP_SAW_DOT   16  /* Saw a dot in current state */
+#define PNG_FP_SAW_E     32  /* Saw an E (or e) in current state */
+#define PNG_FP_SAW_ANY   60  /* Saw any of the above 4 */
+
+/* These three values don't affect the parser.  They are set but not used.
+ */
+#define PNG_FP_WAS_VALID 64  /* Preceding substring is a valid fp number */
+#define PNG_FP_NEGATIVE 128  /* A negative number, including "-0" */
+#define PNG_FP_NONZERO  256  /* A non-zero value */
+#define PNG_FP_STICKY   448  /* The above three flags */
+
+/* This is available for the caller to store in 'state' if required.  Do not
+ * call the parser after setting it (the parser sometimes clears it.)
+ */
+#define PNG_FP_INVALID  512  /* Available for callers as a distinct value */
+
+/* Result codes for the parser (boolean - true meants ok, false means
+ * not ok yet.)
+ */
+#define PNG_FP_MAYBE      0  /* The number may be valid in the future */
+#define PNG_FP_OK         1  /* The number is valid */
+
+/* Tests on the sticky non-zero and negative flags.  To pass these checks
+ * the state must also indicate that the whole number is valid - this is
+ * achieved by testing PNG_FP_SAW_DIGIT (see the implementation for why this
+ * is equivalent to PNG_FP_OK above.)
+ */
+#define PNG_FP_NZ_MASK (PNG_FP_SAW_DIGIT | PNG_FP_NEGATIVE | PNG_FP_NONZERO)
+   /* NZ_MASK: the string is valid and a non-zero negative value */
+#define PNG_FP_Z_MASK (PNG_FP_SAW_DIGIT | PNG_FP_NONZERO)
+   /* Z MASK: the string is valid and a non-zero value. */
+   /* PNG_FP_SAW_DIGIT: the string is valid. */
+#define PNG_FP_IS_ZERO(state) (((state) & PNG_FP_Z_MASK) == PNG_FP_SAW_DIGIT)
+#define PNG_FP_IS_POSITIVE(state) (((state) & PNG_FP_NZ_MASK) == PNG_FP_Z_MASK)
+#define PNG_FP_IS_NEGATIVE(state) (((state) & PNG_FP_NZ_MASK) == PNG_FP_NZ_MASK)
+
+/* The actual parser.  This can be called repeatedly. It updates
+ * the index into the string and the state variable (which must
+ * be initialized to 0).  It returns a result code, as above.  There
+ * is no point calling the parser any more if it fails to advance to
+ * the end of the string - it is stuck on an invalid character (or
+ * terminated by '\0').
+ *
+ * Note that the pointer will consume an E or even an E+ and then leave
+ * a 'maybe' state even though a preceding integer.fraction is valid.
+ * The PNG_FP_WAS_VALID flag indicates that a preceding substring was
+ * a valid number.  It's possible to recover from this by calling
+ * the parser again (from the start, with state 0) but with a string
+ * that omits the last character (i.e. set the size to the index of
+ * the problem character.)  This has not been tested within libpng.
+ */
+PNG_INTERNAL_FUNCTION(int,png_check_fp_number,(png_const_charp string,
+   png_size_t size, int *statep, png_size_tp whereami),PNG_EMPTY);
+
+/* This is the same but it checks a complete string and returns true
+ * only if it just contains a floating point number.  As of 1.5.4 this
+ * function also returns the state at the end of parsing the number if
+ * it was valid (otherwise it returns 0.)  This can be used for testing
+ * for negative or zero values using the sticky flag.
+ */
+PNG_INTERNAL_FUNCTION(int,png_check_fp_string,(png_const_charp string,
+   png_size_t size),PNG_EMPTY);
+#endif /* pCAL || sCAL */
+
+#if defined(PNG_READ_GAMMA_SUPPORTED) ||\
+    defined(PNG_INCH_CONVERSIONS_SUPPORTED) || defined(PNG_READ_pHYs_SUPPORTED)
+/* Added at libpng version 1.5.0 */
+/* This is a utility to provide a*times/div (rounded) and indicate
+ * if there is an overflow.  The result is a boolean - false (0)
+ * for overflow, true (1) if no overflow, in which case *res
+ * holds the result.
+ */
+PNG_INTERNAL_FUNCTION(int,png_muldiv,(png_fixed_point_p res, png_fixed_point a,
+   png_int_32 multiplied_by, png_int_32 divided_by),PNG_EMPTY);
+#endif
+
+#if defined(PNG_READ_GAMMA_SUPPORTED) || defined(PNG_INCH_CONVERSIONS_SUPPORTED)
+/* Same deal, but issue a warning on overflow and return 0. */
+PNG_INTERNAL_FUNCTION(png_fixed_point,png_muldiv_warn,
+   (png_const_structrp png_ptr, png_fixed_point a, png_int_32 multiplied_by,
+   png_int_32 divided_by),PNG_EMPTY);
+#endif
+
+#ifdef PNG_GAMMA_SUPPORTED
+/* Calculate a reciprocal - used for gamma values.  This returns
+ * 0 if the argument is 0 in order to maintain an undefined value;
+ * there are no warnings.
+ */
+PNG_INTERNAL_FUNCTION(png_fixed_point,png_reciprocal,(png_fixed_point a),
+   PNG_EMPTY);
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* The same but gives a reciprocal of the product of two fixed point
+ * values.  Accuracy is suitable for gamma calculations but this is
+ * not exact - use png_muldiv for that.  Only required at present on read.
+ */
+PNG_INTERNAL_FUNCTION(png_fixed_point,png_reciprocal2,(png_fixed_point a,
+   png_fixed_point b),PNG_EMPTY);
+#endif
+
+/* Return true if the gamma value is significantly different from 1.0 */
+PNG_INTERNAL_FUNCTION(int,png_gamma_significant,(png_fixed_point gamma_value),
+   PNG_EMPTY);
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* Internal fixed point gamma correction.  These APIs are called as
+ * required to convert single values - they don't need to be fast,
+ * they are not used when processing image pixel values.
+ *
+ * While the input is an 'unsigned' value it must actually be the
+ * correct bit value - 0..255 or 0..65535 as required.
+ */
+PNG_INTERNAL_FUNCTION(png_uint_16,png_gamma_correct,(png_structrp png_ptr,
+   unsigned int value, png_fixed_point gamma_value),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(png_uint_16,png_gamma_16bit_correct,(unsigned int value,
+   png_fixed_point gamma_value),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(png_byte,png_gamma_8bit_correct,(unsigned int value,
+   png_fixed_point gamma_value),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_destroy_gamma_table,(png_structrp png_ptr),
+   PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_build_gamma_table,(png_structrp png_ptr,
+   int bit_depth),PNG_EMPTY);
+#endif
+
+/* SIMPLIFIED READ/WRITE SUPPORT */
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) ||\
+   defined(PNG_SIMPLIFIED_WRITE_SUPPORTED)
+/* The internal structure that png_image::opaque points to. */
+typedef struct png_control
+{
+   png_structp png_ptr;
+   png_infop   info_ptr;
+   png_voidp   error_buf;           /* Always a jmp_buf at present. */
+
+   png_const_bytep memory;          /* Memory buffer. */
+   png_size_t      size;            /* Size of the memory buffer. */
+
+   unsigned int for_write       :1; /* Otherwise it is a read structure */
+   unsigned int owned_file      :1; /* We own the file in io_ptr */
+} png_control;
+
+/* Return the pointer to the jmp_buf from a png_control: necessary because C
+ * does not reveal the type of the elements of jmp_buf.
+ */
+#ifdef __cplusplus
+#  define png_control_jmp_buf(pc) (((jmp_buf*)((pc)->error_buf))[0])
+#else
+#  define png_control_jmp_buf(pc) ((pc)->error_buf)
+#endif
+
+/* Utility to safely execute a piece of libpng code catching and logging any
+ * errors that might occur.  Returns true on success, false on failure (either
+ * of the function or as a result of a png_error.)
+ */
+PNG_INTERNAL_FUNCTION(void,png_safe_error,(png_structp png_ptr,
+   png_const_charp error_message),PNG_NORETURN);
+
+#ifdef PNG_WARNINGS_SUPPORTED
+PNG_INTERNAL_FUNCTION(void,png_safe_warning,(png_structp png_ptr,
+   png_const_charp warning_message),PNG_EMPTY);
+#else
+#  define png_safe_warning 0/*dummy argument*/
+#endif
+
+PNG_INTERNAL_FUNCTION(int,png_safe_execute,(png_imagep image,
+   int (*function)(png_voidp), png_voidp arg),PNG_EMPTY);
+
+/* Utility to log an error; this also cleans up the png_image; the function
+ * always returns 0 (false).
+ */
+PNG_INTERNAL_FUNCTION(int,png_image_error,(png_imagep image,
+   png_const_charp error_message),PNG_EMPTY);
+
+#ifndef PNG_SIMPLIFIED_READ_SUPPORTED
+/* png_image_free is used by the write code but not exported */
+PNG_INTERNAL_FUNCTION(void, png_image_free, (png_imagep image), PNG_EMPTY);
+#endif /* !SIMPLIFIED_READ */
+
+#endif /* SIMPLIFIED READ/WRITE */
+
+/* These are initialization functions for hardware specific PNG filter
+ * optimizations; list these here then select the appropriate one at compile
+ * time using the macro PNG_FILTER_OPTIMIZATIONS.  If the macro is not defined
+ * the generic code is used.
+ */
+#ifdef PNG_FILTER_OPTIMIZATIONS
+PNG_INTERNAL_FUNCTION(void, PNG_FILTER_OPTIMIZATIONS, (png_structp png_ptr,
+   unsigned int bpp), PNG_EMPTY);
+   /* Just declare the optimization that will be used */
+#else
+   /* List *all* the possible optimizations here - this branch is required if
+    * the builder of libpng passes the definition of PNG_FILTER_OPTIMIZATIONS in
+    * CFLAGS in place of CPPFLAGS *and* uses symbol prefixing.
+    */
+PNG_INTERNAL_FUNCTION(void, png_init_filter_functions_neon,
+   (png_structp png_ptr, unsigned int bpp), PNG_EMPTY);
+#endif
+
+/* Maintainer: Put new private prototypes here ^ */
+
+#include "pngdebug.h"
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* PNG_VERSION_INFO_ONLY */
+#endif /* PNGPRIV_H */
diff --git a/ml/dlib/dlib/external/libpng/pngread.c b/ml/dlib/dlib/external/libpng/pngread.c
new file mode 100644
index 000000000..8f96ca23e
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngread.c
@@ -0,0 +1,4000 @@
+
+/* pngread.c - read a PNG file
+ *
+ * Last changed in libpng 1.6.1 [March 28, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file contains routines that an application calls directly to
+ * read a PNG file or stream.
+ */
+
+#include "pngpriv.h"
+#if defined(PNG_SIMPLIFIED_READ_SUPPORTED) && defined(PNG_STDIO_SUPPORTED)
+#  include <errno.h>
+#endif
+
+#ifdef PNG_READ_SUPPORTED
+
+/* Create a PNG structure for reading, and allocate any memory needed. */
+PNG_FUNCTION(png_structp,PNGAPI
+png_create_read_struct,(png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn),PNG_ALLOCATED)
+{
+#ifndef PNG_USER_MEM_SUPPORTED
+   png_structp png_ptr = png_create_png_struct(user_png_ver, error_ptr,
+      error_fn, warn_fn, NULL, NULL, NULL);
+#else
+   return png_create_read_struct_2(user_png_ver, error_ptr, error_fn,
+       warn_fn, NULL, NULL, NULL);
+}
+
+/* Alternate create PNG structure for reading, and allocate any memory
+ * needed.
+ */
+PNG_FUNCTION(png_structp,PNGAPI
+png_create_read_struct_2,(png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn, png_voidp mem_ptr,
+    png_malloc_ptr malloc_fn, png_free_ptr free_fn),PNG_ALLOCATED)
+{
+   png_structp png_ptr = png_create_png_struct(user_png_ver, error_ptr,
+      error_fn, warn_fn, mem_ptr, malloc_fn, free_fn);
+#endif /* PNG_USER_MEM_SUPPORTED */
+
+   if (png_ptr != NULL)
+   {
+      png_ptr->mode = PNG_IS_READ_STRUCT;
+
+      /* Added in libpng-1.6.0; this can be used to detect a read structure if
+       * required (it will be zero in a write structure.)
+       */
+#     ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+         png_ptr->IDAT_read_size = PNG_IDAT_READ_SIZE;
+#     endif
+
+#     ifdef PNG_BENIGN_READ_ERRORS_SUPPORTED
+         png_ptr->flags |= PNG_FLAG_BENIGN_ERRORS_WARN;
+
+         /* In stable builds only warn if an application error can be completely
+          * handled.
+          */
+#        if PNG_LIBPNG_BUILD_BASE_TYPE >= PNG_LIBPNG_BUILD_RC
+            png_ptr->flags |= PNG_FLAG_APP_WARNINGS_WARN;
+#        endif
+#     endif
+
+      /* TODO: delay this, it can be done in png_init_io (if the app doesn't
+       * do it itself) avoiding setting the default function if it is not
+       * required.
+       */
+      png_set_read_fn(png_ptr, NULL, NULL);
+   }
+
+   return png_ptr;
+}
+
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the information before the actual image data.  This has been
+ * changed in v0.90 to allow reading a file that already has the magic
+ * bytes read from the stream.  You can tell libpng how many bytes have
+ * been read from the beginning of the stream (up to the maximum of 8)
+ * via png_set_sig_bytes(), and we will only check the remaining bytes
+ * here.  The application can then have access to the signature bytes we
+ * read if it is determined that this isn't a valid PNG file.
+ */
+void PNGAPI
+png_read_info(png_structrp png_ptr, png_inforp info_ptr)
+{
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+   int keep;
+#endif
+
+   png_debug(1, "in png_read_info");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   /* Read and check the PNG file signature. */
+   png_read_sig(png_ptr, info_ptr);
+
+   for (;;)
+   {
+      png_uint_32 length = png_read_chunk_header(png_ptr);
+      png_uint_32 chunk_name = png_ptr->chunk_name;
+
+      /* IDAT logic needs to happen here to simplify getting the two flags
+       * right.
+       */
+      if (chunk_name == png_IDAT)
+      {
+         if (!(png_ptr->mode & PNG_HAVE_IHDR))
+            png_chunk_error(png_ptr, "Missing IHDR before IDAT");
+
+         else if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE &&
+             !(png_ptr->mode & PNG_HAVE_PLTE))
+            png_chunk_error(png_ptr, "Missing PLTE before IDAT");
+
+         else if (png_ptr->mode & PNG_AFTER_IDAT)
+            png_chunk_benign_error(png_ptr, "Too many IDATs found");
+
+         png_ptr->mode |= PNG_HAVE_IDAT;
+      }
+
+      else if (png_ptr->mode & PNG_HAVE_IDAT)
+         png_ptr->mode |= PNG_AFTER_IDAT;
+
+      /* This should be a binary subdivision search or a hash for
+       * matching the chunk name rather than a linear search.
+       */
+      if (chunk_name == png_IHDR)
+         png_handle_IHDR(png_ptr, info_ptr, length);
+
+      else if (chunk_name == png_IEND)
+         png_handle_IEND(png_ptr, info_ptr, length);
+
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+      else if ((keep = png_chunk_unknown_handling(png_ptr, chunk_name)) != 0)
+      {
+         png_handle_unknown(png_ptr, info_ptr, length, keep);
+
+         if (chunk_name == png_PLTE)
+            png_ptr->mode |= PNG_HAVE_PLTE;
+
+         else if (chunk_name == png_IDAT)
+         {
+            png_ptr->idat_size = 0; /* It has been consumed */
+            break;
+         }
+      }
+#endif
+      else if (chunk_name == png_PLTE)
+         png_handle_PLTE(png_ptr, info_ptr, length);
+
+      else if (chunk_name == png_IDAT)
+      {
+         png_ptr->idat_size = length;
+         break;
+      }
+
+#ifdef PNG_READ_bKGD_SUPPORTED
+      else if (chunk_name == png_bKGD)
+         png_handle_bKGD(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_cHRM_SUPPORTED
+      else if (chunk_name == png_cHRM)
+         png_handle_cHRM(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_gAMA_SUPPORTED
+      else if (chunk_name == png_gAMA)
+         png_handle_gAMA(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_hIST_SUPPORTED
+      else if (chunk_name == png_hIST)
+         png_handle_hIST(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_oFFs_SUPPORTED
+      else if (chunk_name == png_oFFs)
+         png_handle_oFFs(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_pCAL_SUPPORTED
+      else if (chunk_name == png_pCAL)
+         png_handle_pCAL(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sCAL_SUPPORTED
+      else if (chunk_name == png_sCAL)
+         png_handle_sCAL(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_pHYs_SUPPORTED
+      else if (chunk_name == png_pHYs)
+         png_handle_pHYs(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sBIT_SUPPORTED
+      else if (chunk_name == png_sBIT)
+         png_handle_sBIT(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sRGB_SUPPORTED
+      else if (chunk_name == png_sRGB)
+         png_handle_sRGB(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_iCCP_SUPPORTED
+      else if (chunk_name == png_iCCP)
+         png_handle_iCCP(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sPLT_SUPPORTED
+      else if (chunk_name == png_sPLT)
+         png_handle_sPLT(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tEXt_SUPPORTED
+      else if (chunk_name == png_tEXt)
+         png_handle_tEXt(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tIME_SUPPORTED
+      else if (chunk_name == png_tIME)
+         png_handle_tIME(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tRNS_SUPPORTED
+      else if (chunk_name == png_tRNS)
+         png_handle_tRNS(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_zTXt_SUPPORTED
+      else if (chunk_name == png_zTXt)
+         png_handle_zTXt(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_iTXt_SUPPORTED
+      else if (chunk_name == png_iTXt)
+         png_handle_iTXt(png_ptr, info_ptr, length);
+#endif
+
+      else
+         png_handle_unknown(png_ptr, info_ptr, length,
+            PNG_HANDLE_CHUNK_AS_DEFAULT);
+   }
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+/* Optional call to update the users info_ptr structure */
+void PNGAPI
+png_read_update_info(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_debug(1, "in png_read_update_info");
+
+   if (png_ptr != NULL)
+   {
+      if ((png_ptr->flags & PNG_FLAG_ROW_INIT) == 0)
+      {
+         png_read_start_row(png_ptr);
+
+#        ifdef PNG_READ_TRANSFORMS_SUPPORTED
+            png_read_transform_info(png_ptr, info_ptr);
+#        else
+            PNG_UNUSED(info_ptr)
+#        endif
+      }
+
+      /* New in 1.6.0 this avoids the bug of doing the initializations twice */
+      else
+         png_app_error(png_ptr,
+            "png_read_update_info/png_start_read_image: duplicate call");
+   }
+}
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Initialize palette, background, etc, after transformations
+ * are set, but before any reading takes place.  This allows
+ * the user to obtain a gamma-corrected palette, for example.
+ * If the user doesn't call this, we will do it ourselves.
+ */
+void PNGAPI
+png_start_read_image(png_structrp png_ptr)
+{
+   png_debug(1, "in png_start_read_image");
+
+   if (png_ptr != NULL)
+   {
+      if ((png_ptr->flags & PNG_FLAG_ROW_INIT) == 0)
+         png_read_start_row(png_ptr);
+
+      /* New in 1.6.0 this avoids the bug of doing the initializations twice */
+      else
+         png_app_error(png_ptr,
+            "png_start_read_image/png_read_update_info: duplicate call");
+   }
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+void PNGAPI
+png_read_row(png_structrp png_ptr, png_bytep row, png_bytep dsp_row)
+{
+   png_row_info row_info;
+
+   if (png_ptr == NULL)
+      return;
+
+   png_debug2(1, "in png_read_row (row %lu, pass %d)",
+       (unsigned long)png_ptr->row_number, png_ptr->pass);
+
+   /* png_read_start_row sets the information (in particular iwidth) for this
+    * interlace pass.
+    */
+   if (!(png_ptr->flags & PNG_FLAG_ROW_INIT))
+      png_read_start_row(png_ptr);
+
+   /* 1.5.6: row_info moved out of png_struct to a local here. */
+   row_info.width = png_ptr->iwidth; /* NOTE: width of current interlaced row */
+   row_info.color_type = png_ptr->color_type;
+   row_info.bit_depth = png_ptr->bit_depth;
+   row_info.channels = png_ptr->channels;
+   row_info.pixel_depth = png_ptr->pixel_depth;
+   row_info.rowbytes = PNG_ROWBYTES(row_info.pixel_depth, row_info.width);
+
+   if (png_ptr->row_number == 0 && png_ptr->pass == 0)
+   {
+   /* Check for transforms that have been set but were defined out */
+#if defined(PNG_WRITE_INVERT_SUPPORTED) && !defined(PNG_READ_INVERT_SUPPORTED)
+   if (png_ptr->transformations & PNG_INVERT_MONO)
+      png_warning(png_ptr, "PNG_READ_INVERT_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_FILLER_SUPPORTED) && !defined(PNG_READ_FILLER_SUPPORTED)
+   if (png_ptr->transformations & PNG_FILLER)
+      png_warning(png_ptr, "PNG_READ_FILLER_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_PACKSWAP_SUPPORTED) && \
+    !defined(PNG_READ_PACKSWAP_SUPPORTED)
+   if (png_ptr->transformations & PNG_PACKSWAP)
+      png_warning(png_ptr, "PNG_READ_PACKSWAP_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_PACK_SUPPORTED) && !defined(PNG_READ_PACK_SUPPORTED)
+   if (png_ptr->transformations & PNG_PACK)
+      png_warning(png_ptr, "PNG_READ_PACK_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_SHIFT_SUPPORTED) && !defined(PNG_READ_SHIFT_SUPPORTED)
+   if (png_ptr->transformations & PNG_SHIFT)
+      png_warning(png_ptr, "PNG_READ_SHIFT_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_BGR_SUPPORTED) && !defined(PNG_READ_BGR_SUPPORTED)
+   if (png_ptr->transformations & PNG_BGR)
+      png_warning(png_ptr, "PNG_READ_BGR_SUPPORTED is not defined");
+#endif
+
+#if defined(PNG_WRITE_SWAP_SUPPORTED) && !defined(PNG_READ_SWAP_SUPPORTED)
+   if (png_ptr->transformations & PNG_SWAP_BYTES)
+      png_warning(png_ptr, "PNG_READ_SWAP_SUPPORTED is not defined");
+#endif
+   }
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* If interlaced and we do not need a new row, combine row and return.
+    * Notice that the pixels we have from previous rows have been transformed
+    * already; we can only combine like with like (transformed or
+    * untransformed) and, because of the libpng API for interlaced images, this
+    * means we must transform before de-interlacing.
+    */
+   if (png_ptr->interlaced && (png_ptr->transformations & PNG_INTERLACE))
+   {
+      switch (png_ptr->pass)
+      {
+         case 0:
+            if (png_ptr->row_number & 0x07)
+            {
+               if (dsp_row != NULL)
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 1:
+            if ((png_ptr->row_number & 0x07) || png_ptr->width < 5)
+            {
+               if (dsp_row != NULL)
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 2:
+            if ((png_ptr->row_number & 0x07) != 4)
+            {
+               if (dsp_row != NULL && (png_ptr->row_number & 4))
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 3:
+            if ((png_ptr->row_number & 3) || png_ptr->width < 3)
+            {
+               if (dsp_row != NULL)
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 4:
+            if ((png_ptr->row_number & 3) != 2)
+            {
+               if (dsp_row != NULL && (png_ptr->row_number & 2))
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 5:
+            if ((png_ptr->row_number & 1) || png_ptr->width < 2)
+            {
+               if (dsp_row != NULL)
+                  png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         default:
+         case 6:
+            if (!(png_ptr->row_number & 1))
+            {
+               png_read_finish_row(png_ptr);
+               return;
+            }
+            break;
+      }
+   }
+#endif
+
+   if (!(png_ptr->mode & PNG_HAVE_IDAT))
+      png_error(png_ptr, "Invalid attempt to read row data");
+
+   /* Fill the row with IDAT data: */
+   png_read_IDAT_data(png_ptr, png_ptr->row_buf, row_info.rowbytes + 1);
+
+   if (png_ptr->row_buf[0] > PNG_FILTER_VALUE_NONE)
+   {
+      if (png_ptr->row_buf[0] < PNG_FILTER_VALUE_LAST)
+         png_read_filter_row(png_ptr, &row_info, png_ptr->row_buf + 1,
+            png_ptr->prev_row + 1, png_ptr->row_buf[0]);
+      else
+         png_error(png_ptr, "bad adaptive filter value");
+   }
+
+   /* libpng 1.5.6: the following line was copying png_ptr->rowbytes before
+    * 1.5.6, while the buffer really is this big in current versions of libpng
+    * it may not be in the future, so this was changed just to copy the
+    * interlaced count:
+    */
+   memcpy(png_ptr->prev_row, png_ptr->row_buf, row_info.rowbytes + 1);
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) &&
+       (png_ptr->filter_type == PNG_INTRAPIXEL_DIFFERENCING))
+   {
+      /* Intrapixel differencing */
+      png_do_read_intrapixel(&row_info, png_ptr->row_buf + 1);
+   }
+#endif
+
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   if (png_ptr->transformations)
+      png_do_read_transformations(png_ptr, &row_info);
+#endif
+
+   /* The transformed pixel depth should match the depth now in row_info. */
+   if (png_ptr->transformed_pixel_depth == 0)
+   {
+      png_ptr->transformed_pixel_depth = row_info.pixel_depth;
+      if (row_info.pixel_depth > png_ptr->maximum_pixel_depth)
+         png_error(png_ptr, "sequential row overflow");
+   }
+
+   else if (png_ptr->transformed_pixel_depth != row_info.pixel_depth)
+      png_error(png_ptr, "internal sequential row size calculation error");
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* Blow up interlaced rows to full size */
+   if (png_ptr->interlaced &&
+      (png_ptr->transformations & PNG_INTERLACE))
+   {
+      if (png_ptr->pass < 6)
+         png_do_read_interlace(&row_info, png_ptr->row_buf + 1, png_ptr->pass,
+            png_ptr->transformations);
+
+      if (dsp_row != NULL)
+         png_combine_row(png_ptr, dsp_row, 1/*display*/);
+
+      if (row != NULL)
+         png_combine_row(png_ptr, row, 0/*row*/);
+   }
+
+   else
+#endif
+   {
+      if (row != NULL)
+         png_combine_row(png_ptr, row, -1/*ignored*/);
+
+      if (dsp_row != NULL)
+         png_combine_row(png_ptr, dsp_row, -1/*ignored*/);
+   }
+   png_read_finish_row(png_ptr);
+
+   if (png_ptr->read_row_fn != NULL)
+      (*(png_ptr->read_row_fn))(png_ptr, png_ptr->row_number, png_ptr->pass);
+
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read one or more rows of image data.  If the image is interlaced,
+ * and png_set_interlace_handling() has been called, the rows need to
+ * contain the contents of the rows from the previous pass.  If the
+ * image has alpha or transparency, and png_handle_alpha()[*] has been
+ * called, the rows contents must be initialized to the contents of the
+ * screen.
+ *
+ * "row" holds the actual image, and pixels are placed in it
+ * as they arrive.  If the image is displayed after each pass, it will
+ * appear to "sparkle" in.  "display_row" can be used to display a
+ * "chunky" progressive image, with finer detail added as it becomes
+ * available.  If you do not want this "chunky" display, you may pass
+ * NULL for display_row.  If you do not want the sparkle display, and
+ * you have not called png_handle_alpha(), you may pass NULL for rows.
+ * If you have called png_handle_alpha(), and the image has either an
+ * alpha channel or a transparency chunk, you must provide a buffer for
+ * rows.  In this case, you do not have to provide a display_row buffer
+ * also, but you may.  If the image is not interlaced, or if you have
+ * not called png_set_interlace_handling(), the display_row buffer will
+ * be ignored, so pass NULL to it.
+ *
+ * [*] png_handle_alpha() does not exist yet, as of this version of libpng
+ */
+
+void PNGAPI
+png_read_rows(png_structrp png_ptr, png_bytepp row,
+    png_bytepp display_row, png_uint_32 num_rows)
+{
+   png_uint_32 i;
+   png_bytepp rp;
+   png_bytepp dp;
+
+   png_debug(1, "in png_read_rows");
+
+   if (png_ptr == NULL)
+      return;
+
+   rp = row;
+   dp = display_row;
+   if (rp != NULL && dp != NULL)
+      for (i = 0; i < num_rows; i++)
+      {
+         png_bytep rptr = *rp++;
+         png_bytep dptr = *dp++;
+
+         png_read_row(png_ptr, rptr, dptr);
+      }
+
+   else if (rp != NULL)
+      for (i = 0; i < num_rows; i++)
+      {
+         png_bytep rptr = *rp;
+         png_read_row(png_ptr, rptr, NULL);
+         rp++;
+      }
+
+   else if (dp != NULL)
+      for (i = 0; i < num_rows; i++)
+      {
+         png_bytep dptr = *dp;
+         png_read_row(png_ptr, NULL, dptr);
+         dp++;
+      }
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the entire image.  If the image has an alpha channel or a tRNS
+ * chunk, and you have called png_handle_alpha()[*], you will need to
+ * initialize the image to the current image that PNG will be overlaying.
+ * We set the num_rows again here, in case it was incorrectly set in
+ * png_read_start_row() by a call to png_read_update_info() or
+ * png_start_read_image() if png_set_interlace_handling() wasn't called
+ * prior to either of these functions like it should have been.  You can
+ * only call this function once.  If you desire to have an image for
+ * each pass of a interlaced image, use png_read_rows() instead.
+ *
+ * [*] png_handle_alpha() does not exist yet, as of this version of libpng
+ */
+void PNGAPI
+png_read_image(png_structrp png_ptr, png_bytepp image)
+{
+   png_uint_32 i, image_height;
+   int pass, j;
+   png_bytepp rp;
+
+   png_debug(1, "in png_read_image");
+
+   if (png_ptr == NULL)
+      return;
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   if (!(png_ptr->flags & PNG_FLAG_ROW_INIT))
+   {
+      pass = png_set_interlace_handling(png_ptr);
+      /* And make sure transforms are initialized. */
+      png_start_read_image(png_ptr);
+   }
+   else
+   {
+      if (png_ptr->interlaced && !(png_ptr->transformations & PNG_INTERLACE))
+      {
+         /* Caller called png_start_read_image or png_read_update_info without
+          * first turning on the PNG_INTERLACE transform.  We can fix this here,
+          * but the caller should do it!
+          */
+         png_warning(png_ptr, "Interlace handling should be turned on when "
+            "using png_read_image");
+         /* Make sure this is set correctly */
+         png_ptr->num_rows = png_ptr->height;
+      }
+
+      /* Obtain the pass number, which also turns on the PNG_INTERLACE flag in
+       * the above error case.
+       */
+      pass = png_set_interlace_handling(png_ptr);
+   }
+#else
+   if (png_ptr->interlaced)
+      png_error(png_ptr,
+          "Cannot read interlaced image -- interlace handler disabled");
+
+   pass = 1;
+#endif
+
+   image_height=png_ptr->height;
+
+   for (j = 0; j < pass; j++)
+   {
+      rp = image;
+      for (i = 0; i < image_height; i++)
+      {
+         png_read_row(png_ptr, *rp, NULL);
+         rp++;
+      }
+   }
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+/* Read the end of the PNG file.  Will not read past the end of the
+ * file, will verify the end is accurate, and will read any comments
+ * or time information at the end of the file, if info is not NULL.
+ */
+void PNGAPI
+png_read_end(png_structrp png_ptr, png_inforp info_ptr)
+{
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+   int keep;
+#endif
+
+   png_debug(1, "in png_read_end");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* If png_read_end is called in the middle of reading the rows there may
+    * still be pending IDAT data and an owned zstream.  Deal with this here.
+    */
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+   if (!png_chunk_unknown_handling(png_ptr, png_IDAT))
+#endif
+      png_read_finish_IDAT(png_ptr);
+
+#ifdef PNG_READ_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   /* Report invalid palette index; added at libng-1.5.10 */
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE &&
+      png_ptr->num_palette_max > png_ptr->num_palette)
+     png_benign_error(png_ptr, "Read palette index exceeding num_palette");
+#endif
+
+   do
+   {
+      png_uint_32 length = png_read_chunk_header(png_ptr);
+      png_uint_32 chunk_name = png_ptr->chunk_name;
+
+      if (chunk_name == png_IHDR)
+         png_handle_IHDR(png_ptr, info_ptr, length);
+
+      else if (chunk_name == png_IEND)
+         png_handle_IEND(png_ptr, info_ptr, length);
+
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+      else if ((keep = png_chunk_unknown_handling(png_ptr, chunk_name)) != 0)
+      {
+         if (chunk_name == png_IDAT)
+         {
+            if ((length > 0) || (png_ptr->mode & PNG_HAVE_CHUNK_AFTER_IDAT))
+               png_benign_error(png_ptr, "Too many IDATs found");
+         }
+         png_handle_unknown(png_ptr, info_ptr, length, keep);
+         if (chunk_name == png_PLTE)
+            png_ptr->mode |= PNG_HAVE_PLTE;
+      }
+#endif
+
+      else if (chunk_name == png_IDAT)
+      {
+         /* Zero length IDATs are legal after the last IDAT has been
+          * read, but not after other chunks have been read.
+          */
+         if ((length > 0) || (png_ptr->mode & PNG_HAVE_CHUNK_AFTER_IDAT))
+            png_benign_error(png_ptr, "Too many IDATs found");
+
+         png_crc_finish(png_ptr, length);
+      }
+      else if (chunk_name == png_PLTE)
+         png_handle_PLTE(png_ptr, info_ptr, length);
+
+#ifdef PNG_READ_bKGD_SUPPORTED
+      else if (chunk_name == png_bKGD)
+         png_handle_bKGD(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_cHRM_SUPPORTED
+      else if (chunk_name == png_cHRM)
+         png_handle_cHRM(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_gAMA_SUPPORTED
+      else if (chunk_name == png_gAMA)
+         png_handle_gAMA(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_hIST_SUPPORTED
+      else if (chunk_name == png_hIST)
+         png_handle_hIST(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_oFFs_SUPPORTED
+      else if (chunk_name == png_oFFs)
+         png_handle_oFFs(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_pCAL_SUPPORTED
+      else if (chunk_name == png_pCAL)
+         png_handle_pCAL(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sCAL_SUPPORTED
+      else if (chunk_name == png_sCAL)
+         png_handle_sCAL(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_pHYs_SUPPORTED
+      else if (chunk_name == png_pHYs)
+         png_handle_pHYs(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sBIT_SUPPORTED
+      else if (chunk_name == png_sBIT)
+         png_handle_sBIT(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sRGB_SUPPORTED
+      else if (chunk_name == png_sRGB)
+         png_handle_sRGB(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_iCCP_SUPPORTED
+      else if (chunk_name == png_iCCP)
+         png_handle_iCCP(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_sPLT_SUPPORTED
+      else if (chunk_name == png_sPLT)
+         png_handle_sPLT(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tEXt_SUPPORTED
+      else if (chunk_name == png_tEXt)
+         png_handle_tEXt(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tIME_SUPPORTED
+      else if (chunk_name == png_tIME)
+         png_handle_tIME(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_tRNS_SUPPORTED
+      else if (chunk_name == png_tRNS)
+         png_handle_tRNS(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_zTXt_SUPPORTED
+      else if (chunk_name == png_zTXt)
+         png_handle_zTXt(png_ptr, info_ptr, length);
+#endif
+
+#ifdef PNG_READ_iTXt_SUPPORTED
+      else if (chunk_name == png_iTXt)
+         png_handle_iTXt(png_ptr, info_ptr, length);
+#endif
+
+      else
+         png_handle_unknown(png_ptr, info_ptr, length,
+            PNG_HANDLE_CHUNK_AS_DEFAULT);
+   } while (!(png_ptr->mode & PNG_HAVE_IEND));
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+/* Free all memory used in the read struct */
+static void
+png_read_destroy(png_structrp png_ptr)
+{
+   png_debug(1, "in png_read_destroy");
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   png_destroy_gamma_table(png_ptr);
+#endif
+
+   png_free(png_ptr, png_ptr->big_row_buf);
+   png_free(png_ptr, png_ptr->big_prev_row);
+   png_free(png_ptr, png_ptr->read_buffer);
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+   png_free(png_ptr, png_ptr->palette_lookup);
+   png_free(png_ptr, png_ptr->quantize_index);
+#endif
+
+   if (png_ptr->free_me & PNG_FREE_PLTE)
+      png_zfree(png_ptr, png_ptr->palette);
+   png_ptr->free_me &= ~PNG_FREE_PLTE;
+
+#if defined(PNG_tRNS_SUPPORTED) || \
+    defined(PNG_READ_EXPAND_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED)
+   if (png_ptr->free_me & PNG_FREE_TRNS)
+      png_free(png_ptr, png_ptr->trans_alpha);
+   png_ptr->free_me &= ~PNG_FREE_TRNS;
+#endif
+
+   inflateEnd(&png_ptr->zstream);
+
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+   png_free(png_ptr, png_ptr->save_buffer);
+#endif
+
+#if defined(PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED) &&\
+   defined(PNG_READ_UNKNOWN_CHUNKS_SUPPORTED)
+   png_free(png_ptr, png_ptr->unknown_chunk.data);
+#endif
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+   png_free(png_ptr, png_ptr->chunk_list);
+#endif
+
+   /* NOTE: the 'setjmp' buffer may still be allocated and the memory and error
+    * callbacks are still set at this point.  They are required to complete the
+    * destruction of the png_struct itself.
+    */
+}
+
+/* Free all memory used by the read */
+void PNGAPI
+png_destroy_read_struct(png_structpp png_ptr_ptr, png_infopp info_ptr_ptr,
+    png_infopp end_info_ptr_ptr)
+{
+   png_structrp png_ptr = NULL;
+
+   png_debug(1, "in png_destroy_read_struct");
+
+   if (png_ptr_ptr != NULL)
+      png_ptr = *png_ptr_ptr;
+
+   if (png_ptr == NULL)
+      return;
+
+   /* libpng 1.6.0: use the API to destroy info structs to ensure consistent
+    * behavior.  Prior to 1.6.0 libpng did extra 'info' destruction in this API.
+    * The extra was, apparently, unnecessary yet this hides memory leak bugs.
+    */
+   png_destroy_info_struct(png_ptr, end_info_ptr_ptr);
+   png_destroy_info_struct(png_ptr, info_ptr_ptr);
+
+   *png_ptr_ptr = NULL;
+   png_read_destroy(png_ptr);
+   png_destroy_png_struct(png_ptr);
+}
+
+void PNGAPI
+png_set_read_status_fn(png_structrp png_ptr, png_read_status_ptr read_row_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->read_row_fn = read_row_fn;
+}
+
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+void PNGAPI
+png_read_png(png_structrp png_ptr, png_inforp info_ptr,
+                           int transforms,
+                           voidp params)
+{
+   int row;
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   /* png_read_info() gives us all of the information from the
+    * PNG file before the first IDAT (image data chunk).
+    */
+   png_read_info(png_ptr, info_ptr);
+   if (info_ptr->height > PNG_UINT_32_MAX/(sizeof (png_bytep)))
+      png_error(png_ptr, "Image is too high to process with png_read_png()");
+
+   /* -------------- image transformations start here ------------------- */
+
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+   /* Tell libpng to strip 16-bit/color files down to 8 bits per color.
+    */
+   if (transforms & PNG_TRANSFORM_SCALE_16)
+   {
+     /* Added at libpng-1.5.4. "strip_16" produces the same result that it
+      * did in earlier versions, while "scale_16" is now more accurate.
+      */
+      png_set_scale_16(png_ptr);
+   }
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+   /* If both SCALE and STRIP are required pngrtran will effectively cancel the
+    * latter by doing SCALE first.  This is ok and allows apps not to check for
+    * which is supported to get the right answer.
+    */
+   if (transforms & PNG_TRANSFORM_STRIP_16)
+      png_set_strip_16(png_ptr);
+#endif
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+   /* Strip alpha bytes from the input data without combining with
+    * the background (not recommended).
+    */
+   if (transforms & PNG_TRANSFORM_STRIP_ALPHA)
+      png_set_strip_alpha(png_ptr);
+#endif
+
+#if defined(PNG_READ_PACK_SUPPORTED) && !defined(PNG_READ_EXPAND_SUPPORTED)
+   /* Extract multiple pixels with bit depths of 1, 2, or 4 from a single
+    * byte into separate bytes (useful for paletted and grayscale images).
+    */
+   if (transforms & PNG_TRANSFORM_PACKING)
+      png_set_packing(png_ptr);
+#endif
+
+#ifdef PNG_READ_PACKSWAP_SUPPORTED
+   /* Change the order of packed pixels to least significant bit first
+    * (not useful if you are using png_set_packing).
+    */
+   if (transforms & PNG_TRANSFORM_PACKSWAP)
+      png_set_packswap(png_ptr);
+#endif
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+   /* Expand paletted colors into true RGB triplets
+    * Expand grayscale images to full 8 bits from 1, 2, or 4 bits/pixel
+    * Expand paletted or RGB images with transparency to full alpha
+    * channels so the data will be available as RGBA quartets.
+    */
+   if (transforms & PNG_TRANSFORM_EXPAND)
+      if ((png_ptr->bit_depth < 8) ||
+          (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE) ||
+          (png_get_valid(png_ptr, info_ptr, PNG_INFO_tRNS)))
+         png_set_expand(png_ptr);
+#endif
+
+   /* We don't handle background color or gamma transformation or quantizing.
+    */
+
+#ifdef PNG_READ_INVERT_SUPPORTED
+   /* Invert monochrome files to have 0 as white and 1 as black
+    */
+   if (transforms & PNG_TRANSFORM_INVERT_MONO)
+      png_set_invert_mono(png_ptr);
+#endif
+
+#ifdef PNG_READ_SHIFT_SUPPORTED
+   /* If you want to shift the pixel values from the range [0,255] or
+    * [0,65535] to the original [0,7] or [0,31], or whatever range the
+    * colors were originally in:
+    */
+   if ((transforms & PNG_TRANSFORM_SHIFT)
+       && png_get_valid(png_ptr, info_ptr, PNG_INFO_sBIT))
+   {
+      png_color_8p sig_bit;
+
+      png_get_sBIT(png_ptr, info_ptr, &sig_bit);
+      png_set_shift(png_ptr, sig_bit);
+   }
+#endif
+
+#ifdef PNG_READ_BGR_SUPPORTED
+   /* Flip the RGB pixels to BGR (or RGBA to BGRA) */
+   if (transforms & PNG_TRANSFORM_BGR)
+      png_set_bgr(png_ptr);
+#endif
+
+#ifdef PNG_READ_SWAP_ALPHA_SUPPORTED
+   /* Swap the RGBA or GA data to ARGB or AG (or BGRA to ABGR) */
+   if (transforms & PNG_TRANSFORM_SWAP_ALPHA)
+      png_set_swap_alpha(png_ptr);
+#endif
+
+#ifdef PNG_READ_SWAP_SUPPORTED
+   /* Swap bytes of 16-bit files to least significant byte first */
+   if (transforms & PNG_TRANSFORM_SWAP_ENDIAN)
+      png_set_swap(png_ptr);
+#endif
+
+/* Added at libpng-1.2.41 */
+#ifdef PNG_READ_INVERT_ALPHA_SUPPORTED
+   /* Invert the alpha channel from opacity to transparency */
+   if (transforms & PNG_TRANSFORM_INVERT_ALPHA)
+      png_set_invert_alpha(png_ptr);
+#endif
+
+/* Added at libpng-1.2.41 */
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+   /* Expand grayscale image to RGB */
+   if (transforms & PNG_TRANSFORM_GRAY_TO_RGB)
+      png_set_gray_to_rgb(png_ptr);
+#endif
+
+/* Added at libpng-1.5.4 */
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+   if (transforms & PNG_TRANSFORM_EXPAND_16)
+      png_set_expand_16(png_ptr);
+#endif
+
+   /* We don't handle adding filler bytes */
+
+   /* We use png_read_image and rely on that for interlace handling, but we also
+    * call png_read_update_info therefore must turn on interlace handling now:
+    */
+   (void)png_set_interlace_handling(png_ptr);
+
+   /* Optional call to gamma correct and add the background to the palette
+    * and update info structure.  REQUIRED if you are expecting libpng to
+    * update the palette for you (i.e., you selected such a transform above).
+    */
+   png_read_update_info(png_ptr, info_ptr);
+
+   /* -------------- image transformations end here ------------------- */
+
+   png_free_data(png_ptr, info_ptr, PNG_FREE_ROWS, 0);
+   if (info_ptr->row_pointers == NULL)
+   {
+      png_uint_32 iptr;
+
+      info_ptr->row_pointers = (png_bytepp)png_malloc(png_ptr,
+          info_ptr->height * (sizeof (png_bytep)));
+      for (iptr=0; iptr<info_ptr->height; iptr++)
+         info_ptr->row_pointers[iptr] = NULL;
+
+      info_ptr->free_me |= PNG_FREE_ROWS;
+
+      for (row = 0; row < (int)info_ptr->height; row++)
+         info_ptr->row_pointers[row] = (png_bytep)png_malloc(png_ptr,
+            png_get_rowbytes(png_ptr, info_ptr));
+   }
+
+   png_read_image(png_ptr, info_ptr->row_pointers);
+   info_ptr->valid |= PNG_INFO_IDAT;
+
+   /* Read rest of file, and get additional chunks in info_ptr - REQUIRED */
+   png_read_end(png_ptr, info_ptr);
+
+   PNG_UNUSED(transforms)   /* Quiet compiler warnings */
+   PNG_UNUSED(params)
+
+}
+#endif /* PNG_INFO_IMAGE_SUPPORTED */
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+#ifdef PNG_SIMPLIFIED_READ_SUPPORTED
+/* SIMPLIFIED READ
+ *
+ * This code currently relies on the sequential reader, though it could easily
+ * be made to work with the progressive one.
+ */
+/* Arguments to png_image_finish_read: */
+
+/* Encoding of PNG data (used by the color-map code) */
+/* TODO: change these, dang, ANSI-C reserves the 'E' namespace. */
+#  define E_NOTSET  0 /* File encoding not yet known */
+#  define E_sRGB    1 /* 8-bit encoded to sRGB gamma */
+#  define E_LINEAR  2 /* 16-bit linear: not encoded, NOT pre-multiplied! */
+#  define E_FILE    3 /* 8-bit encoded to file gamma, not sRGB or linear */
+#  define E_LINEAR8 4 /* 8-bit linear: only from a file value */
+
+/* Color-map processing: after libpng has run on the PNG image further
+ * processing may be needed to conver the data to color-map indicies.
+ */
+#define PNG_CMAP_NONE      0
+#define PNG_CMAP_GA        1 /* Process GA data to a color-map with alpha */
+#define PNG_CMAP_TRANS     2 /* Process GA data to a background index */
+#define PNG_CMAP_RGB       3 /* Process RGB data */
+#define PNG_CMAP_RGB_ALPHA 4 /* Process RGBA data */
+
+/* The following document where the background is for each processing case. */
+#define PNG_CMAP_NONE_BACKGROUND      256
+#define PNG_CMAP_GA_BACKGROUND        231
+#define PNG_CMAP_TRANS_BACKGROUND     254
+#define PNG_CMAP_RGB_BACKGROUND       256
+#define PNG_CMAP_RGB_ALPHA_BACKGROUND 216
+
+typedef struct
+{
+   /* Arguments: */
+   png_imagep image;
+   png_voidp  buffer;
+   png_int_32 row_stride;
+   png_voidp  colormap;
+   png_const_colorp background;
+   /* Local variables: */
+   png_voidp       local_row;
+   png_voidp       first_row;
+   ptrdiff_t       row_bytes;           /* step between rows */
+   int             file_encoding;       /* E_ values above */
+   png_fixed_point gamma_to_linear;     /* For E_FILE, reciprocal of gamma */
+   int             colormap_processing; /* PNG_CMAP_ values above */
+} png_image_read_control;
+
+/* Do all the *safe* initialization - 'safe' means that png_error won't be
+ * called, so setting up the jmp_buf is not required.  This means that anything
+ * called from here must *not* call png_malloc - it has to call png_malloc_warn
+ * instead so that control is returned safely back to this routine.
+ */
+static int
+png_image_read_init(png_imagep image)
+{
+   if (image->opaque == NULL)
+   {
+      png_structp png_ptr = png_create_read_struct(PNG_LIBPNG_VER_STRING, image,
+          png_safe_error, png_safe_warning);
+
+      /* And set the rest of the structure to NULL to ensure that the various
+       * fields are consistent.
+       */
+      memset(image, 0, (sizeof *image));
+      image->version = PNG_IMAGE_VERSION;
+
+      if (png_ptr != NULL)
+      {
+         png_infop info_ptr = png_create_info_struct(png_ptr);
+
+         if (info_ptr != NULL)
+         {
+            png_controlp control = png_voidcast(png_controlp,
+               png_malloc_warn(png_ptr, (sizeof *control)));
+
+            if (control != NULL)
+            {
+               memset(control, 0, (sizeof *control));
+
+               control->png_ptr = png_ptr;
+               control->info_ptr = info_ptr;
+               control->for_write = 0;
+
+               image->opaque = control;
+               return 1;
+            }
+
+            /* Error clean up */
+            png_destroy_info_struct(png_ptr, &info_ptr);
+         }
+
+         png_destroy_read_struct(&png_ptr, NULL, NULL);
+      }
+
+      return png_image_error(image, "png_image_read: out of memory");
+   }
+
+   return png_image_error(image, "png_image_read: opaque pointer not NULL");
+}
+
+/* Utility to find the base format of a PNG file from a png_struct. */
+static png_uint_32
+png_image_format(png_structrp png_ptr)
+{
+   png_uint_32 format = 0;
+
+   if (png_ptr->color_type & PNG_COLOR_MASK_COLOR)
+      format |= PNG_FORMAT_FLAG_COLOR;
+
+   if (png_ptr->color_type & PNG_COLOR_MASK_ALPHA)
+      format |= PNG_FORMAT_FLAG_ALPHA;
+
+   /* Use png_ptr here, not info_ptr, because by examination png_handle_tRNS
+    * sets the png_struct fields; that's all we are interested in here.  The
+    * precise interaction with an app call to png_set_tRNS and PNG file reading
+    * is unclear.
+    */
+   else if (png_ptr->num_trans > 0)
+      format |= PNG_FORMAT_FLAG_ALPHA;
+
+   if (png_ptr->bit_depth == 16)
+      format |= PNG_FORMAT_FLAG_LINEAR;
+
+   if (png_ptr->color_type & PNG_COLOR_MASK_PALETTE)
+      format |= PNG_FORMAT_FLAG_COLORMAP;
+
+   return format;
+}
+
+/* Is the given gamma significantly different from sRGB?  The test is the same
+ * one used in pngrtran.c when deciding whether to do gamma correction.  The
+ * arithmetic optimizes the division by using the fact that the inverse of the
+ * file sRGB gamma is 2.2
+ */
+static int
+png_gamma_not_sRGB(png_fixed_point g)
+{
+   if (g < PNG_FP_1)
+   {
+      /* An uninitialized gamma is assumed to be sRGB for the simplified API. */
+      if (g == 0)
+         return 0;
+
+      return png_gamma_significant((g * 11 + 2)/5 /* i.e. *2.2, rounded */);
+   }
+
+   return 1;
+}
+
+/* Do the main body of a 'png_image_begin_read' function; read the PNG file
+ * header and fill in all the information.  This is executed in a safe context,
+ * unlike the init routine above.
+ */
+static int
+png_image_read_header(png_voidp argument)
+{
+   png_imagep image = png_voidcast(png_imagep, argument);
+   png_structrp png_ptr = image->opaque->png_ptr;
+   png_inforp info_ptr = image->opaque->info_ptr;
+
+   png_set_benign_errors(png_ptr, 1/*warn*/);
+   png_read_info(png_ptr, info_ptr);
+
+   /* Do this the fast way; just read directly out of png_struct. */
+   image->width = png_ptr->width;
+   image->height = png_ptr->height;
+
+   {
+      png_uint_32 format = png_image_format(png_ptr);
+
+      image->format = format;
+
+#ifdef PNG_COLORSPACE_SUPPORTED
+      /* Does the colorspace match sRGB?  If there is no color endpoint
+       * (colorant) information assume yes, otherwise require the
+       * 'ENDPOINTS_MATCHE_sRGB' colorspace flag to have been set.  If the
+       * colorspace has been determined to be invalid ignore it.
+       */
+      if ((format & PNG_FORMAT_FLAG_COLOR) != 0 && ((png_ptr->colorspace.flags
+         & (PNG_COLORSPACE_HAVE_ENDPOINTS|PNG_COLORSPACE_ENDPOINTS_MATCH_sRGB|
+            PNG_COLORSPACE_INVALID)) == PNG_COLORSPACE_HAVE_ENDPOINTS))
+         image->flags |= PNG_IMAGE_FLAG_COLORSPACE_NOT_sRGB;
+#endif
+   }
+
+   /* We need the maximum number of entries regardless of the format the
+    * application sets here.
+    */
+   {
+      png_uint_32 cmap_entries;
+
+      switch (png_ptr->color_type)
+      {
+         case PNG_COLOR_TYPE_GRAY:
+            cmap_entries = 1U << png_ptr->bit_depth;
+            break;
+
+         case PNG_COLOR_TYPE_PALETTE:
+            cmap_entries = png_ptr->num_palette;
+            break;
+
+         default:
+            cmap_entries = 256;
+            break;
+      }
+
+      if (cmap_entries > 256)
+         cmap_entries = 256;
+
+      image->colormap_entries = cmap_entries;
+   }
+
+   return 1;
+}
+
+#ifdef PNG_STDIO_SUPPORTED
+int PNGAPI
+png_image_begin_read_from_stdio(png_imagep image, FILE* file)
+{
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      if (file != NULL)
+      {
+         if (png_image_read_init(image))
+         {
+            /* This is slightly evil, but png_init_io doesn't do anything other
+             * than this and we haven't changed the standard IO functions so
+             * this saves a 'safe' function.
+             */
+            image->opaque->png_ptr->io_ptr = file;
+            return png_safe_execute(image, png_image_read_header, image);
+         }
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_begin_read_from_stdio: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_begin_read_from_stdio: incorrect PNG_IMAGE_VERSION");
+
+   return 0;
+}
+
+int PNGAPI
+png_image_begin_read_from_file(png_imagep image, const char *file_name)
+{
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      if (file_name != NULL)
+      {
+         FILE *fp = fopen(file_name, "rb");
+
+         if (fp != NULL)
+         {
+            if (png_image_read_init(image))
+            {
+               image->opaque->png_ptr->io_ptr = fp;
+               image->opaque->owned_file = 1;
+               return png_safe_execute(image, png_image_read_header, image);
+            }
+
+            /* Clean up: just the opened file. */
+            (void)fclose(fp);
+         }
+
+         else
+            return png_image_error(image, strerror(errno));
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_begin_read_from_file: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_begin_read_from_file: incorrect PNG_IMAGE_VERSION");
+
+   return 0;
+}
+#endif /* PNG_STDIO_SUPPORTED */
+
+static void PNGCBAPI
+png_image_memory_read(png_structp png_ptr, png_bytep out, png_size_t need)
+{
+   if (png_ptr != NULL)
+   {
+      png_imagep image = png_voidcast(png_imagep, png_ptr->io_ptr);
+      if (image != NULL)
+      {
+         png_controlp cp = image->opaque;
+         if (cp != NULL)
+         {
+            png_const_bytep memory = cp->memory;
+            png_size_t size = cp->size;
+
+            if (memory != NULL && size >= need)
+            {
+               memcpy(out, memory, need);
+               cp->memory = memory + need;
+               cp->size = size - need;
+               return;
+            }
+
+            png_error(png_ptr, "read beyond end of data");
+         }
+      }
+
+      png_error(png_ptr, "invalid memory read");
+   }
+}
+
+int PNGAPI png_image_begin_read_from_memory(png_imagep image,
+   png_const_voidp memory, png_size_t size)
+{
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      if (memory != NULL && size > 0)
+      {
+         if (png_image_read_init(image))
+         {
+            /* Now set the IO functions to read from the memory buffer and
+             * store it into io_ptr.  Again do this in-place to avoid calling a
+             * libpng function that requires error handling.
+             */
+            image->opaque->memory = png_voidcast(png_const_bytep, memory);
+            image->opaque->size = size;
+            image->opaque->png_ptr->io_ptr = image;
+            image->opaque->png_ptr->read_data_fn = png_image_memory_read;
+
+            return png_safe_execute(image, png_image_read_header, image);
+         }
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_begin_read_from_memory: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_begin_read_from_memory: incorrect PNG_IMAGE_VERSION");
+
+   return 0;
+}
+
+/* Utility function to skip chunks that are not used by the simplified image
+ * read functions and an appropriate macro to call it.
+ */
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+static void
+png_image_skip_unused_chunks(png_structrp png_ptr)
+{
+   /* Prepare the reader to ignore all recognized chunks whose data will not
+    * be used, i.e., all chunks recognized by libpng except for those
+    * involved in basic image reading:
+    *
+    *    IHDR, PLTE, IDAT, IEND
+    *
+    * Or image data handling:
+    *
+    *    tRNS, bKGD, gAMA, cHRM, sRGB, iCCP and sBIT.
+    *
+    * This provides a small performance improvement and eliminates any
+    * potential vulnerability to security problems in the unused chunks.
+    */
+   {
+         static PNG_CONST png_byte chunks_to_process[] = {
+            98,  75,  71,  68, '\0',  /* bKGD */
+            99,  72,  82,  77, '\0',  /* cHRM */
+           103,  65,  77,  65, '\0',  /* gAMA */
+           105,  67,  67,  80, '\0',  /* iCCP */
+           115,  66,  73,  84, '\0',  /* sBIT */
+           115,  82,  71,  66, '\0',  /* sRGB */
+           };
+
+       /* Ignore unknown chunks and all other chunks except for the
+        * IHDR, PLTE, tRNS, IDAT, and IEND chunks.
+        */
+       png_set_keep_unknown_chunks(png_ptr, PNG_HANDLE_CHUNK_NEVER,
+         NULL, -1);
+
+       /* But do not ignore image data handling chunks */
+       png_set_keep_unknown_chunks(png_ptr, PNG_HANDLE_CHUNK_AS_DEFAULT,
+         chunks_to_process, (sizeof chunks_to_process)/5);
+    }
+}
+
+#  define PNG_SKIP_CHUNKS(p) png_image_skip_unused_chunks(p)
+#else
+#  define PNG_SKIP_CHUNKS(p) ((void)0)
+#endif /* PNG_HANDLE_AS_UNKNOWN_SUPPORTED */
+
+/* The following macro gives the exact rounded answer for all values in the
+ * range 0..255 (it actually divides by 51.2, but the rounding still generates
+ * the correct numbers 0..5
+ */
+#define PNG_DIV51(v8) (((v8) * 5 + 130) >> 8)
+
+/* Utility functions to make particular color-maps */
+static void
+set_file_encoding(png_image_read_control *display)
+{
+   png_fixed_point g = display->image->opaque->png_ptr->colorspace.gamma;
+   if (png_gamma_significant(g))
+   {
+      if (png_gamma_not_sRGB(g))
+      {
+         display->file_encoding = E_FILE;
+         display->gamma_to_linear = png_reciprocal(g);
+      }
+
+      else
+         display->file_encoding = E_sRGB;
+   }
+
+   else
+      display->file_encoding = E_LINEAR8;
+}
+
+static unsigned int
+decode_gamma(png_image_read_control *display, png_uint_32 value, int encoding)
+{
+   if (encoding == E_FILE) /* double check */
+      encoding = display->file_encoding;
+
+   if (encoding == E_NOTSET) /* must be the file encoding */
+   {
+      set_file_encoding(display);
+      encoding = display->file_encoding;
+   }
+
+   switch (encoding)
+   {
+      case E_FILE:
+         value = png_gamma_16bit_correct(value*257, display->gamma_to_linear);
+         break;
+
+      case E_sRGB:
+         value = png_sRGB_table[value];
+         break;
+
+      case E_LINEAR:
+         break;
+
+      case E_LINEAR8:
+         value *= 257;
+         break;
+
+      default:
+         png_error(display->image->opaque->png_ptr,
+            "unexpected encoding (internal error)");
+         break;
+   }
+
+   return value;
+}
+
+static png_uint_32
+png_colormap_compose(png_image_read_control *display,
+   png_uint_32 foreground, int foreground_encoding, png_uint_32 alpha,
+   png_uint_32 background, int encoding)
+{
+   /* The file value is composed on the background, the background has the given
+    * encoding and so does the result, the file is encoded with E_FILE and the
+    * file and alpha are 8-bit values.  The (output) encoding will always be
+    * E_LINEAR or E_sRGB.
+    */
+   png_uint_32 f = decode_gamma(display, foreground, foreground_encoding);
+   png_uint_32 b = decode_gamma(display, background, encoding);
+
+   /* The alpha is always an 8-bit value (it comes from the palette), the value
+    * scaled by 255 is what PNG_sRGB_FROM_LINEAR requires.
+    */
+   f = f * alpha + b * (255-alpha);
+
+   if (encoding == E_LINEAR)
+   {
+      /* Scale to 65535; divide by 255, approximately (in fact this is extremely
+       * accurate, it divides by 255.00000005937181414556, with no overflow.)
+       */
+      f *= 257; /* Now scaled by 65535 */
+      f += f >> 16;
+      f = (f+32768) >> 16;
+   }
+
+   else /* E_sRGB */
+      f = PNG_sRGB_FROM_LINEAR(f);
+
+   return f;
+}
+
+/* NOTE: E_LINEAR values to this routine must be 16-bit, but E_FILE values must
+ * be 8-bit.
+ */
+static void
+png_create_colormap_entry(png_image_read_control *display,
+   png_uint_32 ip, png_uint_32 red, png_uint_32 green, png_uint_32 blue,
+   png_uint_32 alpha, int encoding)
+{
+   png_imagep image = display->image;
+   const int output_encoding = (image->format & PNG_FORMAT_FLAG_LINEAR) ?
+      E_LINEAR : E_sRGB;
+   const int convert_to_Y = (image->format & PNG_FORMAT_FLAG_COLOR) == 0 &&
+      (red != green || green != blue);
+
+   if (ip > 255)
+      png_error(image->opaque->png_ptr, "color-map index out of range");
+
+   /* Update the cache with whether the file gamma is significantly different
+    * from sRGB.
+    */
+   if (encoding == E_FILE)
+   {
+      if (display->file_encoding == E_NOTSET)
+         set_file_encoding(display);
+
+      /* Note that the cached value may be E_FILE too, but if it is then the
+       * gamma_to_linear member has been set.
+       */
+      encoding = display->file_encoding;
+   }
+
+   if (encoding == E_FILE)
+   {
+      png_fixed_point g = display->gamma_to_linear;
+
+      red = png_gamma_16bit_correct(red*257, g);
+      green = png_gamma_16bit_correct(green*257, g);
+      blue = png_gamma_16bit_correct(blue*257, g);
+
+      if (convert_to_Y || output_encoding == E_LINEAR)
+      {
+         alpha *= 257;
+         encoding = E_LINEAR;
+      }
+
+      else
+      {
+         red = PNG_sRGB_FROM_LINEAR(red * 255);
+         green = PNG_sRGB_FROM_LINEAR(green * 255);
+         blue = PNG_sRGB_FROM_LINEAR(blue * 255);
+         encoding = E_sRGB;
+      }
+   }
+
+   else if (encoding == E_LINEAR8)
+   {
+      /* This encoding occurs quite frequently in test cases because PngSuite
+       * includes a gAMA 1.0 chunk with most images.
+       */
+      red *= 257;
+      green *= 257;
+      blue *= 257;
+      alpha *= 257;
+      encoding = E_LINEAR;
+   }
+
+   else if (encoding == E_sRGB && (convert_to_Y || output_encoding == E_LINEAR))
+   {
+      /* The values are 8-bit sRGB values, but must be converted to 16-bit
+       * linear.
+       */
+      red = png_sRGB_table[red];
+      green = png_sRGB_table[green];
+      blue = png_sRGB_table[blue];
+      alpha *= 257;
+      encoding = E_LINEAR;
+   }
+
+   /* This is set if the color isn't gray but the output is. */
+   if (encoding == E_LINEAR)
+   {
+      if (convert_to_Y)
+      {
+         /* NOTE: these values are copied from png_do_rgb_to_gray */
+         png_uint_32 y = (png_uint_32)6968 * red  + (png_uint_32)23434 * green +
+            (png_uint_32)2366 * blue;
+
+         if (output_encoding == E_LINEAR)
+            y = (y + 16384) >> 15;
+
+         else
+         {
+            /* y is scaled by 32768, we need it scaled by 255: */
+            y = (y + 128) >> 8;
+            y *= 255;
+            y = PNG_sRGB_FROM_LINEAR((y + 64) >> 7);
+            encoding = E_sRGB;
+         }
+
+         blue = red = green = y;
+      }
+
+      else if (output_encoding == E_sRGB)
+      {
+         red = PNG_sRGB_FROM_LINEAR(red * 255);
+         green = PNG_sRGB_FROM_LINEAR(green * 255);
+         blue = PNG_sRGB_FROM_LINEAR(blue * 255);
+         alpha = PNG_DIV257(alpha);
+         encoding = E_sRGB;
+      }
+   }
+
+   if (encoding != output_encoding)
+      png_error(image->opaque->png_ptr, "bad encoding (internal error)");
+
+   /* Store the value. */
+   {
+#     ifdef PNG_FORMAT_BGR_SUPPORTED
+         const int afirst = (image->format & PNG_FORMAT_FLAG_AFIRST) != 0 &&
+            (image->format & PNG_FORMAT_FLAG_ALPHA) != 0;
+#     else
+#        define afirst 0
+#     endif
+#     ifdef PNG_FORMAT_BGR_SUPPORTED
+         const int bgr = (image->format & PNG_FORMAT_FLAG_BGR) ? 2 : 0;
+#     else
+#        define bgr 0
+#     endif
+
+      if (output_encoding == E_LINEAR)
+      {
+         png_uint_16p entry = png_voidcast(png_uint_16p, display->colormap);
+
+         entry += ip * PNG_IMAGE_SAMPLE_CHANNELS(image->format);
+
+         /* The linear 16-bit values must be pre-multiplied by the alpha channel
+          * value, if less than 65535 (this is, effectively, composite on black
+          * if the alpha channel is removed.)
+          */
+         switch (PNG_IMAGE_SAMPLE_CHANNELS(image->format))
+         {
+            case 4:
+               entry[afirst ? 0 : 3] = (png_uint_16)alpha;
+               /* FALL THROUGH */
+
+            case 3:
+               if (alpha < 65535)
+               {
+                  if (alpha > 0)
+                  {
+                     blue = (blue * alpha + 32767U)/65535U;
+                     green = (green * alpha + 32767U)/65535U;
+                     red = (red * alpha + 32767U)/65535U;
+                  }
+
+                  else
+                     red = green = blue = 0;
+               }
+               entry[afirst + (2 ^ bgr)] = (png_uint_16)blue;
+               entry[afirst + 1] = (png_uint_16)green;
+               entry[afirst + bgr] = (png_uint_16)red;
+               break;
+
+            case 2:
+               entry[1 ^ afirst] = (png_uint_16)alpha;
+               /* FALL THROUGH */
+
+            case 1:
+               if (alpha < 65535)
+               {
+                  if (alpha > 0)
+                     green = (green * alpha + 32767U)/65535U;
+
+                  else
+                     green = 0;
+               }
+               entry[afirst] = (png_uint_16)green;
+               break;
+
+            default:
+               break;
+         }
+      }
+
+      else /* output encoding is E_sRGB */
+      {
+         png_bytep entry = png_voidcast(png_bytep, display->colormap);
+
+         entry += ip * PNG_IMAGE_SAMPLE_CHANNELS(image->format);
+
+         switch (PNG_IMAGE_SAMPLE_CHANNELS(image->format))
+         {
+            case 4:
+               entry[afirst ? 0 : 3] = (png_byte)alpha;
+            case 3:
+               entry[afirst + (2 ^ bgr)] = (png_byte)blue;
+               entry[afirst + 1] = (png_byte)green;
+               entry[afirst + bgr] = (png_byte)red;
+               break;
+
+            case 2:
+               entry[1 ^ afirst] = (png_byte)alpha;
+            case 1:
+               entry[afirst] = (png_byte)green;
+               break;
+
+            default:
+               break;
+         }
+      }
+
+#     ifdef afirst
+#        undef afirst
+#     endif
+#     ifdef bgr
+#        undef bgr
+#     endif
+   }
+}
+
+static int
+make_gray_file_colormap(png_image_read_control *display)
+{
+   unsigned int i;
+
+   for (i=0; i<256; ++i)
+      png_create_colormap_entry(display, i, i, i, i, 255, E_FILE);
+
+   return i;
+}
+
+static int
+make_gray_colormap(png_image_read_control *display)
+{
+   unsigned int i;
+
+   for (i=0; i<256; ++i)
+      png_create_colormap_entry(display, i, i, i, i, 255, E_sRGB);
+
+   return i;
+}
+#define PNG_GRAY_COLORMAP_ENTRIES 256
+
+static int
+make_ga_colormap(png_image_read_control *display)
+{
+   unsigned int i, a;
+
+   /* Alpha is retained, the output will be a color-map with entries
+    * selected by six levels of alpha.  One transparent entry, 6 gray
+    * levels for all the intermediate alpha values, leaving 230 entries
+    * for the opaque grays.  The color-map entries are the six values
+    * [0..5]*51, the GA processing uses PNG_DIV51(value) to find the
+    * relevant entry.
+    *
+    * if (alpha > 229) // opaque
+    * {
+    *    // The 231 entries are selected to make the math below work:
+    *    base = 0;
+    *    entry = (231 * gray + 128) >> 8;
+    * }
+    * else if (alpha < 26) // transparent
+    * {
+    *    base = 231;
+    *    entry = 0;
+    * }
+    * else // partially opaque
+    * {
+    *    base = 226 + 6 * PNG_DIV51(alpha);
+    *    entry = PNG_DIV51(gray);
+    * }
+    */
+   i = 0;
+   while (i < 231)
+   {
+      unsigned int gray = (i * 256 + 115) / 231;
+      png_create_colormap_entry(display, i++, gray, gray, gray, 255, E_sRGB);
+   }
+
+   /* 255 is used here for the component values for consistency with the code
+    * that undoes premultiplication in pngwrite.c.
+    */
+   png_create_colormap_entry(display, i++, 255, 255, 255, 0, E_sRGB);
+
+   for (a=1; a<5; ++a)
+   {
+      unsigned int g;
+
+      for (g=0; g<6; ++g)
+         png_create_colormap_entry(display, i++, g*51, g*51, g*51, a*51,
+            E_sRGB);
+   }
+
+   return i;
+}
+
+#define PNG_GA_COLORMAP_ENTRIES 256
+
+static int
+make_rgb_colormap(png_image_read_control *display)
+{
+   unsigned int i, r;
+
+   /* Build a 6x6x6 opaque RGB cube */
+   for (i=r=0; r<6; ++r)
+   {
+      unsigned int g;
+
+      for (g=0; g<6; ++g)
+      {
+         unsigned int b;
+
+         for (b=0; b<6; ++b)
+            png_create_colormap_entry(display, i++, r*51, g*51, b*51, 255,
+               E_sRGB);
+      }
+   }
+
+   return i;
+}
+
+#define PNG_RGB_COLORMAP_ENTRIES 216
+
+/* Return a palette index to the above palette given three 8-bit sRGB values. */
+#define PNG_RGB_INDEX(r,g,b) \
+   ((png_byte)(6 * (6 * PNG_DIV51(r) + PNG_DIV51(g)) + PNG_DIV51(b)))
+
+static int
+png_image_read_colormap(png_voidp argument)
+{
+   png_image_read_control *display =
+      png_voidcast(png_image_read_control*, argument);
+   const png_imagep image = display->image;
+
+   const png_structrp png_ptr = image->opaque->png_ptr;
+   const png_uint_32 output_format = image->format;
+   const int output_encoding = (output_format & PNG_FORMAT_FLAG_LINEAR) ?
+      E_LINEAR : E_sRGB;
+
+   unsigned int cmap_entries;
+   unsigned int output_processing;        /* Output processing option */
+   unsigned int data_encoding = E_NOTSET; /* Encoding libpng must produce */
+
+   /* Background information; the background color and the index of this color
+    * in the color-map if it exists (else 256).
+    */
+   unsigned int background_index = 256;
+   png_uint_32 back_r, back_g, back_b;
+
+   /* Flags to accumulate things that need to be done to the input. */
+   int expand_tRNS = 0;
+
+   /* Exclude the NYI feature of compositing onto a color-mapped buffer; it is
+    * very difficult to do, the results look awful, and it is difficult to see
+    * what possible use it is because the application can't control the
+    * color-map.
+    */
+   if (((png_ptr->color_type & PNG_COLOR_MASK_ALPHA) != 0 ||
+         png_ptr->num_trans > 0) /* alpha in input */ &&
+      ((output_format & PNG_FORMAT_FLAG_ALPHA) == 0) /* no alpha in output */)
+   {
+      if (output_encoding == E_LINEAR) /* compose on black */
+         back_b = back_g = back_r = 0;
+
+      else if (display->background == NULL /* no way to remove it */)
+         png_error(png_ptr,
+            "a background color must be supplied to remove alpha/transparency");
+
+      /* Get a copy of the background color (this avoids repeating the checks
+       * below.)  The encoding is 8-bit sRGB or 16-bit linear, depending on the
+       * output format.
+       */
+      else
+      {
+         back_g = display->background->green;
+         if (output_format & PNG_FORMAT_FLAG_COLOR)
+         {
+            back_r = display->background->red;
+            back_b = display->background->blue;
+         }
+         else
+            back_b = back_r = back_g;
+      }
+   }
+
+   else if (output_encoding == E_LINEAR)
+      back_b = back_r = back_g = 65535;
+
+   else
+      back_b = back_r = back_g = 255;
+
+   /* Default the input file gamma if required - this is necessary because
+    * libpng assumes that if no gamma information is present the data is in the
+    * output format, but the simplified API deduces the gamma from the input
+    * format.
+    */
+   if ((png_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_GAMMA) == 0)
+   {
+      /* Do this directly, not using the png_colorspace functions, to ensure
+       * that it happens even if the colorspace is invalid (though probably if
+       * it is the setting will be ignored)  Note that the same thing can be
+       * achieved at the application interface with png_set_gAMA.
+       */
+      if (png_ptr->bit_depth == 16 &&
+         (image->flags & PNG_IMAGE_FLAG_16BIT_sRGB) == 0)
+         png_ptr->colorspace.gamma = PNG_GAMMA_LINEAR;
+
+      else
+         png_ptr->colorspace.gamma = PNG_GAMMA_sRGB_INVERSE;
+
+      png_ptr->colorspace.flags |= PNG_COLORSPACE_HAVE_GAMMA;
+   }
+
+   /* Decide what to do based on the PNG color type of the input data.  The
+    * utility function png_create_colormap_entry deals with most aspects of the
+    * output transformations; this code works out how to produce bytes of
+    * color-map entries from the original format.
+    */
+   switch (png_ptr->color_type)
+   {
+      case PNG_COLOR_TYPE_GRAY:
+         if (png_ptr->bit_depth <= 8)
+         {
+            /* There at most 256 colors in the output, regardless of
+             * transparency.
+             */
+            unsigned int step, i, val, trans = 256/*ignore*/, back_alpha = 0;
+
+            cmap_entries = 1U << png_ptr->bit_depth;
+            if (cmap_entries > image->colormap_entries)
+               png_error(png_ptr, "gray[8] color-map: too few entries");
+
+            step = 255 / (cmap_entries - 1);
+            output_processing = PNG_CMAP_NONE;
+
+            /* If there is a tRNS chunk then this either selects a transparent
+             * value or, if the output has no alpha, the background color.
+             */
+            if (png_ptr->num_trans > 0)
+            {
+               trans = png_ptr->trans_color.gray;
+
+               if ((output_format & PNG_FORMAT_FLAG_ALPHA) == 0)
+                  back_alpha = output_encoding == E_LINEAR ? 65535 : 255;
+            }
+
+            /* png_create_colormap_entry just takes an RGBA and writes the
+             * corresponding color-map entry using the format from 'image',
+             * including the required conversion to sRGB or linear as
+             * appropriate.  The input values are always either sRGB (if the
+             * gamma correction flag is 0) or 0..255 scaled file encoded values
+             * (if the function must gamma correct them).
+             */
+            for (i=val=0; i<cmap_entries; ++i, val += step)
+            {
+               /* 'i' is a file value.  While this will result in duplicated
+                * entries for 8-bit non-sRGB encoded files it is necessary to
+                * have non-gamma corrected values to do tRNS handling.
+                */
+               if (i != trans)
+                  png_create_colormap_entry(display, i, val, val, val, 255,
+                     E_FILE/*8-bit with file gamma*/);
+
+               /* Else this entry is transparent.  The colors don't matter if
+                * there is an alpha channel (back_alpha == 0), but it does no
+                * harm to pass them in; the values are not set above so this
+                * passes in white.
+                *
+                * NOTE: this preserves the full precision of the application
+                * supplied background color when it is used.
+                */
+               else
+                  png_create_colormap_entry(display, i, back_r, back_g, back_b,
+                     back_alpha, output_encoding);
+            }
+
+            /* We need libpng to preserve the original encoding. */
+            data_encoding = E_FILE;
+
+            /* The rows from libpng, while technically gray values, are now also
+             * color-map indicies; however, they may need to be expanded to 1
+             * byte per pixel.  This is what png_set_packing does (i.e., it
+             * unpacks the bit values into bytes.)
+             */
+            if (png_ptr->bit_depth < 8)
+               png_set_packing(png_ptr);
+         }
+
+         else /* bit depth is 16 */
+         {
+            /* The 16-bit input values can be converted directly to 8-bit gamma
+             * encoded values; however, if a tRNS chunk is present 257 color-map
+             * entries are required.  This means that the extra entry requires
+             * special processing; add an alpha channel, sacrifice gray level
+             * 254 and convert transparent (alpha==0) entries to that.
+             *
+             * Use libpng to chop the data to 8 bits.  Convert it to sRGB at the
+             * same time to minimize quality loss.  If a tRNS chunk is present
+             * this means libpng must handle it too; otherwise it is impossible
+             * to do the exact match on the 16-bit value.
+             *
+             * If the output has no alpha channel *and* the background color is
+             * gray then it is possible to let libpng handle the substitution by
+             * ensuring that the corresponding gray level matches the background
+             * color exactly.
+             */
+            data_encoding = E_sRGB;
+
+            if (PNG_GRAY_COLORMAP_ENTRIES > image->colormap_entries)
+               png_error(png_ptr, "gray[16] color-map: too few entries");
+
+            cmap_entries = make_gray_colormap(display);
+
+            if (png_ptr->num_trans > 0)
+            {
+               unsigned int back_alpha;
+
+               if (output_format & PNG_FORMAT_FLAG_ALPHA)
+                  back_alpha = 0;
+
+               else
+               {
+                  if (back_r == back_g && back_g == back_b)
+                  {
+                     /* Background is gray; no special processing will be
+                      * required.
+                      */
+                     png_color_16 c;
+                     png_uint_32 gray = back_g;
+
+                     if (output_encoding == E_LINEAR)
+                     {
+                        gray = PNG_sRGB_FROM_LINEAR(gray * 255);
+
+                        /* And make sure the corresponding palette entry
+                         * matches.
+                         */
+                        png_create_colormap_entry(display, gray, back_g, back_g,
+                           back_g, 65535, E_LINEAR);
+                     }
+
+                     /* The background passed to libpng, however, must be the
+                      * sRGB value.
+                      */
+                     c.index = 0; /*unused*/
+                     c.gray = c.red = c.green = c.blue = (png_uint_16)gray;
+
+                     /* NOTE: does this work without expanding tRNS to alpha?
+                      * It should be the color->gray case below apparently
+                      * doesn't.
+                      */
+                     png_set_background_fixed(png_ptr, &c,
+                        PNG_BACKGROUND_GAMMA_SCREEN, 0/*need_expand*/,
+                        0/*gamma: not used*/);
+
+                     output_processing = PNG_CMAP_NONE;
+                     break;
+                  }
+
+                  back_alpha = output_encoding == E_LINEAR ? 65535 : 255;
+               }
+
+               /* output_processing means that the libpng-processed row will be
+                * 8-bit GA and it has to be processing to single byte color-map
+                * values.  Entry 254 is replaced by either a completely
+                * transparent entry or by the background color at full
+                * precision (and the background color is not a simple gray leve
+                * in this case.)
+                */
+               expand_tRNS = 1;
+               output_processing = PNG_CMAP_TRANS;
+               background_index = 254;
+
+               /* And set (overwrite) color-map entry 254 to the actual
+                * background color at full precision.
+                */
+               png_create_colormap_entry(display, 254, back_r, back_g, back_b,
+                  back_alpha, output_encoding);
+            }
+
+            else
+               output_processing = PNG_CMAP_NONE;
+         }
+         break;
+
+      case PNG_COLOR_TYPE_GRAY_ALPHA:
+         /* 8-bit or 16-bit PNG with two channels - gray and alpha.  A minimum
+          * of 65536 combinations.  If, however, the alpha channel is to be
+          * removed there are only 256 possibilities if the background is gray.
+          * (Otherwise there is a subset of the 65536 possibilities defined by
+          * the triangle between black, white and the background color.)
+          *
+          * Reduce 16-bit files to 8-bit and sRGB encode the result.  No need to
+          * worry about tRNS matching - tRNS is ignored if there is an alpha
+          * channel.
+          */
+         data_encoding = E_sRGB;
+
+         if (output_format & PNG_FORMAT_FLAG_ALPHA)
+         {
+            if (PNG_GA_COLORMAP_ENTRIES > image->colormap_entries)
+               png_error(png_ptr, "gray+alpha color-map: too few entries");
+
+            cmap_entries = make_ga_colormap(display);
+
+            background_index = PNG_CMAP_GA_BACKGROUND;
+            output_processing = PNG_CMAP_GA;
+         }
+
+         else /* alpha is removed */
+         {
+            /* Alpha must be removed as the PNG data is processed when the
+             * background is a color because the G and A channels are
+             * independent and the vector addition (non-parallel vectors) is a
+             * 2-D problem.
+             *
+             * This can be reduced to the same algorithm as above by making a
+             * colormap containing gray levels (for the opaque grays), a
+             * background entry (for a transparent pixel) and a set of four six
+             * level color values, one set for each intermediate alpha value.
+             * See the comments in make_ga_colormap for how this works in the
+             * per-pixel processing.
+             *
+             * If the background is gray, however, we only need a 256 entry gray
+             * level color map.  It is sufficient to make the entry generated
+             * for the background color be exactly the color specified.
+             */
+            if ((output_format & PNG_FORMAT_FLAG_COLOR) == 0 ||
+               (back_r == back_g && back_g == back_b))
+            {
+               /* Background is gray; no special processing will be required. */
+               png_color_16 c;
+               png_uint_32 gray = back_g;
+
+               if (PNG_GRAY_COLORMAP_ENTRIES > image->colormap_entries)
+                  png_error(png_ptr, "gray-alpha color-map: too few entries");
+
+               cmap_entries = make_gray_colormap(display);
+
+               if (output_encoding == E_LINEAR)
+               {
+                  gray = PNG_sRGB_FROM_LINEAR(gray * 255);
+
+                  /* And make sure the corresponding palette entry matches. */
+                  png_create_colormap_entry(display, gray, back_g, back_g,
+                     back_g, 65535, E_LINEAR);
+               }
+
+               /* The background passed to libpng, however, must be the sRGB
+                * value.
+                */
+               c.index = 0; /*unused*/
+               c.gray = c.red = c.green = c.blue = (png_uint_16)gray;
+
+               png_set_background_fixed(png_ptr, &c,
+                  PNG_BACKGROUND_GAMMA_SCREEN, 0/*need_expand*/,
+                  0/*gamma: not used*/);
+
+               output_processing = PNG_CMAP_NONE;
+            }
+
+            else
+            {
+               png_uint_32 i, a;
+
+               /* This is the same as png_make_ga_colormap, above, except that
+                * the entries are all opaque.
+                */
+               if (PNG_GA_COLORMAP_ENTRIES > image->colormap_entries)
+                  png_error(png_ptr, "ga-alpha color-map: too few entries");
+
+               i = 0;
+               while (i < 231)
+               {
+                  png_uint_32 gray = (i * 256 + 115) / 231;
+                  png_create_colormap_entry(display, i++, gray, gray, gray,
+                     255, E_sRGB);
+               }
+
+               /* NOTE: this preserves the full precision of the application
+                * background color.
+                */
+               background_index = i;
+               png_create_colormap_entry(display, i++, back_r, back_g, back_b,
+                  output_encoding == E_LINEAR ? 65535U : 255U, output_encoding);
+
+               /* For non-opaque input composite on the sRGB background - this
+                * requires inverting the encoding for each component.  The input
+                * is still converted to the sRGB encoding because this is a
+                * reasonable approximate to the logarithmic curve of human
+                * visual sensitivity, at least over the narrow range which PNG
+                * represents.  Consequently 'G' is always sRGB encoded, while
+                * 'A' is linear.  We need the linear background colors.
+                */
+               if (output_encoding == E_sRGB) /* else already linear */
+               {
+                  /* This may produce a value not exactly matching the
+                   * background, but that's ok because these numbers are only
+                   * used when alpha != 0
+                   */
+                  back_r = png_sRGB_table[back_r];
+                  back_g = png_sRGB_table[back_g];
+                  back_b = png_sRGB_table[back_b];
+               }
+
+               for (a=1; a<5; ++a)
+               {
+                  unsigned int g;
+
+                  /* PNG_sRGB_FROM_LINEAR expects a 16-bit linear value scaled
+                   * by an 8-bit alpha value (0..255).
+                   */
+                  png_uint_32 alpha = 51 * a;
+                  png_uint_32 back_rx = (255-alpha) * back_r;
+                  png_uint_32 back_gx = (255-alpha) * back_g;
+                  png_uint_32 back_bx = (255-alpha) * back_b;
+
+                  for (g=0; g<6; ++g)
+                  {
+                     png_uint_32 gray = png_sRGB_table[g*51] * alpha;
+
+                     png_create_colormap_entry(display, i++,
+                        PNG_sRGB_FROM_LINEAR(gray + back_rx),
+                        PNG_sRGB_FROM_LINEAR(gray + back_gx),
+                        PNG_sRGB_FROM_LINEAR(gray + back_bx), 255, E_sRGB);
+                  }
+               }
+
+               cmap_entries = i;
+               output_processing = PNG_CMAP_GA;
+            }
+         }
+         break;
+
+      case PNG_COLOR_TYPE_RGB:
+      case PNG_COLOR_TYPE_RGB_ALPHA:
+         /* Exclude the case where the output is gray; we can always handle this
+          * with the cases above.
+          */
+         if ((output_format & PNG_FORMAT_FLAG_COLOR) == 0)
+         {
+            /* The color-map will be grayscale, so we may as well convert the
+             * input RGB values to a simple grayscale and use the grayscale
+             * code above.
+             *
+             * NOTE: calling this apparently damages the recognition of the
+             * transparent color in background color handling; call
+             * png_set_tRNS_to_alpha before png_set_background_fixed.
+             */
+            png_set_rgb_to_gray_fixed(png_ptr, PNG_ERROR_ACTION_NONE, -1,
+               -1);
+            data_encoding = E_sRGB;
+
+            /* The output will now be one or two 8-bit gray or gray+alpha
+             * channels.  The more complex case arises when the input has alpha.
+             */
+            if ((png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+               png_ptr->num_trans > 0) &&
+               (output_format & PNG_FORMAT_FLAG_ALPHA) != 0)
+            {
+               /* Both input and output have an alpha channel, so no background
+                * processing is required; just map the GA bytes to the right
+                * color-map entry.
+                */
+               expand_tRNS = 1;
+
+               if (PNG_GA_COLORMAP_ENTRIES > image->colormap_entries)
+                  png_error(png_ptr, "rgb[ga] color-map: too few entries");
+
+               cmap_entries = make_ga_colormap(display);
+               background_index = PNG_CMAP_GA_BACKGROUND;
+               output_processing = PNG_CMAP_GA;
+            }
+
+            else
+            {
+               /* Either the input or the output has no alpha channel, so there
+                * will be no non-opaque pixels in the color-map; it will just be
+                * grayscale.
+                */
+               if (PNG_GRAY_COLORMAP_ENTRIES > image->colormap_entries)
+                  png_error(png_ptr, "rgb[gray] color-map: too few entries");
+
+               /* Ideally this code would use libpng to do the gamma correction,
+                * but if an input alpha channel is to be removed we will hit the
+                * libpng bug in gamma+compose+rgb-to-gray (the double gamma
+                * correction bug).  Fix this by dropping the gamma correction in
+                * this case and doing it in the palette; this will result in
+                * duplicate palette entries, but that's better than the
+                * alternative of double gamma correction.
+                */
+               if ((png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+                  png_ptr->num_trans > 0) &&
+                  png_gamma_not_sRGB(png_ptr->colorspace.gamma))
+               {
+                  cmap_entries = make_gray_file_colormap(display);
+                  data_encoding = E_FILE;
+               }
+
+               else
+                  cmap_entries = make_gray_colormap(display);
+
+               /* But if the input has alpha or transparency it must be removed
+                */
+               if (png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+                  png_ptr->num_trans > 0)
+               {
+                  png_color_16 c;
+                  png_uint_32 gray = back_g;
+
+                  /* We need to ensure that the application background exists in
+                   * the colormap and that completely transparent pixels map to
+                   * it.  Achieve this simply by ensuring that the entry
+                   * selected for the background really is the background color.
+                   */
+                  if (data_encoding == E_FILE) /* from the fixup above */
+                  {
+                     /* The app supplied a gray which is in output_encoding, we
+                      * need to convert it to a value of the input (E_FILE)
+                      * encoding then set this palette entry to the required
+                      * output encoding.
+                      */
+                     if (output_encoding == E_sRGB)
+                        gray = png_sRGB_table[gray]; /* now E_LINEAR */
+
+                     gray = PNG_DIV257(png_gamma_16bit_correct(gray,
+                        png_ptr->colorspace.gamma)); /* now E_FILE */
+
+                     /* And make sure the corresponding palette entry contains
+                      * exactly the required sRGB value.
+                      */
+                     png_create_colormap_entry(display, gray, back_g, back_g,
+                        back_g, 0/*unused*/, output_encoding);
+                  }
+
+                  else if (output_encoding == E_LINEAR)
+                  {
+                     gray = PNG_sRGB_FROM_LINEAR(gray * 255);
+
+                     /* And make sure the corresponding palette entry matches.
+                      */
+                     png_create_colormap_entry(display, gray, back_g, back_g,
+                        back_g, 0/*unused*/, E_LINEAR);
+                  }
+
+                  /* The background passed to libpng, however, must be the
+                   * output (normally sRGB) value.
+                   */
+                  c.index = 0; /*unused*/
+                  c.gray = c.red = c.green = c.blue = (png_uint_16)gray;
+
+                  /* NOTE: the following is apparently a bug in libpng. Without
+                   * it the transparent color recognition in
+                   * png_set_background_fixed seems to go wrong.
+                   */
+                  expand_tRNS = 1;
+                  png_set_background_fixed(png_ptr, &c,
+                     PNG_BACKGROUND_GAMMA_SCREEN, 0/*need_expand*/,
+                     0/*gamma: not used*/);
+               }
+
+               output_processing = PNG_CMAP_NONE;
+            }
+         }
+
+         else /* output is color */
+         {
+            /* We could use png_quantize here so long as there is no transparent
+             * color or alpha; png_quantize ignores alpha.  Easier overall just
+             * to do it once and using PNG_DIV51 on the 6x6x6 reduced RGB cube.
+             * Consequently we always want libpng to produce sRGB data.
+             */
+            data_encoding = E_sRGB;
+
+            /* Is there any transparency or alpha? */
+            if (png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+               png_ptr->num_trans > 0)
+            {
+               /* Is there alpha in the output too?  If so all four channels are
+                * processed into a special RGB cube with alpha support.
+                */
+               if (output_format & PNG_FORMAT_FLAG_ALPHA)
+               {
+                  png_uint_32 r;
+
+                  if (PNG_RGB_COLORMAP_ENTRIES+1+27 > image->colormap_entries)
+                     png_error(png_ptr, "rgb+alpha color-map: too few entries");
+
+                  cmap_entries = make_rgb_colormap(display);
+
+                  /* Add a transparent entry. */
+                  png_create_colormap_entry(display, cmap_entries, 255, 255,
+                     255, 0, E_sRGB);
+
+                  /* This is stored as the background index for the processing
+                   * algorithm.
+                   */
+                  background_index = cmap_entries++;
+
+                  /* Add 27 r,g,b entries each with alpha 0.5. */
+                  for (r=0; r<256; r = (r << 1) | 0x7f)
+                  {
+                     png_uint_32 g;
+
+                     for (g=0; g<256; g = (g << 1) | 0x7f)
+                     {
+                        png_uint_32 b;
+
+                        /* This generates components with the values 0, 127 and
+                         * 255
+                         */
+                        for (b=0; b<256; b = (b << 1) | 0x7f)
+                           png_create_colormap_entry(display, cmap_entries++,
+                              r, g, b, 128, E_sRGB);
+                     }
+                  }
+
+                  expand_tRNS = 1;
+                  output_processing = PNG_CMAP_RGB_ALPHA;
+               }
+
+               else
+               {
+                  /* Alpha/transparency must be removed.  The background must
+                   * exist in the color map (achieved by setting adding it after
+                   * the 666 color-map).  If the standard processing code will
+                   * pick up this entry automatically that's all that is
+                   * required; libpng can be called to do the background
+                   * processing.
+                   */
+                  unsigned int sample_size =
+                     PNG_IMAGE_SAMPLE_SIZE(output_format);
+                  png_uint_32 r, g, b; /* sRGB background */
+
+                  if (PNG_RGB_COLORMAP_ENTRIES+1+27 > image->colormap_entries)
+                     png_error(png_ptr, "rgb-alpha color-map: too few entries");
+
+                  cmap_entries = make_rgb_colormap(display);
+
+                  png_create_colormap_entry(display, cmap_entries, back_r,
+                        back_g, back_b, 0/*unused*/, output_encoding);
+
+                  if (output_encoding == E_LINEAR)
+                  {
+                     r = PNG_sRGB_FROM_LINEAR(back_r * 255);
+                     g = PNG_sRGB_FROM_LINEAR(back_g * 255);
+                     b = PNG_sRGB_FROM_LINEAR(back_b * 255);
+                  }
+
+                  else
+                  {
+                     r = back_r;
+                     g = back_g;
+                     b = back_g;
+                  }
+
+                  /* Compare the newly-created color-map entry with the one the
+                   * PNG_CMAP_RGB algorithm will use.  If the two entries don't
+                   * match, add the new one and set this as the background
+                   * index.
+                   */
+                  if (memcmp((png_const_bytep)display->colormap +
+                        sample_size * cmap_entries,
+                     (png_const_bytep)display->colormap +
+                        sample_size * PNG_RGB_INDEX(r,g,b),
+                     sample_size) != 0)
+                  {
+                     /* The background color must be added. */
+                     background_index = cmap_entries++;
+
+                     /* Add 27 r,g,b entries each with created by composing with
+                      * the background at alpha 0.5.
+                      */
+                     for (r=0; r<256; r = (r << 1) | 0x7f)
+                     {
+                        for (g=0; g<256; g = (g << 1) | 0x7f)
+                        {
+                           /* This generates components with the values 0, 127
+                            * and 255
+                            */
+                           for (b=0; b<256; b = (b << 1) | 0x7f)
+                              png_create_colormap_entry(display, cmap_entries++,
+                                 png_colormap_compose(display, r, E_sRGB, 128,
+                                    back_r, output_encoding),
+                                 png_colormap_compose(display, g, E_sRGB, 128,
+                                    back_g, output_encoding),
+                                 png_colormap_compose(display, b, E_sRGB, 128,
+                                    back_b, output_encoding),
+                                 0/*unused*/, output_encoding);
+                        }
+                     }
+
+                     expand_tRNS = 1;
+                     output_processing = PNG_CMAP_RGB_ALPHA;
+                  }
+
+                  else /* background color is in the standard color-map */
+                  {
+                     png_color_16 c;
+
+                     c.index = 0; /*unused*/
+                     c.red = (png_uint_16)back_r;
+                     c.gray = c.green = (png_uint_16)back_g;
+                     c.blue = (png_uint_16)back_b;
+
+                     png_set_background_fixed(png_ptr, &c,
+                        PNG_BACKGROUND_GAMMA_SCREEN, 0/*need_expand*/,
+                        0/*gamma: not used*/);
+
+                     output_processing = PNG_CMAP_RGB;
+                  }
+               }
+            }
+
+            else /* no alpha or transparency in the input */
+            {
+               /* Alpha in the output is irrelevant, simply map the opaque input
+                * pixels to the 6x6x6 color-map.
+                */
+               if (PNG_RGB_COLORMAP_ENTRIES > image->colormap_entries)
+                  png_error(png_ptr, "rgb color-map: too few entries");
+
+               cmap_entries = make_rgb_colormap(display);
+               output_processing = PNG_CMAP_RGB;
+            }
+         }
+         break;
+
+      case PNG_COLOR_TYPE_PALETTE:
+         /* It's already got a color-map.  It may be necessary to eliminate the
+          * tRNS entries though.
+          */
+         {
+            unsigned int num_trans = png_ptr->num_trans;
+            png_const_bytep trans = num_trans > 0 ? png_ptr->trans_alpha : NULL;
+            png_const_colorp colormap = png_ptr->palette;
+            const int do_background = trans != NULL &&
+               (output_format & PNG_FORMAT_FLAG_ALPHA) == 0;
+            unsigned int i;
+
+            /* Just in case: */
+            if (trans == NULL)
+               num_trans = 0;
+
+            output_processing = PNG_CMAP_NONE;
+            data_encoding = E_FILE; /* Don't change from color-map indicies */
+            cmap_entries = png_ptr->num_palette;
+            if (cmap_entries > 256)
+               cmap_entries = 256;
+
+            if (cmap_entries > image->colormap_entries)
+               png_error(png_ptr, "palette color-map: too few entries");
+
+            for (i=0; i < cmap_entries; ++i)
+            {
+               if (do_background && i < num_trans && trans[i] < 255)
+               {
+                  if (trans[i] == 0)
+                     png_create_colormap_entry(display, i, back_r, back_g,
+                        back_b, 0, output_encoding);
+
+                  else
+                  {
+                     /* Must compose the PNG file color in the color-map entry
+                      * on the sRGB color in 'back'.
+                      */
+                     png_create_colormap_entry(display, i,
+                        png_colormap_compose(display, colormap[i].red, E_FILE,
+                           trans[i], back_r, output_encoding),
+                        png_colormap_compose(display, colormap[i].green, E_FILE,
+                           trans[i], back_g, output_encoding),
+                        png_colormap_compose(display, colormap[i].blue, E_FILE,
+                           trans[i], back_b, output_encoding),
+                        output_encoding == E_LINEAR ? trans[i] * 257U :
+                           trans[i],
+                        output_encoding);
+                  }
+               }
+
+               else
+                  png_create_colormap_entry(display, i, colormap[i].red,
+                     colormap[i].green, colormap[i].blue,
+                     i < num_trans ? trans[i] : 255U, E_FILE/*8-bit*/);
+            }
+
+            /* The PNG data may have indicies packed in fewer than 8 bits, it
+             * must be expanded if so.
+             */
+            if (png_ptr->bit_depth < 8)
+               png_set_packing(png_ptr);
+         }
+         break;
+
+      default:
+         png_error(png_ptr, "invalid PNG color type");
+         /*NOT REACHED*/
+         break;
+   }
+
+   /* Now deal with the output processing */
+   if (expand_tRNS && png_ptr->num_trans > 0 &&
+      (png_ptr->color_type & PNG_COLOR_MASK_ALPHA) == 0)
+      png_set_tRNS_to_alpha(png_ptr);
+
+   switch (data_encoding)
+   {
+      default:
+         png_error(png_ptr, "bad data option (internal error)");
+         break;
+
+      case E_sRGB:
+         /* Change to 8-bit sRGB */
+         png_set_alpha_mode_fixed(png_ptr, PNG_ALPHA_PNG, PNG_GAMMA_sRGB);
+         /* FALL THROUGH */
+
+      case E_FILE:
+         if (png_ptr->bit_depth > 8)
+            png_set_scale_16(png_ptr);
+         break;
+   }
+
+   if (cmap_entries > 256 || cmap_entries > image->colormap_entries)
+      png_error(png_ptr, "color map overflow (BAD internal error)");
+
+   image->colormap_entries = cmap_entries;
+
+   /* Double check using the recorded background index */
+   switch (output_processing)
+   {
+      case PNG_CMAP_NONE:
+         if (background_index != PNG_CMAP_NONE_BACKGROUND)
+            goto bad_background;
+         break;
+
+      case PNG_CMAP_GA:
+         if (background_index != PNG_CMAP_GA_BACKGROUND)
+            goto bad_background;
+         break;
+
+      case PNG_CMAP_TRANS:
+         if (background_index >= cmap_entries ||
+            background_index != PNG_CMAP_TRANS_BACKGROUND)
+            goto bad_background;
+         break;
+
+      case PNG_CMAP_RGB:
+         if (background_index != PNG_CMAP_RGB_BACKGROUND)
+            goto bad_background;
+         break;
+
+      case PNG_CMAP_RGB_ALPHA:
+         if (background_index != PNG_CMAP_RGB_ALPHA_BACKGROUND)
+            goto bad_background;
+         break;
+
+      default:
+         png_error(png_ptr, "bad processing option (internal error)");
+
+      bad_background:
+         png_error(png_ptr, "bad background index (internal error)");
+   }
+
+   display->colormap_processing = output_processing;
+
+   return 1/*ok*/;
+}
+
+/* The final part of the color-map read called from png_image_finish_read. */
+static int
+png_image_read_and_map(png_voidp argument)
+{
+   png_image_read_control *display = png_voidcast(png_image_read_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+   int passes;
+
+   /* Called when the libpng data must be transformed into the color-mapped
+    * form.  There is a local row buffer in display->local and this routine must
+    * do the interlace handling.
+    */
+   switch (png_ptr->interlaced)
+   {
+      case PNG_INTERLACE_NONE:
+         passes = 1;
+         break;
+
+      case PNG_INTERLACE_ADAM7:
+         passes = PNG_INTERLACE_ADAM7_PASSES;
+         break;
+
+      default:
+         passes = 0;
+         png_error(png_ptr, "unknown interlace type");
+   }
+
+   {
+      png_uint_32  height = image->height;
+      png_uint_32  width = image->width;
+      int          proc = display->colormap_processing;
+      png_bytep    first_row = png_voidcast(png_bytep, display->first_row);
+      ptrdiff_t    step_row = display->row_bytes;
+      int pass;
+
+      for (pass = 0; pass < passes; ++pass)
+      {
+         unsigned int     startx, stepx, stepy;
+         png_uint_32      y;
+
+         if (png_ptr->interlaced == PNG_INTERLACE_ADAM7)
+         {
+            /* The row may be empty for a short image: */
+            if (PNG_PASS_COLS(width, pass) == 0)
+               continue;
+
+            startx = PNG_PASS_START_COL(pass);
+            stepx = PNG_PASS_COL_OFFSET(pass);
+            y = PNG_PASS_START_ROW(pass);
+            stepy = PNG_PASS_ROW_OFFSET(pass);
+         }
+
+         else
+         {
+            y = 0;
+            startx = 0;
+            stepx = stepy = 1;
+         }
+
+         for (; y<height; y += stepy)
+         {
+            png_bytep inrow = png_voidcast(png_bytep, display->local_row);
+            png_bytep outrow = first_row + y * step_row;
+            png_const_bytep end_row = outrow + width;
+
+            /* Read read the libpng data into the temporary buffer. */
+            png_read_row(png_ptr, inrow, NULL);
+
+            /* Now process the row according to the processing option, note
+             * that the caller verifies that the format of the libpng output
+             * data is as required.
+             */
+            outrow += startx;
+            switch (proc)
+            {
+               case PNG_CMAP_GA:
+                  for (; outrow < end_row; outrow += stepx)
+                  {
+                     /* The data is always in the PNG order */
+                     unsigned int gray = *inrow++;
+                     unsigned int alpha = *inrow++;
+                     unsigned int entry;
+
+                     /* NOTE: this code is copied as a comment in
+                      * make_ga_colormap above.  Please update the
+                      * comment if you change this code!
+                      */
+                     if (alpha > 229) /* opaque */
+                     {
+                        entry = (231 * gray + 128) >> 8;
+                     }
+                     else if (alpha < 26) /* transparent */
+                     {
+                        entry = 231;
+                     }
+                     else /* partially opaque */
+                     {
+                        entry = 226 + 6 * PNG_DIV51(alpha) + PNG_DIV51(gray);
+                     }
+
+                     *outrow = (png_byte)entry;
+                  }
+                  break;
+
+               case PNG_CMAP_TRANS:
+                  for (; outrow < end_row; outrow += stepx)
+                  {
+                     png_byte gray = *inrow++;
+                     png_byte alpha = *inrow++;
+
+                     if (alpha == 0)
+                        *outrow = PNG_CMAP_TRANS_BACKGROUND;
+
+                     else if (gray != PNG_CMAP_TRANS_BACKGROUND)
+                        *outrow = gray;
+
+                     else
+                        *outrow = (png_byte)(PNG_CMAP_TRANS_BACKGROUND+1);
+                  }
+                  break;
+
+               case PNG_CMAP_RGB:
+                  for (; outrow < end_row; outrow += stepx)
+                  {
+                     *outrow = PNG_RGB_INDEX(inrow[0], inrow[1], inrow[2]);
+                     inrow += 3;
+                  }
+                  break;
+
+               case PNG_CMAP_RGB_ALPHA:
+                  for (; outrow < end_row; outrow += stepx)
+                  {
+                     unsigned int alpha = inrow[3];
+
+                     /* Because the alpha entries only hold alpha==0.5 values
+                      * split the processing at alpha==0.25 (64) and 0.75
+                      * (196).
+                      */
+
+                     if (alpha >= 196)
+                        *outrow = PNG_RGB_INDEX(inrow[0], inrow[1],
+                           inrow[2]);
+
+                     else if (alpha < 64)
+                        *outrow = PNG_CMAP_RGB_ALPHA_BACKGROUND;
+
+                     else
+                     {
+                        /* Likewise there are three entries for each of r, g
+                         * and b.  We could select the entry by popcount on
+                         * the top two bits on those architectures that
+                         * support it, this is what the code below does,
+                         * crudely.
+                         */
+                        unsigned int back_i = PNG_CMAP_RGB_ALPHA_BACKGROUND+1;
+
+                        /* Here are how the values map:
+                         *
+                         * 0x00 .. 0x3f -> 0
+                         * 0x40 .. 0xbf -> 1
+                         * 0xc0 .. 0xff -> 2
+                         *
+                         * So, as above with the explicit alpha checks, the
+                         * breakpoints are at 64 and 196.
+                         */
+                        if (inrow[0] & 0x80) back_i += 9; /* red */
+                        if (inrow[0] & 0x40) back_i += 9;
+                        if (inrow[0] & 0x80) back_i += 3; /* green */
+                        if (inrow[0] & 0x40) back_i += 3;
+                        if (inrow[0] & 0x80) back_i += 1; /* blue */
+                        if (inrow[0] & 0x40) back_i += 1;
+
+                        *outrow = (png_byte)back_i;
+                     }
+
+                     inrow += 4;
+                  }
+                  break;
+
+               default:
+                  break;
+            }
+         }
+      }
+   }
+
+   return 1;
+}
+
+static int
+png_image_read_colormapped(png_voidp argument)
+{
+   png_image_read_control *display = png_voidcast(png_image_read_control*,
+      argument);
+   png_imagep image = display->image;
+   png_controlp control = image->opaque;
+   png_structrp png_ptr = control->png_ptr;
+   png_inforp info_ptr = control->info_ptr;
+
+   int passes = 0; /* As a flag */
+
+   PNG_SKIP_CHUNKS(png_ptr);
+
+   /* Update the 'info' structure and make sure the result is as required; first
+    * make sure to turn on the interlace handling if it will be required
+    * (because it can't be turned on *after* the call to png_read_update_info!)
+    */
+   if (display->colormap_processing == PNG_CMAP_NONE)
+      passes = png_set_interlace_handling(png_ptr);
+
+   png_read_update_info(png_ptr, info_ptr);
+
+   /* The expected output can be deduced from the colormap_processing option. */
+   switch (display->colormap_processing)
+   {
+      case PNG_CMAP_NONE:
+         /* Output must be one channel and one byte per pixel, the output
+          * encoding can be anything.
+          */
+         if ((info_ptr->color_type == PNG_COLOR_TYPE_PALETTE ||
+            info_ptr->color_type == PNG_COLOR_TYPE_GRAY) &&
+            info_ptr->bit_depth == 8)
+            break;
+
+         goto bad_output;
+
+      case PNG_CMAP_TRANS:
+      case PNG_CMAP_GA:
+         /* Output must be two channels and the 'G' one must be sRGB, the latter
+          * can be checked with an exact number because it should have been set
+          * to this number above!
+          */
+         if (info_ptr->color_type == PNG_COLOR_TYPE_GRAY_ALPHA &&
+            info_ptr->bit_depth == 8 &&
+            png_ptr->screen_gamma == PNG_GAMMA_sRGB &&
+            image->colormap_entries == 256)
+            break;
+
+         goto bad_output;
+
+      case PNG_CMAP_RGB:
+         /* Output must be 8-bit sRGB encoded RGB */
+         if (info_ptr->color_type == PNG_COLOR_TYPE_RGB &&
+            info_ptr->bit_depth == 8 &&
+            png_ptr->screen_gamma == PNG_GAMMA_sRGB &&
+            image->colormap_entries == 216)
+            break;
+
+         goto bad_output;
+
+      case PNG_CMAP_RGB_ALPHA:
+         /* Output must be 8-bit sRGB encoded RGBA */
+         if (info_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA &&
+            info_ptr->bit_depth == 8 &&
+            png_ptr->screen_gamma == PNG_GAMMA_sRGB &&
+            image->colormap_entries == 244 /* 216 + 1 + 27 */)
+            break;
+
+         /* goto bad_output; */
+         /* FALL THROUGH */
+
+      default:
+      bad_output:
+         png_error(png_ptr, "bad color-map processing (internal error)");
+   }
+
+   /* Now read the rows.  Do this here if it is possible to read directly into
+    * the output buffer, otherwise allocate a local row buffer of the maximum
+    * size libpng requires and call the relevant processing routine safely.
+    */
+   {
+      png_voidp first_row = display->buffer;
+      ptrdiff_t row_bytes = display->row_stride;
+
+      /* The following expression is designed to work correctly whether it gives
+       * a signed or an unsigned result.
+       */
+      if (row_bytes < 0)
+      {
+         char *ptr = png_voidcast(char*, first_row);
+         ptr += (image->height-1) * (-row_bytes);
+         first_row = png_voidcast(png_voidp, ptr);
+      }
+
+      display->first_row = first_row;
+      display->row_bytes = row_bytes;
+   }
+
+   if (passes == 0)
+   {
+      int result;
+      png_voidp row = png_malloc(png_ptr, png_get_rowbytes(png_ptr, info_ptr));
+
+      display->local_row = row;
+      result = png_safe_execute(image, png_image_read_and_map, display);
+      display->local_row = NULL;
+      png_free(png_ptr, row);
+
+      return result;
+   }
+
+   else
+   {
+      png_alloc_size_t row_bytes = display->row_bytes;
+
+      while (--passes >= 0)
+      {
+         png_uint_32      y = image->height;
+         png_bytep        row = png_voidcast(png_bytep, display->first_row);
+
+         while (y-- > 0)
+         {
+            png_read_row(png_ptr, row, NULL);
+            row += row_bytes;
+         }
+      }
+
+      return 1;
+   }
+}
+
+/* Just the row reading part of png_image_read. */
+static int
+png_image_read_composite(png_voidp argument)
+{
+   png_image_read_control *display = png_voidcast(png_image_read_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+   int passes;
+
+   switch (png_ptr->interlaced)
+   {
+      case PNG_INTERLACE_NONE:
+         passes = 1;
+         break;
+
+      case PNG_INTERLACE_ADAM7:
+         passes = PNG_INTERLACE_ADAM7_PASSES;
+         break;
+
+      default:
+         passes = 0;
+         png_error(png_ptr, "unknown interlace type");
+   }
+
+   {
+      png_uint_32  height = image->height;
+      png_uint_32  width = image->width;
+      ptrdiff_t    step_row = display->row_bytes;
+      unsigned int channels = (image->format & PNG_FORMAT_FLAG_COLOR) ? 3 : 1;
+      int pass;
+
+      for (pass = 0; pass < passes; ++pass)
+      {
+         unsigned int     startx, stepx, stepy;
+         png_uint_32      y;
+
+         if (png_ptr->interlaced == PNG_INTERLACE_ADAM7)
+         {
+            /* The row may be empty for a short image: */
+            if (PNG_PASS_COLS(width, pass) == 0)
+               continue;
+
+            startx = PNG_PASS_START_COL(pass) * channels;
+            stepx = PNG_PASS_COL_OFFSET(pass) * channels;
+            y = PNG_PASS_START_ROW(pass);
+            stepy = PNG_PASS_ROW_OFFSET(pass);
+         }
+
+         else
+         {
+            y = 0;
+            startx = 0;
+            stepx = channels;
+            stepy = 1;
+         }
+
+         for (; y<height; y += stepy)
+         {
+            png_bytep inrow = png_voidcast(png_bytep, display->local_row);
+            png_bytep outrow;
+            png_const_bytep end_row;
+
+            /* Read the row, which is packed: */
+            png_read_row(png_ptr, inrow, NULL);
+
+            outrow = png_voidcast(png_bytep, display->first_row);
+            outrow += y * step_row;
+            end_row = outrow + width * channels;
+
+            /* Now do the composition on each pixel in this row. */
+            outrow += startx;
+            for (; outrow < end_row; outrow += stepx)
+            {
+               png_byte alpha = inrow[channels];
+
+               if (alpha > 0) /* else no change to the output */
+               {
+                  unsigned int c;
+
+                  for (c=0; c<channels; ++c)
+                  {
+                     png_uint_32 component = inrow[c];
+
+                     if (alpha < 255) /* else just use component */
+                     {
+                        /* This is PNG_OPTIMIZED_ALPHA, the component value
+                         * is a linear 8-bit value.  Combine this with the
+                         * current outrow[c] value which is sRGB encoded.
+                         * Arithmetic here is 16-bits to preserve the output
+                         * values correctly.
+                         */
+                        component *= 257*255; /* =65535 */
+                        component += (255-alpha)*png_sRGB_table[outrow[c]];
+
+                        /* So 'component' is scaled by 255*65535 and is
+                         * therefore appropriate for the sRGB to linear
+                         * conversion table.
+                         */
+                        component = PNG_sRGB_FROM_LINEAR(component);
+                     }
+
+                     outrow[c] = (png_byte)component;
+                  }
+               }
+
+               inrow += channels+1; /* components and alpha channel */
+            }
+         }
+      }
+   }
+
+   return 1;
+}
+
+/* The do_local_background case; called when all the following transforms are to
+ * be done:
+ *
+ * PNG_RGB_TO_GRAY
+ * PNG_COMPOSITE
+ * PNG_GAMMA
+ *
+ * This is a work-round for the fact that both the PNG_RGB_TO_GRAY and
+ * PNG_COMPOSITE code performs gamma correction, so we get double gamma
+ * correction.  The fix-up is to prevent the PNG_COMPOSITE operation happening
+ * inside libpng, so this routine sees an 8 or 16-bit gray+alpha row and handles
+ * the removal or pre-multiplication of the alpha channel.
+ */
+static int
+png_image_read_background(png_voidp argument)
+{
+   png_image_read_control *display = png_voidcast(png_image_read_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+   png_inforp info_ptr = image->opaque->info_ptr;
+   png_uint_32 height = image->height;
+   png_uint_32 width = image->width;
+   int pass, passes;
+
+   /* Double check the convoluted logic below.  We expect to get here with
+    * libpng doing rgb to gray and gamma correction but background processing
+    * left to the png_image_read_background function.  The rows libpng produce
+    * might be 8 or 16-bit but should always have two channels; gray plus alpha.
+    */
+   if ((png_ptr->transformations & PNG_RGB_TO_GRAY) == 0)
+      png_error(png_ptr, "lost rgb to gray");
+
+   if ((png_ptr->transformations & PNG_COMPOSE) != 0)
+      png_error(png_ptr, "unexpected compose");
+
+   if (png_get_channels(png_ptr, info_ptr) != 2)
+      png_error(png_ptr, "lost/gained channels");
+
+   /* Expect the 8-bit case to always remove the alpha channel */
+   if ((image->format & PNG_FORMAT_FLAG_LINEAR) == 0 &&
+      (image->format & PNG_FORMAT_FLAG_ALPHA) != 0)
+      png_error(png_ptr, "unexpected 8-bit transformation");
+
+   switch (png_ptr->interlaced)
+   {
+      case PNG_INTERLACE_NONE:
+         passes = 1;
+         break;
+
+      case PNG_INTERLACE_ADAM7:
+         passes = PNG_INTERLACE_ADAM7_PASSES;
+         break;
+
+      default:
+         passes = 0;
+         png_error(png_ptr, "unknown interlace type");
+   }
+
+   switch (png_get_bit_depth(png_ptr, info_ptr))
+   {
+      default:
+         png_error(png_ptr, "unexpected bit depth");
+         break;
+
+      case 8:
+         /* 8-bit sRGB gray values with an alpha channel; the alpha channel is
+          * to be removed by composing on a background: either the row if
+          * display->background is NULL or display->background->green if not.
+          * Unlike the code above ALPHA_OPTIMIZED has *not* been done.
+          */
+         {
+            png_bytep first_row = png_voidcast(png_bytep, display->first_row);
+            ptrdiff_t step_row = display->row_bytes;
+
+            for (pass = 0; pass < passes; ++pass)
+            {
+               png_bytep        row = png_voidcast(png_bytep,
+                                                   display->first_row);
+               unsigned int     startx, stepx, stepy;
+               png_uint_32      y;
+
+               if (png_ptr->interlaced == PNG_INTERLACE_ADAM7)
+               {
+                  /* The row may be empty for a short image: */
+                  if (PNG_PASS_COLS(width, pass) == 0)
+                     continue;
+
+                  startx = PNG_PASS_START_COL(pass);
+                  stepx = PNG_PASS_COL_OFFSET(pass);
+                  y = PNG_PASS_START_ROW(pass);
+                  stepy = PNG_PASS_ROW_OFFSET(pass);
+               }
+
+               else
+               {
+                  y = 0;
+                  startx = 0;
+                  stepx = stepy = 1;
+               }
+
+               if (display->background == NULL)
+               {
+                  for (; y<height; y += stepy)
+                  {
+                     png_bytep inrow = png_voidcast(png_bytep,
+                        display->local_row);
+                     png_bytep outrow = first_row + y * step_row;
+                     png_const_bytep end_row = outrow + width;
+
+                     /* Read the row, which is packed: */
+                     png_read_row(png_ptr, inrow, NULL);
+
+                     /* Now do the composition on each pixel in this row. */
+                     outrow += startx;
+                     for (; outrow < end_row; outrow += stepx)
+                     {
+                        png_byte alpha = inrow[1];
+
+                        if (alpha > 0) /* else no change to the output */
+                        {
+                           png_uint_32 component = inrow[0];
+
+                           if (alpha < 255) /* else just use component */
+                           {
+                              /* Since PNG_OPTIMIZED_ALPHA was not set it is
+                               * necessary to invert the sRGB transfer
+                               * function and multiply the alpha out.
+                               */
+                              component = png_sRGB_table[component] * alpha;
+                              component += png_sRGB_table[outrow[0]] *
+                                 (255-alpha);
+                              component = PNG_sRGB_FROM_LINEAR(component);
+                           }
+
+                           outrow[0] = (png_byte)component;
+                        }
+
+                        inrow += 2; /* gray and alpha channel */
+                     }
+                  }
+               }
+
+               else /* constant background value */
+               {
+                  png_byte background8 = display->background->green;
+                  png_uint_16 background = png_sRGB_table[background8];
+
+                  for (; y<height; y += stepy)
+                  {
+                     png_bytep inrow = png_voidcast(png_bytep,
+                        display->local_row);
+                     png_bytep outrow = first_row + y * step_row;
+                     png_const_bytep end_row = outrow + width;
+
+                     /* Read the row, which is packed: */
+                     png_read_row(png_ptr, inrow, NULL);
+
+                     /* Now do the composition on each pixel in this row. */
+                     outrow += startx;
+                     for (; outrow < end_row; outrow += stepx)
+                     {
+                        png_byte alpha = inrow[1];
+
+                        if (alpha > 0) /* else use background */
+                        {
+                           png_uint_32 component = inrow[0];
+
+                           if (alpha < 255) /* else just use component */
+                           {
+                              component = png_sRGB_table[component] * alpha;
+                              component += background * (255-alpha);
+                              component = PNG_sRGB_FROM_LINEAR(component);
+                           }
+
+                           outrow[0] = (png_byte)component;
+                        }
+
+                        else
+                           outrow[0] = background8;
+
+                        inrow += 2; /* gray and alpha channel */
+                     }
+
+                     row += display->row_bytes;
+                  }
+               }
+            }
+         }
+         break;
+
+      case 16:
+         /* 16-bit linear with pre-multiplied alpha; the pre-multiplication must
+          * still be done and, maybe, the alpha channel removed.  This code also
+          * handles the alpha-first option.
+          */
+         {
+            png_uint_16p first_row = png_voidcast(png_uint_16p,
+               display->first_row);
+            /* The division by two is safe because the caller passed in a
+             * stride which was multiplied by 2 (below) to get row_bytes.
+             */
+            ptrdiff_t    step_row = display->row_bytes / 2;
+            int preserve_alpha = (image->format & PNG_FORMAT_FLAG_ALPHA) != 0;
+            unsigned int outchannels = 1+preserve_alpha;
+            int swap_alpha = 0;
+
+            if (preserve_alpha && (image->format & PNG_FORMAT_FLAG_AFIRST))
+               swap_alpha = 1;
+
+            for (pass = 0; pass < passes; ++pass)
+            {
+               unsigned int     startx, stepx, stepy;
+               png_uint_32      y;
+
+               /* The 'x' start and step are adjusted to output components here.
+                */
+               if (png_ptr->interlaced == PNG_INTERLACE_ADAM7)
+               {
+                  /* The row may be empty for a short image: */
+                  if (PNG_PASS_COLS(width, pass) == 0)
+                     continue;
+
+                  startx = PNG_PASS_START_COL(pass) * outchannels;
+                  stepx = PNG_PASS_COL_OFFSET(pass) * outchannels;
+                  y = PNG_PASS_START_ROW(pass);
+                  stepy = PNG_PASS_ROW_OFFSET(pass);
+               }
+
+               else
+               {
+                  y = 0;
+                  startx = 0;
+                  stepx = outchannels;
+                  stepy = 1;
+               }
+
+               for (; y<height; y += stepy)
+               {
+                  png_const_uint_16p inrow;
+                  png_uint_16p outrow = first_row + y*step_row;
+                  png_uint_16p end_row = outrow + width * outchannels;
+
+                  /* Read the row, which is packed: */
+                  png_read_row(png_ptr, png_voidcast(png_bytep,
+                     display->local_row), NULL);
+                  inrow = png_voidcast(png_const_uint_16p, display->local_row);
+
+                  /* Now do the pre-multiplication on each pixel in this row.
+                   */
+                  outrow += startx;
+                  for (; outrow < end_row; outrow += stepx)
+                  {
+                     png_uint_32 component = inrow[0];
+                     png_uint_16 alpha = inrow[1];
+
+                     if (alpha > 0) /* else 0 */
+                     {
+                        if (alpha < 65535) /* else just use component */
+                        {
+                           component *= alpha;
+                           component += 32767;
+                           component /= 65535;
+                        }
+                     }
+
+                     else
+                        component = 0;
+
+                     outrow[swap_alpha] = (png_uint_16)component;
+                     if (preserve_alpha)
+                        outrow[1 ^ swap_alpha] = alpha;
+
+                     inrow += 2; /* components and alpha channel */
+                  }
+               }
+            }
+         }
+         break;
+   }
+
+   return 1;
+}
+
+/* The guts of png_image_finish_read as a png_safe_execute callback. */
+static int
+png_image_read_direct(png_voidp argument)
+{
+   png_image_read_control *display = png_voidcast(png_image_read_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+   png_inforp info_ptr = image->opaque->info_ptr;
+
+   png_uint_32 format = image->format;
+   int linear = (format & PNG_FORMAT_FLAG_LINEAR) != 0;
+   int do_local_compose = 0;
+   int do_local_background = 0; /* to avoid double gamma correction bug */
+   int passes = 0;
+
+   /* Add transforms to ensure the correct output format is produced then check
+    * that the required implementation support is there.  Always expand; always
+    * need 8 bits minimum, no palette and expanded tRNS.
+    */
+   png_set_expand(png_ptr);
+
+   /* Now check the format to see if it was modified. */
+   {
+      png_uint_32 base_format = png_image_format(png_ptr) &
+         ~PNG_FORMAT_FLAG_COLORMAP /* removed by png_set_expand */;
+      png_uint_32 change = format ^ base_format;
+      png_fixed_point output_gamma;
+      int mode; /* alpha mode */
+
+      /* Do this first so that we have a record if rgb to gray is happening. */
+      if (change & PNG_FORMAT_FLAG_COLOR)
+      {
+         /* gray<->color transformation required. */
+         if (format & PNG_FORMAT_FLAG_COLOR)
+            png_set_gray_to_rgb(png_ptr);
+
+         else
+         {
+            /* libpng can't do both rgb to gray and
+             * background/pre-multiplication if there is also significant gamma
+             * correction, because both operations require linear colors and
+             * the code only supports one transform doing the gamma correction.
+             * Handle this by doing the pre-multiplication or background
+             * operation in this code, if necessary.
+             *
+             * TODO: fix this by rewriting pngrtran.c (!)
+             *
+             * For the moment (given that fixing this in pngrtran.c is an
+             * enormous change) 'do_local_background' is used to indicate that
+             * the problem exists.
+             */
+            if (base_format & PNG_FORMAT_FLAG_ALPHA)
+               do_local_background = 1/*maybe*/;
+
+            png_set_rgb_to_gray_fixed(png_ptr, PNG_ERROR_ACTION_NONE,
+               PNG_RGB_TO_GRAY_DEFAULT, PNG_RGB_TO_GRAY_DEFAULT);
+         }
+
+         change &= ~PNG_FORMAT_FLAG_COLOR;
+      }
+
+      /* Set the gamma appropriately, linear for 16-bit input, sRGB otherwise.
+       */
+      {
+         png_fixed_point input_gamma_default;
+
+         if ((base_format & PNG_FORMAT_FLAG_LINEAR) &&
+            (image->flags & PNG_IMAGE_FLAG_16BIT_sRGB) == 0)
+            input_gamma_default = PNG_GAMMA_LINEAR;
+         else
+            input_gamma_default = PNG_DEFAULT_sRGB;
+
+         /* Call png_set_alpha_mode to set the default for the input gamma; the
+          * output gamma is set by a second call below.
+          */
+         png_set_alpha_mode_fixed(png_ptr, PNG_ALPHA_PNG, input_gamma_default);
+      }
+
+      if (linear)
+      {
+         /* If there *is* an alpha channel in the input it must be multiplied
+          * out; use PNG_ALPHA_STANDARD, otherwise just use PNG_ALPHA_PNG.
+          */
+         if (base_format & PNG_FORMAT_FLAG_ALPHA)
+            mode = PNG_ALPHA_STANDARD; /* associated alpha */
+
+         else
+            mode = PNG_ALPHA_PNG;
+
+         output_gamma = PNG_GAMMA_LINEAR;
+      }
+
+      else
+      {
+         mode = PNG_ALPHA_PNG;
+         output_gamma = PNG_DEFAULT_sRGB;
+      }
+
+      /* If 'do_local_background' is set check for the presence of gamma
+       * correction; this is part of the work-round for the libpng bug
+       * described above.
+       *
+       * TODO: fix libpng and remove this.
+       */
+      if (do_local_background)
+      {
+         png_fixed_point gtest;
+
+         /* This is 'png_gamma_threshold' from pngrtran.c; the test used for
+          * gamma correction, the screen gamma hasn't been set on png_struct
+          * yet; it's set below.  png_struct::gamma, however, is set to the
+          * final value.
+          */
+         if (png_muldiv(&gtest, output_gamma, png_ptr->colorspace.gamma,
+               PNG_FP_1) && !png_gamma_significant(gtest))
+            do_local_background = 0;
+
+         else if (mode == PNG_ALPHA_STANDARD)
+         {
+            do_local_background = 2/*required*/;
+            mode = PNG_ALPHA_PNG; /* prevent libpng doing it */
+         }
+
+         /* else leave as 1 for the checks below */
+      }
+
+      /* If the bit-depth changes then handle that here. */
+      if (change & PNG_FORMAT_FLAG_LINEAR)
+      {
+         if (linear /*16-bit output*/)
+            png_set_expand_16(png_ptr);
+
+         else /* 8-bit output */
+            png_set_scale_16(png_ptr);
+
+         change &= ~PNG_FORMAT_FLAG_LINEAR;
+      }
+
+      /* Now the background/alpha channel changes. */
+      if (change & PNG_FORMAT_FLAG_ALPHA)
+      {
+         /* Removing an alpha channel requires composition for the 8-bit
+          * formats; for the 16-bit it is already done, above, by the
+          * pre-multiplication and the channel just needs to be stripped.
+          */
+         if (base_format & PNG_FORMAT_FLAG_ALPHA)
+         {
+            /* If RGB->gray is happening the alpha channel must be left and the
+             * operation completed locally.
+             *
+             * TODO: fix libpng and remove this.
+             */
+            if (do_local_background)
+               do_local_background = 2/*required*/;
+
+            /* 16-bit output: just remove the channel */
+            else if (linear) /* compose on black (well, pre-multiply) */
+               png_set_strip_alpha(png_ptr);
+
+            /* 8-bit output: do an appropriate compose */
+            else if (display->background != NULL)
+            {
+               png_color_16 c;
+
+               c.index = 0; /*unused*/
+               c.red = display->background->red;
+               c.green = display->background->green;
+               c.blue = display->background->blue;
+               c.gray = display->background->green;
+
+               /* This is always an 8-bit sRGB value, using the 'green' channel
+                * for gray is much better than calculating the luminance here;
+                * we can get off-by-one errors in that calculation relative to
+                * the app expectations and that will show up in transparent
+                * pixels.
+                */
+               png_set_background_fixed(png_ptr, &c,
+                  PNG_BACKGROUND_GAMMA_SCREEN, 0/*need_expand*/,
+                  0/*gamma: not used*/);
+            }
+
+            else /* compose on row: implemented below. */
+            {
+               do_local_compose = 1;
+               /* This leaves the alpha channel in the output, so it has to be
+                * removed by the code below.  Set the encoding to the 'OPTIMIZE'
+                * one so the code only has to hack on the pixels that require
+                * composition.
+                */
+               mode = PNG_ALPHA_OPTIMIZED;
+            }
+         }
+
+         else /* output needs an alpha channel */
+         {
+            /* This is tricky because it happens before the swap operation has
+             * been accomplished; however, the swap does *not* swap the added
+             * alpha channel (weird API), so it must be added in the correct
+             * place.
+             */
+            png_uint_32 filler; /* opaque filler */
+            int where;
+
+            if (linear)
+               filler = 65535;
+
+            else
+               filler = 255;
+
+#           ifdef PNG_FORMAT_AFIRST_SUPPORTED
+               if (format & PNG_FORMAT_FLAG_AFIRST)
+               {
+                  where = PNG_FILLER_BEFORE;
+                  change &= ~PNG_FORMAT_FLAG_AFIRST;
+               }
+
+               else
+#           endif
+               where = PNG_FILLER_AFTER;
+
+            png_set_add_alpha(png_ptr, filler, where);
+         }
+
+         /* This stops the (irrelevant) call to swap_alpha below. */
+         change &= ~PNG_FORMAT_FLAG_ALPHA;
+      }
+
+      /* Now set the alpha mode correctly; this is always done, even if there is
+       * no alpha channel in either the input or the output because it correctly
+       * sets the output gamma.
+       */
+      png_set_alpha_mode_fixed(png_ptr, mode, output_gamma);
+
+#     ifdef PNG_FORMAT_BGR_SUPPORTED
+         if (change & PNG_FORMAT_FLAG_BGR)
+         {
+            /* Check only the output format; PNG is never BGR; don't do this if
+             * the output is gray, but fix up the 'format' value in that case.
+             */
+            if (format & PNG_FORMAT_FLAG_COLOR)
+               png_set_bgr(png_ptr);
+
+            else
+               format &= ~PNG_FORMAT_FLAG_BGR;
+
+            change &= ~PNG_FORMAT_FLAG_BGR;
+         }
+#     endif
+
+#     ifdef PNG_FORMAT_AFIRST_SUPPORTED
+         if (change & PNG_FORMAT_FLAG_AFIRST)
+         {
+            /* Only relevant if there is an alpha channel - it's particularly
+             * important to handle this correctly because do_local_compose may
+             * be set above and then libpng will keep the alpha channel for this
+             * code to remove.
+             */
+            if (format & PNG_FORMAT_FLAG_ALPHA)
+            {
+               /* Disable this if doing a local background,
+                * TODO: remove this when local background is no longer required.
+                */
+               if (do_local_background != 2)
+                  png_set_swap_alpha(png_ptr);
+            }
+
+            else
+               format &= ~PNG_FORMAT_FLAG_AFIRST;
+
+            change &= ~PNG_FORMAT_FLAG_AFIRST;
+         }
+#     endif
+
+      /* If the *output* is 16-bit then we need to check for a byte-swap on this
+       * architecture.
+       */
+      if (linear)
+      {
+         PNG_CONST png_uint_16 le = 0x0001;
+
+         if (*(png_const_bytep)&le)
+            png_set_swap(png_ptr);
+      }
+
+      /* If change is not now 0 some transformation is missing - error out. */
+      if (change)
+         png_error(png_ptr, "png_read_image: unsupported transformation");
+   }
+
+   PNG_SKIP_CHUNKS(png_ptr);
+
+   /* Update the 'info' structure and make sure the result is as required; first
+    * make sure to turn on the interlace handling if it will be required
+    * (because it can't be turned on *after* the call to png_read_update_info!)
+    *
+    * TODO: remove the do_local_background fixup below.
+    */
+   if (!do_local_compose && do_local_background != 2)
+      passes = png_set_interlace_handling(png_ptr);
+
+   png_read_update_info(png_ptr, info_ptr);
+
+   {
+      png_uint_32 info_format = 0;
+
+      if (info_ptr->color_type & PNG_COLOR_MASK_COLOR)
+         info_format |= PNG_FORMAT_FLAG_COLOR;
+
+      if (info_ptr->color_type & PNG_COLOR_MASK_ALPHA)
+      {
+         /* do_local_compose removes this channel below. */
+         if (!do_local_compose)
+         {
+            /* do_local_background does the same if required. */
+            if (do_local_background != 2 ||
+               (format & PNG_FORMAT_FLAG_ALPHA) != 0)
+               info_format |= PNG_FORMAT_FLAG_ALPHA;
+         }
+      }
+
+      else if (do_local_compose) /* internal error */
+         png_error(png_ptr, "png_image_read: alpha channel lost");
+
+      if (info_ptr->bit_depth == 16)
+         info_format |= PNG_FORMAT_FLAG_LINEAR;
+
+#     ifdef PNG_FORMAT_BGR_SUPPORTED
+         if (png_ptr->transformations & PNG_BGR)
+            info_format |= PNG_FORMAT_FLAG_BGR;
+#     endif
+
+#     ifdef PNG_FORMAT_AFIRST_SUPPORTED
+         if (do_local_background == 2)
+         {
+            if (format & PNG_FORMAT_FLAG_AFIRST)
+               info_format |= PNG_FORMAT_FLAG_AFIRST;
+         }
+
+         if ((png_ptr->transformations & PNG_SWAP_ALPHA) != 0 ||
+            ((png_ptr->transformations & PNG_ADD_ALPHA) != 0 &&
+            (png_ptr->flags & PNG_FLAG_FILLER_AFTER) == 0))
+         {
+            if (do_local_background == 2)
+               png_error(png_ptr, "unexpected alpha swap transformation");
+
+            info_format |= PNG_FORMAT_FLAG_AFIRST;
+         }
+#     endif
+
+      /* This is actually an internal error. */
+      if (info_format != format)
+         png_error(png_ptr, "png_read_image: invalid transformations");
+   }
+
+   /* Now read the rows.  If do_local_compose is set then it is necessary to use
+    * a local row buffer.  The output will be GA, RGBA or BGRA and must be
+    * converted to G, RGB or BGR as appropriate.  The 'local_row' member of the
+    * display acts as a flag.
+    */
+   {
+      png_voidp first_row = display->buffer;
+      ptrdiff_t row_bytes = display->row_stride;
+
+      if (linear)
+         row_bytes *= 2;
+
+      /* The following expression is designed to work correctly whether it gives
+       * a signed or an unsigned result.
+       */
+      if (row_bytes < 0)
+      {
+         char *ptr = png_voidcast(char*, first_row);
+         ptr += (image->height-1) * (-row_bytes);
+         first_row = png_voidcast(png_voidp, ptr);
+      }
+
+      display->first_row = first_row;
+      display->row_bytes = row_bytes;
+   }
+
+   if (do_local_compose)
+   {
+      int result;
+      png_voidp row = png_malloc(png_ptr, png_get_rowbytes(png_ptr, info_ptr));
+
+      display->local_row = row;
+      result = png_safe_execute(image, png_image_read_composite, display);
+      display->local_row = NULL;
+      png_free(png_ptr, row);
+
+      return result;
+   }
+
+   else if (do_local_background == 2)
+   {
+      int result;
+      png_voidp row = png_malloc(png_ptr, png_get_rowbytes(png_ptr, info_ptr));
+
+      display->local_row = row;
+      result = png_safe_execute(image, png_image_read_background, display);
+      display->local_row = NULL;
+      png_free(png_ptr, row);
+
+      return result;
+   }
+
+   else
+   {
+      png_alloc_size_t row_bytes = display->row_bytes;
+
+      while (--passes >= 0)
+      {
+         png_uint_32      y = image->height;
+         png_bytep        row = png_voidcast(png_bytep, display->first_row);
+
+         while (y-- > 0)
+         {
+            png_read_row(png_ptr, row, NULL);
+            row += row_bytes;
+         }
+      }
+
+      return 1;
+   }
+}
+
+int PNGAPI
+png_image_finish_read(png_imagep image, png_const_colorp background,
+   void *buffer, png_int_32 row_stride, void *colormap)
+{
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      png_uint_32 check;
+
+      if (row_stride == 0)
+         row_stride = PNG_IMAGE_ROW_STRIDE(*image);
+
+      if (row_stride < 0)
+         check = -row_stride;
+
+      else
+         check = row_stride;
+
+      if (image->opaque != NULL && buffer != NULL &&
+         check >= PNG_IMAGE_ROW_STRIDE(*image))
+      {
+         if ((image->format & PNG_FORMAT_FLAG_COLORMAP) == 0 ||
+            (image->colormap_entries > 0 && colormap != NULL))
+         {
+            int result;
+            png_image_read_control display;
+
+            memset(&display, 0, (sizeof display));
+            display.image = image;
+            display.buffer = buffer;
+            display.row_stride = row_stride;
+            display.colormap = colormap;
+            display.background = background;
+            display.local_row = NULL;
+
+            /* Choose the correct 'end' routine; for the color-map case all the
+             * setup has already been done.
+             */
+            if (image->format & PNG_FORMAT_FLAG_COLORMAP)
+               result =
+                  png_safe_execute(image, png_image_read_colormap, &display) &&
+                  png_safe_execute(image, png_image_read_colormapped, &display);
+
+            else
+               result =
+                  png_safe_execute(image, png_image_read_direct, &display);
+
+            png_image_free(image);
+            return result;
+         }
+
+         else
+            return png_image_error(image,
+               "png_image_finish_read[color-map]: no color-map");
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_finish_read: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_finish_read: damaged PNG_IMAGE_VERSION");
+
+   return 0;
+}
+
+#endif /* PNG_SIMPLIFIED_READ_SUPPORTED */
+#endif /* PNG_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngrio.c b/ml/dlib/dlib/external/libpng/pngrio.c
new file mode 100644
index 000000000..d7864407b
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngrio.c
@@ -0,0 +1,118 @@
+
+/* pngrio.c - functions for data input
+ *
+ * Last changed in libpng 1.6.0 [February 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file provides a location for all input.  Users who need
+ * special handling are expected to write a function that has the same
+ * arguments as this and performs a similar function, but that possibly
+ * has a different input method.  Note that you shouldn't change this
+ * function, but rather write a replacement function and then make
+ * libpng use it at run time with png_set_read_fn(...).
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+/* Read the data from whatever input you are using.  The default routine
+ * reads from a file pointer.  Note that this routine sometimes gets called
+ * with very small lengths, so you should implement some kind of simple
+ * buffering if you are using unbuffered reads.  This should never be asked
+ * to read more then 64K on a 16 bit machine.
+ */
+void /* PRIVATE */
+png_read_data(png_structrp png_ptr, png_bytep data, png_size_t length)
+{
+   png_debug1(4, "reading %d bytes", (int)length);
+
+   if (png_ptr->read_data_fn != NULL)
+      (*(png_ptr->read_data_fn))(png_ptr, data, length);
+
+   else
+      png_error(png_ptr, "Call to NULL read function");
+}
+
+#ifdef PNG_STDIO_SUPPORTED
+/* This is the function that does the actual reading of data.  If you are
+ * not reading from a standard C stream, you should create a replacement
+ * read_data function and use it at run time with png_set_read_fn(), rather
+ * than changing the library.
+ */
+void PNGCBAPI
+png_default_read_data(png_structp png_ptr, png_bytep data, png_size_t length)
+{
+   png_size_t check;
+
+   if (png_ptr == NULL)
+      return;
+
+   /* fread() returns 0 on error, so it is OK to store this in a png_size_t
+    * instead of an int, which is what fread() actually returns.
+    */
+   check = fread(data, 1, length, png_voidcast(png_FILE_p, png_ptr->io_ptr));
+
+   if (check != length)
+      png_error(png_ptr, "Read Error");
+}
+#endif
+
+/* This function allows the application to supply a new input function
+ * for libpng if standard C streams aren't being used.
+ *
+ * This function takes as its arguments:
+ *
+ * png_ptr      - pointer to a png input data structure
+ *
+ * io_ptr       - pointer to user supplied structure containing info about
+ *                the input functions.  May be NULL.
+ *
+ * read_data_fn - pointer to a new input function that takes as its
+ *                arguments a pointer to a png_struct, a pointer to
+ *                a location where input data can be stored, and a 32-bit
+ *                unsigned int that is the number of bytes to be read.
+ *                To exit and output any fatal error messages the new write
+ *                function should call png_error(png_ptr, "Error msg").
+ *                May be NULL, in which case libpng's default function will
+ *                be used.
+ */
+void PNGAPI
+png_set_read_fn(png_structrp png_ptr, png_voidp io_ptr,
+   png_rw_ptr read_data_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->io_ptr = io_ptr;
+
+#ifdef PNG_STDIO_SUPPORTED
+   if (read_data_fn != NULL)
+      png_ptr->read_data_fn = read_data_fn;
+
+   else
+      png_ptr->read_data_fn = png_default_read_data;
+#else
+   png_ptr->read_data_fn = read_data_fn;
+#endif
+
+   /* It is an error to write to a read device */
+   if (png_ptr->write_data_fn != NULL)
+   {
+      png_ptr->write_data_fn = NULL;
+      png_warning(png_ptr,
+          "Can't set both read_data_fn and write_data_fn in the"
+          " same structure");
+   }
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+   png_ptr->output_flush_fn = NULL;
+#endif
+}
+#endif /* PNG_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngrtran.c b/ml/dlib/dlib/external/libpng/pngrtran.c
new file mode 100644
index 000000000..3b7d484fc
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngrtran.c
@@ -0,0 +1,5110 @@
+
+/* pngrtran.c - transforms the data in a row for PNG readers
+ *
+ * Last changed in libpng 1.6.4 [August 21, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file contains functions optionally called by an application
+ * in order to tell libpng how to handle data when reading a PNG.
+ * Transformations that are used in both reading and writing are
+ * in pngtrans.c.
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+/* Set the action on getting a CRC error for an ancillary or critical chunk. */
+void PNGAPI
+png_set_crc_action(png_structrp png_ptr, int crit_action, int ancil_action)
+{
+   png_debug(1, "in png_set_crc_action");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* Tell libpng how we react to CRC errors in critical chunks */
+   switch (crit_action)
+   {
+      case PNG_CRC_NO_CHANGE:                        /* Leave setting as is */
+         break;
+
+      case PNG_CRC_WARN_USE:                               /* Warn/use data */
+         png_ptr->flags &= ~PNG_FLAG_CRC_CRITICAL_MASK;
+         png_ptr->flags |= PNG_FLAG_CRC_CRITICAL_USE;
+         break;
+
+      case PNG_CRC_QUIET_USE:                             /* Quiet/use data */
+         png_ptr->flags &= ~PNG_FLAG_CRC_CRITICAL_MASK;
+         png_ptr->flags |= PNG_FLAG_CRC_CRITICAL_USE |
+                           PNG_FLAG_CRC_CRITICAL_IGNORE;
+         break;
+
+      case PNG_CRC_WARN_DISCARD:    /* Not a valid action for critical data */
+         png_warning(png_ptr,
+            "Can't discard critical data on CRC error");
+      case PNG_CRC_ERROR_QUIT:                                /* Error/quit */
+
+      case PNG_CRC_DEFAULT:
+      default:
+         png_ptr->flags &= ~PNG_FLAG_CRC_CRITICAL_MASK;
+         break;
+   }
+
+   /* Tell libpng how we react to CRC errors in ancillary chunks */
+   switch (ancil_action)
+   {
+      case PNG_CRC_NO_CHANGE:                       /* Leave setting as is */
+         break;
+
+      case PNG_CRC_WARN_USE:                              /* Warn/use data */
+         png_ptr->flags &= ~PNG_FLAG_CRC_ANCILLARY_MASK;
+         png_ptr->flags |= PNG_FLAG_CRC_ANCILLARY_USE;
+         break;
+
+      case PNG_CRC_QUIET_USE:                            /* Quiet/use data */
+         png_ptr->flags &= ~PNG_FLAG_CRC_ANCILLARY_MASK;
+         png_ptr->flags |= PNG_FLAG_CRC_ANCILLARY_USE |
+                           PNG_FLAG_CRC_ANCILLARY_NOWARN;
+         break;
+
+      case PNG_CRC_ERROR_QUIT:                               /* Error/quit */
+         png_ptr->flags &= ~PNG_FLAG_CRC_ANCILLARY_MASK;
+         png_ptr->flags |= PNG_FLAG_CRC_ANCILLARY_NOWARN;
+         break;
+
+      case PNG_CRC_WARN_DISCARD:                      /* Warn/discard data */
+
+      case PNG_CRC_DEFAULT:
+      default:
+         png_ptr->flags &= ~PNG_FLAG_CRC_ANCILLARY_MASK;
+         break;
+   }
+}
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+/* Is it OK to set a transformation now?  Only if png_start_read_image or
+ * png_read_update_info have not been called.  It is not necessary for the IHDR
+ * to have been read in all cases, the parameter allows for this check too.
+ */
+static int
+png_rtran_ok(png_structrp png_ptr, int need_IHDR)
+{
+   if (png_ptr != NULL)
+   {
+      if (png_ptr->flags & PNG_FLAG_ROW_INIT)
+         png_app_error(png_ptr,
+            "invalid after png_start_read_image or png_read_update_info");
+
+      else if (need_IHDR && (png_ptr->mode & PNG_HAVE_IHDR) == 0)
+         png_app_error(png_ptr, "invalid before the PNG header has been read");
+
+      else
+      {
+         /* Turn on failure to initialize correctly for all transforms. */
+         png_ptr->flags |= PNG_FLAG_DETECT_UNINITIALIZED;
+
+         return 1; /* Ok */
+      }
+   }
+
+   return 0; /* no png_error possible! */
+}
+#endif
+
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+/* Handle alpha and tRNS via a background color */
+void PNGFAPI
+png_set_background_fixed(png_structrp png_ptr,
+    png_const_color_16p background_color, int background_gamma_code,
+    int need_expand, png_fixed_point background_gamma)
+{
+   png_debug(1, "in png_set_background_fixed");
+
+   if (!png_rtran_ok(png_ptr, 0) || background_color == NULL)
+      return;
+
+   if (background_gamma_code == PNG_BACKGROUND_GAMMA_UNKNOWN)
+   {
+      png_warning(png_ptr, "Application must supply a known background gamma");
+      return;
+   }
+
+   png_ptr->transformations |= PNG_COMPOSE | PNG_STRIP_ALPHA;
+   png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+   png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+
+   png_ptr->background = *background_color;
+   png_ptr->background_gamma = background_gamma;
+   png_ptr->background_gamma_type = (png_byte)(background_gamma_code);
+   if (need_expand)
+      png_ptr->transformations |= PNG_BACKGROUND_EXPAND;
+   else
+      png_ptr->transformations &= ~PNG_BACKGROUND_EXPAND;
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_background(png_structrp png_ptr,
+    png_const_color_16p background_color, int background_gamma_code,
+    int need_expand, double background_gamma)
+{
+   png_set_background_fixed(png_ptr, background_color, background_gamma_code,
+      need_expand, png_fixed(png_ptr, background_gamma, "png_set_background"));
+}
+#  endif  /* FLOATING_POINT */
+#endif /* READ_BACKGROUND */
+
+/* Scale 16-bit depth files to 8-bit depth.  If both of these are set then the
+ * one that pngrtran does first (scale) happens.  This is necessary to allow the
+ * TRANSFORM and API behavior to be somewhat consistent, and it's simpler.
+ */
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+void PNGAPI
+png_set_scale_16(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_scale_16");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= PNG_SCALE_16_TO_8;
+}
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+/* Chop 16-bit depth files to 8-bit depth */
+void PNGAPI
+png_set_strip_16(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_strip_16");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= PNG_16_TO_8;
+}
+#endif
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+void PNGAPI
+png_set_strip_alpha(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_strip_alpha");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= PNG_STRIP_ALPHA;
+}
+#endif
+
+#if defined(PNG_READ_ALPHA_MODE_SUPPORTED) || defined(PNG_READ_GAMMA_SUPPORTED)
+static png_fixed_point
+translate_gamma_flags(png_structrp png_ptr, png_fixed_point output_gamma,
+   int is_screen)
+{
+   /* Check for flag values.  The main reason for having the old Mac value as a
+    * flag is that it is pretty near impossible to work out what the correct
+    * value is from Apple documentation - a working Mac system is needed to
+    * discover the value!
+    */
+   if (output_gamma == PNG_DEFAULT_sRGB ||
+      output_gamma == PNG_FP_1 / PNG_DEFAULT_sRGB)
+   {
+      /* If there is no sRGB support this just sets the gamma to the standard
+       * sRGB value.  (This is a side effect of using this function!)
+       */
+#     ifdef PNG_READ_sRGB_SUPPORTED
+         png_ptr->flags |= PNG_FLAG_ASSUME_sRGB;
+#     else
+         PNG_UNUSED(png_ptr)
+#     endif
+      if (is_screen)
+         output_gamma = PNG_GAMMA_sRGB;
+      else
+         output_gamma = PNG_GAMMA_sRGB_INVERSE;
+   }
+
+   else if (output_gamma == PNG_GAMMA_MAC_18 ||
+      output_gamma == PNG_FP_1 / PNG_GAMMA_MAC_18)
+   {
+      if (is_screen)
+         output_gamma = PNG_GAMMA_MAC_OLD;
+      else
+         output_gamma = PNG_GAMMA_MAC_INVERSE;
+   }
+
+   return output_gamma;
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+static png_fixed_point
+convert_gamma_value(png_structrp png_ptr, double output_gamma)
+{
+   /* The following silently ignores cases where fixed point (times 100,000)
+    * gamma values are passed to the floating point API.  This is safe and it
+    * means the fixed point constants work just fine with the floating point
+    * API.  The alternative would just lead to undetected errors and spurious
+    * bug reports.  Negative values fail inside the _fixed API unless they
+    * correspond to the flag values.
+    */
+   if (output_gamma > 0 && output_gamma < 128)
+      output_gamma *= PNG_FP_1;
+
+   /* This preserves -1 and -2 exactly: */
+   output_gamma = floor(output_gamma + .5);
+
+   if (output_gamma > PNG_FP_MAX || output_gamma < PNG_FP_MIN)
+      png_fixed_error(png_ptr, "gamma value");
+
+   return (png_fixed_point)output_gamma;
+}
+#  endif
+#endif /* READ_ALPHA_MODE || READ_GAMMA */
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+void PNGFAPI
+png_set_alpha_mode_fixed(png_structrp png_ptr, int mode,
+   png_fixed_point output_gamma)
+{
+   int compose = 0;
+   png_fixed_point file_gamma;
+
+   png_debug(1, "in png_set_alpha_mode");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   output_gamma = translate_gamma_flags(png_ptr, output_gamma, 1/*screen*/);
+
+   /* Validate the value to ensure it is in a reasonable range. The value
+    * is expected to be 1 or greater, but this range test allows for some
+    * viewing correction values.  The intent is to weed out users of this API
+    * who use the inverse of the gamma value accidentally!  Since some of these
+    * values are reasonable this may have to be changed.
+    */
+   if (output_gamma < 70000 || output_gamma > 300000)
+      png_error(png_ptr, "output gamma out of expected range");
+
+   /* The default file gamma is the inverse of the output gamma; the output
+    * gamma may be changed below so get the file value first:
+    */
+   file_gamma = png_reciprocal(output_gamma);
+
+   /* There are really 8 possibilities here, composed of any combination
+    * of:
+    *
+    *    premultiply the color channels
+    *    do not encode non-opaque pixels
+    *    encode the alpha as well as the color channels
+    *
+    * The differences disappear if the input/output ('screen') gamma is 1.0,
+    * because then the encoding is a no-op and there is only the choice of
+    * premultiplying the color channels or not.
+    *
+    * png_set_alpha_mode and png_set_background interact because both use
+    * png_compose to do the work.  Calling both is only useful when
+    * png_set_alpha_mode is used to set the default mode - PNG_ALPHA_PNG - along
+    * with a default gamma value.  Otherwise PNG_COMPOSE must not be set.
+    */
+   switch (mode)
+   {
+      case PNG_ALPHA_PNG:        /* default: png standard */
+         /* No compose, but it may be set by png_set_background! */
+         png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+         png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+         break;
+
+      case PNG_ALPHA_ASSOCIATED: /* color channels premultiplied */
+         compose = 1;
+         png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+         png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+         /* The output is linear: */
+         output_gamma = PNG_FP_1;
+         break;
+
+      case PNG_ALPHA_OPTIMIZED:  /* associated, non-opaque pixels linear */
+         compose = 1;
+         png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+         png_ptr->flags |= PNG_FLAG_OPTIMIZE_ALPHA;
+         /* output_gamma records the encoding of opaque pixels! */
+         break;
+
+      case PNG_ALPHA_BROKEN:     /* associated, non-linear, alpha encoded */
+         compose = 1;
+         png_ptr->transformations |= PNG_ENCODE_ALPHA;
+         png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+         break;
+
+      default:
+         png_error(png_ptr, "invalid alpha mode");
+   }
+
+   /* Only set the default gamma if the file gamma has not been set (this has
+    * the side effect that the gamma in a second call to png_set_alpha_mode will
+    * be ignored.)
+    */
+   if (png_ptr->colorspace.gamma == 0)
+   {
+      png_ptr->colorspace.gamma = file_gamma;
+      png_ptr->colorspace.flags |= PNG_COLORSPACE_HAVE_GAMMA;
+   }
+
+   /* But always set the output gamma: */
+   png_ptr->screen_gamma = output_gamma;
+
+   /* Finally, if pre-multiplying, set the background fields to achieve the
+    * desired result.
+    */
+   if (compose)
+   {
+      /* And obtain alpha pre-multiplication by composing on black: */
+      memset(&png_ptr->background, 0, (sizeof png_ptr->background));
+      png_ptr->background_gamma = png_ptr->colorspace.gamma; /* just in case */
+      png_ptr->background_gamma_type = PNG_BACKGROUND_GAMMA_FILE;
+      png_ptr->transformations &= ~PNG_BACKGROUND_EXPAND;
+
+      if (png_ptr->transformations & PNG_COMPOSE)
+         png_error(png_ptr,
+            "conflicting calls to set alpha mode and background");
+
+      png_ptr->transformations |= PNG_COMPOSE;
+   }
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_alpha_mode(png_structrp png_ptr, int mode, double output_gamma)
+{
+   png_set_alpha_mode_fixed(png_ptr, mode, convert_gamma_value(png_ptr,
+      output_gamma));
+}
+#  endif
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+/* Dither file to 8-bit.  Supply a palette, the current number
+ * of elements in the palette, the maximum number of elements
+ * allowed, and a histogram if possible.  If the current number
+ * of colors is greater then the maximum number, the palette will be
+ * modified to fit in the maximum number.  "full_quantize" indicates
+ * whether we need a quantizing cube set up for RGB images, or if we
+ * simply are reducing the number of colors in a paletted image.
+ */
+
+typedef struct png_dsort_struct
+{
+   struct png_dsort_struct * next;
+   png_byte left;
+   png_byte right;
+} png_dsort;
+typedef png_dsort *   png_dsortp;
+typedef png_dsort * * png_dsortpp;
+
+void PNGAPI
+png_set_quantize(png_structrp png_ptr, png_colorp palette,
+    int num_palette, int maximum_colors, png_const_uint_16p histogram,
+    int full_quantize)
+{
+   png_debug(1, "in png_set_quantize");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= PNG_QUANTIZE;
+
+   if (!full_quantize)
+   {
+      int i;
+
+      png_ptr->quantize_index = (png_bytep)png_malloc(png_ptr,
+          (png_uint_32)(num_palette * (sizeof (png_byte))));
+      for (i = 0; i < num_palette; i++)
+         png_ptr->quantize_index[i] = (png_byte)i;
+   }
+
+   if (num_palette > maximum_colors)
+   {
+      if (histogram != NULL)
+      {
+         /* This is easy enough, just throw out the least used colors.
+          * Perhaps not the best solution, but good enough.
+          */
+
+         int i;
+
+         /* Initialize an array to sort colors */
+         png_ptr->quantize_sort = (png_bytep)png_malloc(png_ptr,
+             (png_uint_32)(num_palette * (sizeof (png_byte))));
+
+         /* Initialize the quantize_sort array */
+         for (i = 0; i < num_palette; i++)
+            png_ptr->quantize_sort[i] = (png_byte)i;
+
+         /* Find the least used palette entries by starting a
+          * bubble sort, and running it until we have sorted
+          * out enough colors.  Note that we don't care about
+          * sorting all the colors, just finding which are
+          * least used.
+          */
+
+         for (i = num_palette - 1; i >= maximum_colors; i--)
+         {
+            int done; /* To stop early if the list is pre-sorted */
+            int j;
+
+            done = 1;
+            for (j = 0; j < i; j++)
+            {
+               if (histogram[png_ptr->quantize_sort[j]]
+                   < histogram[png_ptr->quantize_sort[j + 1]])
+               {
+                  png_byte t;
+
+                  t = png_ptr->quantize_sort[j];
+                  png_ptr->quantize_sort[j] = png_ptr->quantize_sort[j + 1];
+                  png_ptr->quantize_sort[j + 1] = t;
+                  done = 0;
+               }
+            }
+
+            if (done)
+               break;
+         }
+
+         /* Swap the palette around, and set up a table, if necessary */
+         if (full_quantize)
+         {
+            int j = num_palette;
+
+            /* Put all the useful colors within the max, but don't
+             * move the others.
+             */
+            for (i = 0; i < maximum_colors; i++)
+            {
+               if ((int)png_ptr->quantize_sort[i] >= maximum_colors)
+               {
+                  do
+                     j--;
+                  while ((int)png_ptr->quantize_sort[j] >= maximum_colors);
+
+                  palette[i] = palette[j];
+               }
+            }
+         }
+         else
+         {
+            int j = num_palette;
+
+            /* Move all the used colors inside the max limit, and
+             * develop a translation table.
+             */
+            for (i = 0; i < maximum_colors; i++)
+            {
+               /* Only move the colors we need to */
+               if ((int)png_ptr->quantize_sort[i] >= maximum_colors)
+               {
+                  png_color tmp_color;
+
+                  do
+                     j--;
+                  while ((int)png_ptr->quantize_sort[j] >= maximum_colors);
+
+                  tmp_color = palette[j];
+                  palette[j] = palette[i];
+                  palette[i] = tmp_color;
+                  /* Indicate where the color went */
+                  png_ptr->quantize_index[j] = (png_byte)i;
+                  png_ptr->quantize_index[i] = (png_byte)j;
+               }
+            }
+
+            /* Find closest color for those colors we are not using */
+            for (i = 0; i < num_palette; i++)
+            {
+               if ((int)png_ptr->quantize_index[i] >= maximum_colors)
+               {
+                  int min_d, k, min_k, d_index;
+
+                  /* Find the closest color to one we threw out */
+                  d_index = png_ptr->quantize_index[i];
+                  min_d = PNG_COLOR_DIST(palette[d_index], palette[0]);
+                  for (k = 1, min_k = 0; k < maximum_colors; k++)
+                  {
+                     int d;
+
+                     d = PNG_COLOR_DIST(palette[d_index], palette[k]);
+
+                     if (d < min_d)
+                     {
+                        min_d = d;
+                        min_k = k;
+                     }
+                  }
+                  /* Point to closest color */
+                  png_ptr->quantize_index[i] = (png_byte)min_k;
+               }
+            }
+         }
+         png_free(png_ptr, png_ptr->quantize_sort);
+         png_ptr->quantize_sort = NULL;
+      }
+      else
+      {
+         /* This is much harder to do simply (and quickly).  Perhaps
+          * we need to go through a median cut routine, but those
+          * don't always behave themselves with only a few colors
+          * as input.  So we will just find the closest two colors,
+          * and throw out one of them (chosen somewhat randomly).
+          * [We don't understand this at all, so if someone wants to
+          *  work on improving it, be our guest - AED, GRP]
+          */
+         int i;
+         int max_d;
+         int num_new_palette;
+         png_dsortp t;
+         png_dsortpp hash;
+
+         t = NULL;
+
+         /* Initialize palette index arrays */
+         png_ptr->index_to_palette = (png_bytep)png_malloc(png_ptr,
+             (png_uint_32)(num_palette * (sizeof (png_byte))));
+         png_ptr->palette_to_index = (png_bytep)png_malloc(png_ptr,
+             (png_uint_32)(num_palette * (sizeof (png_byte))));
+
+         /* Initialize the sort array */
+         for (i = 0; i < num_palette; i++)
+         {
+            png_ptr->index_to_palette[i] = (png_byte)i;
+            png_ptr->palette_to_index[i] = (png_byte)i;
+         }
+
+         hash = (png_dsortpp)png_calloc(png_ptr, (png_uint_32)(769 *
+             (sizeof (png_dsortp))));
+
+         num_new_palette = num_palette;
+
+         /* Initial wild guess at how far apart the farthest pixel
+          * pair we will be eliminating will be.  Larger
+          * numbers mean more areas will be allocated, Smaller
+          * numbers run the risk of not saving enough data, and
+          * having to do this all over again.
+          *
+          * I have not done extensive checking on this number.
+          */
+         max_d = 96;
+
+         while (num_new_palette > maximum_colors)
+         {
+            for (i = 0; i < num_new_palette - 1; i++)
+            {
+               int j;
+
+               for (j = i + 1; j < num_new_palette; j++)
+               {
+                  int d;
+
+                  d = PNG_COLOR_DIST(palette[i], palette[j]);
+
+                  if (d <= max_d)
+                  {
+
+                     t = (png_dsortp)png_malloc_warn(png_ptr,
+                         (png_uint_32)(sizeof (png_dsort)));
+
+                     if (t == NULL)
+                         break;
+
+                     t->next = hash[d];
+                     t->left = (png_byte)i;
+                     t->right = (png_byte)j;
+                     hash[d] = t;
+                  }
+               }
+               if (t == NULL)
+                  break;
+            }
+
+            if (t != NULL)
+            for (i = 0; i <= max_d; i++)
+            {
+               if (hash[i] != NULL)
+               {
+                  png_dsortp p;
+
+                  for (p = hash[i]; p; p = p->next)
+                  {
+                     if ((int)png_ptr->index_to_palette[p->left]
+                         < num_new_palette &&
+                         (int)png_ptr->index_to_palette[p->right]
+                         < num_new_palette)
+                     {
+                        int j, next_j;
+
+                        if (num_new_palette & 0x01)
+                        {
+                           j = p->left;
+                           next_j = p->right;
+                        }
+                        else
+                        {
+                           j = p->right;
+                           next_j = p->left;
+                        }
+
+                        num_new_palette--;
+                        palette[png_ptr->index_to_palette[j]]
+                            = palette[num_new_palette];
+                        if (!full_quantize)
+                        {
+                           int k;
+
+                           for (k = 0; k < num_palette; k++)
+                           {
+                              if (png_ptr->quantize_index[k] ==
+                                  png_ptr->index_to_palette[j])
+                                 png_ptr->quantize_index[k] =
+                                     png_ptr->index_to_palette[next_j];
+
+                              if ((int)png_ptr->quantize_index[k] ==
+                                  num_new_palette)
+                                 png_ptr->quantize_index[k] =
+                                     png_ptr->index_to_palette[j];
+                           }
+                        }
+
+                        png_ptr->index_to_palette[png_ptr->palette_to_index
+                            [num_new_palette]] = png_ptr->index_to_palette[j];
+
+                        png_ptr->palette_to_index[png_ptr->index_to_palette[j]]
+                            = png_ptr->palette_to_index[num_new_palette];
+
+                        png_ptr->index_to_palette[j] =
+                            (png_byte)num_new_palette;
+
+                        png_ptr->palette_to_index[num_new_palette] =
+                            (png_byte)j;
+                     }
+                     if (num_new_palette <= maximum_colors)
+                        break;
+                  }
+                  if (num_new_palette <= maximum_colors)
+                     break;
+               }
+            }
+
+            for (i = 0; i < 769; i++)
+            {
+               if (hash[i] != NULL)
+               {
+                  png_dsortp p = hash[i];
+                  while (p)
+                  {
+                     t = p->next;
+                     png_free(png_ptr, p);
+                     p = t;
+                  }
+               }
+               hash[i] = 0;
+            }
+            max_d += 96;
+         }
+         png_free(png_ptr, hash);
+         png_free(png_ptr, png_ptr->palette_to_index);
+         png_free(png_ptr, png_ptr->index_to_palette);
+         png_ptr->palette_to_index = NULL;
+         png_ptr->index_to_palette = NULL;
+      }
+      num_palette = maximum_colors;
+   }
+   if (png_ptr->palette == NULL)
+   {
+      png_ptr->palette = palette;
+   }
+   png_ptr->num_palette = (png_uint_16)num_palette;
+
+   if (full_quantize)
+   {
+      int i;
+      png_bytep distance;
+      int total_bits = PNG_QUANTIZE_RED_BITS + PNG_QUANTIZE_GREEN_BITS +
+          PNG_QUANTIZE_BLUE_BITS;
+      int num_red = (1 << PNG_QUANTIZE_RED_BITS);
+      int num_green = (1 << PNG_QUANTIZE_GREEN_BITS);
+      int num_blue = (1 << PNG_QUANTIZE_BLUE_BITS);
+      png_size_t num_entries = ((png_size_t)1 << total_bits);
+
+      png_ptr->palette_lookup = (png_bytep)png_calloc(png_ptr,
+          (png_uint_32)(num_entries * (sizeof (png_byte))));
+
+      distance = (png_bytep)png_malloc(png_ptr, (png_uint_32)(num_entries *
+          (sizeof (png_byte))));
+
+      memset(distance, 0xff, num_entries * (sizeof (png_byte)));
+
+      for (i = 0; i < num_palette; i++)
+      {
+         int ir, ig, ib;
+         int r = (palette[i].red >> (8 - PNG_QUANTIZE_RED_BITS));
+         int g = (palette[i].green >> (8 - PNG_QUANTIZE_GREEN_BITS));
+         int b = (palette[i].blue >> (8 - PNG_QUANTIZE_BLUE_BITS));
+
+         for (ir = 0; ir < num_red; ir++)
+         {
+            /* int dr = abs(ir - r); */
+            int dr = ((ir > r) ? ir - r : r - ir);
+            int index_r = (ir << (PNG_QUANTIZE_BLUE_BITS +
+                PNG_QUANTIZE_GREEN_BITS));
+
+            for (ig = 0; ig < num_green; ig++)
+            {
+               /* int dg = abs(ig - g); */
+               int dg = ((ig > g) ? ig - g : g - ig);
+               int dt = dr + dg;
+               int dm = ((dr > dg) ? dr : dg);
+               int index_g = index_r | (ig << PNG_QUANTIZE_BLUE_BITS);
+
+               for (ib = 0; ib < num_blue; ib++)
+               {
+                  int d_index = index_g | ib;
+                  /* int db = abs(ib - b); */
+                  int db = ((ib > b) ? ib - b : b - ib);
+                  int dmax = ((dm > db) ? dm : db);
+                  int d = dmax + dt + db;
+
+                  if (d < (int)distance[d_index])
+                  {
+                     distance[d_index] = (png_byte)d;
+                     png_ptr->palette_lookup[d_index] = (png_byte)i;
+                  }
+               }
+            }
+         }
+      }
+
+      png_free(png_ptr, distance);
+   }
+}
+#endif /* PNG_READ_QUANTIZE_SUPPORTED */
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+void PNGFAPI
+png_set_gamma_fixed(png_structrp png_ptr, png_fixed_point scrn_gamma,
+   png_fixed_point file_gamma)
+{
+   png_debug(1, "in png_set_gamma_fixed");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   /* New in libpng-1.5.4 - reserve particular negative values as flags. */
+   scrn_gamma = translate_gamma_flags(png_ptr, scrn_gamma, 1/*screen*/);
+   file_gamma = translate_gamma_flags(png_ptr, file_gamma, 0/*file*/);
+
+   /* Checking the gamma values for being >0 was added in 1.5.4 along with the
+    * premultiplied alpha support; this actually hides an undocumented feature
+    * of the previous implementation which allowed gamma processing to be
+    * disabled in background handling.  There is no evidence (so far) that this
+    * was being used; however, png_set_background itself accepted and must still
+    * accept '0' for the gamma value it takes, because it isn't always used.
+    *
+    * Since this is an API change (albeit a very minor one that removes an
+    * undocumented API feature) the following checks were only enabled in
+    * libpng-1.6.0.
+    */
+   if (file_gamma <= 0)
+      png_error(png_ptr, "invalid file gamma in png_set_gamma");
+
+   if (scrn_gamma <= 0)
+      png_error(png_ptr, "invalid screen gamma in png_set_gamma");
+
+   /* Set the gamma values unconditionally - this overrides the value in the PNG
+    * file if a gAMA chunk was present.  png_set_alpha_mode provides a
+    * different, easier, way to default the file gamma.
+    */
+   png_ptr->colorspace.gamma = file_gamma;
+   png_ptr->colorspace.flags |= PNG_COLORSPACE_HAVE_GAMMA;
+   png_ptr->screen_gamma = scrn_gamma;
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_gamma(png_structrp png_ptr, double scrn_gamma, double file_gamma)
+{
+   png_set_gamma_fixed(png_ptr, convert_gamma_value(png_ptr, scrn_gamma),
+      convert_gamma_value(png_ptr, file_gamma));
+}
+#  endif /* FLOATING_POINT_SUPPORTED */
+#endif /* READ_GAMMA */
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+/* Expand paletted images to RGB, expand grayscale images of
+ * less than 8-bit depth to 8-bit depth, and expand tRNS chunks
+ * to alpha channels.
+ */
+void PNGAPI
+png_set_expand(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_expand");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= (PNG_EXPAND | PNG_EXPAND_tRNS);
+}
+
+/* GRR 19990627:  the following three functions currently are identical
+ *  to png_set_expand().  However, it is entirely reasonable that someone
+ *  might wish to expand an indexed image to RGB but *not* expand a single,
+ *  fully transparent palette entry to a full alpha channel--perhaps instead
+ *  convert tRNS to the grayscale/RGB format (16-bit RGB value), or replace
+ *  the transparent color with a particular RGB value, or drop tRNS entirely.
+ *  IOW, a future version of the library may make the transformations flag
+ *  a bit more fine-grained, with separate bits for each of these three
+ *  functions.
+ *
+ *  More to the point, these functions make it obvious what libpng will be
+ *  doing, whereas "expand" can (and does) mean any number of things.
+ *
+ *  GRP 20060307: In libpng-1.2.9, png_set_gray_1_2_4_to_8() was modified
+ *  to expand only the sample depth but not to expand the tRNS to alpha
+ *  and its name was changed to png_set_expand_gray_1_2_4_to_8().
+ */
+
+/* Expand paletted images to RGB. */
+void PNGAPI
+png_set_palette_to_rgb(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_palette_to_rgb");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= (PNG_EXPAND | PNG_EXPAND_tRNS);
+}
+
+/* Expand grayscale images of less than 8-bit depth to 8 bits. */
+void PNGAPI
+png_set_expand_gray_1_2_4_to_8(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_expand_gray_1_2_4_to_8");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= PNG_EXPAND;
+}
+
+/* Expand tRNS chunks to alpha channels. */
+void PNGAPI
+png_set_tRNS_to_alpha(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_tRNS_to_alpha");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= (PNG_EXPAND | PNG_EXPAND_tRNS);
+}
+#endif /* defined(PNG_READ_EXPAND_SUPPORTED) */
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+/* Expand to 16-bit channels, expand the tRNS chunk too (because otherwise
+ * it may not work correctly.)
+ */
+void PNGAPI
+png_set_expand_16(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_expand_16");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   png_ptr->transformations |= (PNG_EXPAND_16 | PNG_EXPAND | PNG_EXPAND_tRNS);
+}
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+void PNGAPI
+png_set_gray_to_rgb(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_gray_to_rgb");
+
+   if (!png_rtran_ok(png_ptr, 0))
+      return;
+
+   /* Because rgb must be 8 bits or more: */
+   png_set_expand_gray_1_2_4_to_8(png_ptr);
+   png_ptr->transformations |= PNG_GRAY_TO_RGB;
+}
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+void PNGFAPI
+png_set_rgb_to_gray_fixed(png_structrp png_ptr, int error_action,
+    png_fixed_point red, png_fixed_point green)
+{
+   png_debug(1, "in png_set_rgb_to_gray");
+
+   /* Need the IHDR here because of the check on color_type below. */
+   /* TODO: fix this */
+   if (!png_rtran_ok(png_ptr, 1))
+      return;
+
+   switch(error_action)
+   {
+      case PNG_ERROR_ACTION_NONE:
+         png_ptr->transformations |= PNG_RGB_TO_GRAY;
+         break;
+
+      case PNG_ERROR_ACTION_WARN:
+         png_ptr->transformations |= PNG_RGB_TO_GRAY_WARN;
+         break;
+
+      case PNG_ERROR_ACTION_ERROR:
+         png_ptr->transformations |= PNG_RGB_TO_GRAY_ERR;
+         break;
+
+      default:
+         png_error(png_ptr, "invalid error action to rgb_to_gray");
+         break;
+   }
+
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+#ifdef PNG_READ_EXPAND_SUPPORTED
+      png_ptr->transformations |= PNG_EXPAND;
+#else
+   {
+      /* Make this an error in 1.6 because otherwise the application may assume
+       * that it just worked and get a memory overwrite.
+       */
+      png_error(png_ptr,
+        "Cannot do RGB_TO_GRAY without EXPAND_SUPPORTED");
+
+      /* png_ptr->transformations &= ~PNG_RGB_TO_GRAY; */
+   }
+#endif
+   {
+      if (red >= 0 && green >= 0 && red + green <= PNG_FP_1)
+      {
+         png_uint_16 red_int, green_int;
+
+         /* NOTE: this calculation does not round, but this behavior is retained
+          * for consistency, the inaccuracy is very small.  The code here always
+          * overwrites the coefficients, regardless of whether they have been
+          * defaulted or set already.
+          */
+         red_int = (png_uint_16)(((png_uint_32)red*32768)/100000);
+         green_int = (png_uint_16)(((png_uint_32)green*32768)/100000);
+
+         png_ptr->rgb_to_gray_red_coeff   = red_int;
+         png_ptr->rgb_to_gray_green_coeff = green_int;
+         png_ptr->rgb_to_gray_coefficients_set = 1;
+      }
+
+      else
+      {
+         if (red >= 0 && green >= 0)
+            png_app_warning(png_ptr,
+               "ignoring out of range rgb_to_gray coefficients");
+
+         /* Use the defaults, from the cHRM chunk if set, else the historical
+          * values which are close to the sRGB/HDTV/ITU-Rec 709 values.  See
+          * png_do_rgb_to_gray for more discussion of the values.  In this case
+          * the coefficients are not marked as 'set' and are not overwritten if
+          * something has already provided a default.
+          */
+         if (png_ptr->rgb_to_gray_red_coeff == 0 &&
+            png_ptr->rgb_to_gray_green_coeff == 0)
+         {
+            png_ptr->rgb_to_gray_red_coeff   = 6968;
+            png_ptr->rgb_to_gray_green_coeff = 23434;
+            /* png_ptr->rgb_to_gray_blue_coeff  = 2366; */
+         }
+      }
+   }
+}
+
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+/* Convert a RGB image to a grayscale of the same width.  This allows us,
+ * for example, to convert a 24 bpp RGB image into an 8 bpp grayscale image.
+ */
+
+void PNGAPI
+png_set_rgb_to_gray(png_structrp png_ptr, int error_action, double red,
+   double green)
+{
+   png_set_rgb_to_gray_fixed(png_ptr, error_action,
+      png_fixed(png_ptr, red, "rgb to gray red coefficient"),
+      png_fixed(png_ptr, green, "rgb to gray green coefficient"));
+}
+#endif /* FLOATING POINT */
+
+#endif /* RGB_TO_GRAY */
+
+#if defined(PNG_READ_USER_TRANSFORM_SUPPORTED) || \
+    defined(PNG_WRITE_USER_TRANSFORM_SUPPORTED)
+void PNGAPI
+png_set_read_user_transform_fn(png_structrp png_ptr, png_user_transform_ptr
+    read_user_transform_fn)
+{
+   png_debug(1, "in png_set_read_user_transform_fn");
+
+#ifdef PNG_READ_USER_TRANSFORM_SUPPORTED
+   png_ptr->transformations |= PNG_USER_TRANSFORM;
+   png_ptr->read_user_transform_fn = read_user_transform_fn;
+#endif
+}
+#endif
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* In the case of gamma transformations only do transformations on images where
+ * the [file] gamma and screen_gamma are not close reciprocals, otherwise it
+ * slows things down slightly, and also needlessly introduces small errors.
+ */
+static int /* PRIVATE */
+png_gamma_threshold(png_fixed_point screen_gamma, png_fixed_point file_gamma)
+{
+   /* PNG_GAMMA_THRESHOLD is the threshold for performing gamma
+    * correction as a difference of the overall transform from 1.0
+    *
+    * We want to compare the threshold with s*f - 1, if we get
+    * overflow here it is because of wacky gamma values so we
+    * turn on processing anyway.
+    */
+   png_fixed_point gtest;
+   return !png_muldiv(&gtest, screen_gamma, file_gamma, PNG_FP_1) ||
+       png_gamma_significant(gtest);
+}
+#endif
+
+/* Initialize everything needed for the read.  This includes modifying
+ * the palette.
+ */
+
+/*For the moment 'png_init_palette_transformations' and
+ * 'png_init_rgb_transformations' only do some flag canceling optimizations.
+ * The intent is that these two routines should have palette or rgb operations
+ * extracted from 'png_init_read_transformations'.
+ */
+static void /* PRIVATE */
+png_init_palette_transformations(png_structrp png_ptr)
+{
+   /* Called to handle the (input) palette case.  In png_do_read_transformations
+    * the first step is to expand the palette if requested, so this code must
+    * take care to only make changes that are invariant with respect to the
+    * palette expansion, or only do them if there is no expansion.
+    *
+    * STRIP_ALPHA has already been handled in the caller (by setting num_trans
+    * to 0.)
+    */
+   int input_has_alpha = 0;
+   int input_has_transparency = 0;
+
+   if (png_ptr->num_trans > 0)
+   {
+      int i;
+
+      /* Ignore if all the entries are opaque (unlikely!) */
+      for (i=0; i<png_ptr->num_trans; ++i)
+      {
+         if (png_ptr->trans_alpha[i] == 255)
+            continue;
+         else if (png_ptr->trans_alpha[i] == 0)
+            input_has_transparency = 1;
+         else
+         {
+            input_has_transparency = 1;
+            input_has_alpha = 1;
+            break;
+         }
+      }
+   }
+
+   /* If no alpha we can optimize. */
+   if (!input_has_alpha)
+   {
+      /* Any alpha means background and associative alpha processing is
+       * required, however if the alpha is 0 or 1 throughout OPTIIMIZE_ALPHA
+       * and ENCODE_ALPHA are irrelevant.
+       */
+      png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+      png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+
+      if (!input_has_transparency)
+         png_ptr->transformations &= ~(PNG_COMPOSE | PNG_BACKGROUND_EXPAND);
+   }
+
+#if defined(PNG_READ_EXPAND_SUPPORTED) && defined(PNG_READ_BACKGROUND_SUPPORTED)
+   /* png_set_background handling - deals with the complexity of whether the
+    * background color is in the file format or the screen format in the case
+    * where an 'expand' will happen.
+    */
+
+   /* The following code cannot be entered in the alpha pre-multiplication case
+    * because PNG_BACKGROUND_EXPAND is cancelled below.
+    */
+   if ((png_ptr->transformations & PNG_BACKGROUND_EXPAND) &&
+       (png_ptr->transformations & PNG_EXPAND))
+   {
+      {
+         png_ptr->background.red   =
+             png_ptr->palette[png_ptr->background.index].red;
+         png_ptr->background.green =
+             png_ptr->palette[png_ptr->background.index].green;
+         png_ptr->background.blue  =
+             png_ptr->palette[png_ptr->background.index].blue;
+
+#ifdef PNG_READ_INVERT_ALPHA_SUPPORTED
+        if (png_ptr->transformations & PNG_INVERT_ALPHA)
+        {
+           if (!(png_ptr->transformations & PNG_EXPAND_tRNS))
+           {
+              /* Invert the alpha channel (in tRNS) unless the pixels are
+               * going to be expanded, in which case leave it for later
+               */
+              int i, istop = png_ptr->num_trans;
+
+              for (i=0; i<istop; i++)
+                 png_ptr->trans_alpha[i] = (png_byte)(255 -
+                    png_ptr->trans_alpha[i]);
+           }
+        }
+#endif /* PNG_READ_INVERT_ALPHA_SUPPORTED */
+      }
+   } /* background expand and (therefore) no alpha association. */
+#endif /* PNG_READ_EXPAND_SUPPORTED && PNG_READ_BACKGROUND_SUPPORTED */
+}
+
+static void /* PRIVATE */
+png_init_rgb_transformations(png_structrp png_ptr)
+{
+   /* Added to libpng-1.5.4: check the color type to determine whether there
+    * is any alpha or transparency in the image and simply cancel the
+    * background and alpha mode stuff if there isn't.
+    */
+   int input_has_alpha = (png_ptr->color_type & PNG_COLOR_MASK_ALPHA) != 0;
+   int input_has_transparency = png_ptr->num_trans > 0;
+
+   /* If no alpha we can optimize. */
+   if (!input_has_alpha)
+   {
+      /* Any alpha means background and associative alpha processing is
+       * required, however if the alpha is 0 or 1 throughout OPTIIMIZE_ALPHA
+       * and ENCODE_ALPHA are irrelevant.
+       */
+#     ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+         png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+         png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+#     endif
+
+      if (!input_has_transparency)
+         png_ptr->transformations &= ~(PNG_COMPOSE | PNG_BACKGROUND_EXPAND);
+   }
+
+#if defined(PNG_READ_EXPAND_SUPPORTED) && defined(PNG_READ_BACKGROUND_SUPPORTED)
+   /* png_set_background handling - deals with the complexity of whether the
+    * background color is in the file format or the screen format in the case
+    * where an 'expand' will happen.
+    */
+
+   /* The following code cannot be entered in the alpha pre-multiplication case
+    * because PNG_BACKGROUND_EXPAND is cancelled below.
+    */
+   if ((png_ptr->transformations & PNG_BACKGROUND_EXPAND) &&
+       (png_ptr->transformations & PNG_EXPAND) &&
+       !(png_ptr->color_type & PNG_COLOR_MASK_COLOR))
+       /* i.e., GRAY or GRAY_ALPHA */
+   {
+      {
+         /* Expand background and tRNS chunks */
+         int gray = png_ptr->background.gray;
+         int trans_gray = png_ptr->trans_color.gray;
+
+         switch (png_ptr->bit_depth)
+         {
+            case 1:
+               gray *= 0xff;
+               trans_gray *= 0xff;
+               break;
+
+            case 2:
+               gray *= 0x55;
+               trans_gray *= 0x55;
+               break;
+
+            case 4:
+               gray *= 0x11;
+               trans_gray *= 0x11;
+               break;
+
+            default:
+
+            case 8:
+               /* FALL THROUGH (Already 8 bits) */
+
+            case 16:
+               /* Already a full 16 bits */
+               break;
+         }
+
+         png_ptr->background.red = png_ptr->background.green =
+            png_ptr->background.blue = (png_uint_16)gray;
+
+         if (!(png_ptr->transformations & PNG_EXPAND_tRNS))
+         {
+            png_ptr->trans_color.red = png_ptr->trans_color.green =
+               png_ptr->trans_color.blue = (png_uint_16)trans_gray;
+         }
+      }
+   } /* background expand and (therefore) no alpha association. */
+#endif /* PNG_READ_EXPAND_SUPPORTED && PNG_READ_BACKGROUND_SUPPORTED */
+}
+
+void /* PRIVATE */
+png_init_read_transformations(png_structrp png_ptr)
+{
+   png_debug(1, "in png_init_read_transformations");
+
+   /* This internal function is called from png_read_start_row in pngrutil.c
+    * and it is called before the 'rowbytes' calculation is done, so the code
+    * in here can change or update the transformations flags.
+    *
+    * First do updates that do not depend on the details of the PNG image data
+    * being processed.
+    */
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   /* Prior to 1.5.4 these tests were performed from png_set_gamma, 1.5.4 adds
+    * png_set_alpha_mode and this is another source for a default file gamma so
+    * the test needs to be performed later - here.  In addition prior to 1.5.4
+    * the tests were repeated for the PALETTE color type here - this is no
+    * longer necessary (and doesn't seem to have been necessary before.)
+    */
+   {
+      /* The following temporary indicates if overall gamma correction is
+       * required.
+       */
+      int gamma_correction = 0;
+
+      if (png_ptr->colorspace.gamma != 0) /* has been set */
+      {
+         if (png_ptr->screen_gamma != 0) /* screen set too */
+            gamma_correction = png_gamma_threshold(png_ptr->colorspace.gamma,
+               png_ptr->screen_gamma);
+
+         else
+            /* Assume the output matches the input; a long time default behavior
+             * of libpng, although the standard has nothing to say about this.
+             */
+            png_ptr->screen_gamma = png_reciprocal(png_ptr->colorspace.gamma);
+      }
+
+      else if (png_ptr->screen_gamma != 0)
+         /* The converse - assume the file matches the screen, note that this
+          * perhaps undesireable default can (from 1.5.4) be changed by calling
+          * png_set_alpha_mode (even if the alpha handling mode isn't required
+          * or isn't changed from the default.)
+          */
+         png_ptr->colorspace.gamma = png_reciprocal(png_ptr->screen_gamma);
+
+      else /* neither are set */
+         /* Just in case the following prevents any processing - file and screen
+          * are both assumed to be linear and there is no way to introduce a
+          * third gamma value other than png_set_background with 'UNIQUE', and,
+          * prior to 1.5.4
+          */
+         png_ptr->screen_gamma = png_ptr->colorspace.gamma = PNG_FP_1;
+
+      /* We have a gamma value now. */
+      png_ptr->colorspace.flags |= PNG_COLORSPACE_HAVE_GAMMA;
+
+      /* Now turn the gamma transformation on or off as appropriate.  Notice
+       * that PNG_GAMMA just refers to the file->screen correction.  Alpha
+       * composition may independently cause gamma correction because it needs
+       * linear data (e.g. if the file has a gAMA chunk but the screen gamma
+       * hasn't been specified.)  In any case this flag may get turned off in
+       * the code immediately below if the transform can be handled outside the
+       * row loop.
+       */
+      if (gamma_correction)
+         png_ptr->transformations |= PNG_GAMMA;
+
+      else
+         png_ptr->transformations &= ~PNG_GAMMA;
+   }
+#endif
+
+   /* Certain transformations have the effect of preventing other
+    * transformations that happen afterward in png_do_read_transformations,
+    * resolve the interdependencies here.  From the code of
+    * png_do_read_transformations the order is:
+    *
+    *  1) PNG_EXPAND (including PNG_EXPAND_tRNS)
+    *  2) PNG_STRIP_ALPHA (if no compose)
+    *  3) PNG_RGB_TO_GRAY
+    *  4) PNG_GRAY_TO_RGB iff !PNG_BACKGROUND_IS_GRAY
+    *  5) PNG_COMPOSE
+    *  6) PNG_GAMMA
+    *  7) PNG_STRIP_ALPHA (if compose)
+    *  8) PNG_ENCODE_ALPHA
+    *  9) PNG_SCALE_16_TO_8
+    * 10) PNG_16_TO_8
+    * 11) PNG_QUANTIZE (converts to palette)
+    * 12) PNG_EXPAND_16
+    * 13) PNG_GRAY_TO_RGB iff PNG_BACKGROUND_IS_GRAY
+    * 14) PNG_INVERT_MONO
+    * 15) PNG_SHIFT
+    * 16) PNG_PACK
+    * 17) PNG_BGR
+    * 18) PNG_PACKSWAP
+    * 19) PNG_FILLER (includes PNG_ADD_ALPHA)
+    * 20) PNG_INVERT_ALPHA
+    * 21) PNG_SWAP_ALPHA
+    * 22) PNG_SWAP_BYTES
+    * 23) PNG_USER_TRANSFORM [must be last]
+    */
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+   if ((png_ptr->transformations & PNG_STRIP_ALPHA) &&
+      !(png_ptr->transformations & PNG_COMPOSE))
+   {
+      /* Stripping the alpha channel happens immediately after the 'expand'
+       * transformations, before all other transformation, so it cancels out
+       * the alpha handling.  It has the side effect negating the effect of
+       * PNG_EXPAND_tRNS too:
+       */
+      png_ptr->transformations &= ~(PNG_BACKGROUND_EXPAND | PNG_ENCODE_ALPHA |
+         PNG_EXPAND_tRNS);
+      png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+
+      /* Kill the tRNS chunk itself too.  Prior to 1.5.4 this did not happen
+       * so transparency information would remain just so long as it wasn't
+       * expanded.  This produces unexpected API changes if the set of things
+       * that do PNG_EXPAND_tRNS changes (perfectly possible given the
+       * documentation - which says ask for what you want, accept what you
+       * get.)  This makes the behavior consistent from 1.5.4:
+       */
+      png_ptr->num_trans = 0;
+   }
+#endif /* STRIP_ALPHA supported, no COMPOSE */
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+   /* If the screen gamma is about 1.0 then the OPTIMIZE_ALPHA and ENCODE_ALPHA
+    * settings will have no effect.
+    */
+   if (!png_gamma_significant(png_ptr->screen_gamma))
+   {
+      png_ptr->transformations &= ~PNG_ENCODE_ALPHA;
+      png_ptr->flags &= ~PNG_FLAG_OPTIMIZE_ALPHA;
+   }
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+   /* Make sure the coefficients for the rgb to gray conversion are set
+    * appropriately.
+    */
+   if (png_ptr->transformations & PNG_RGB_TO_GRAY)
+      png_colorspace_set_rgb_coefficients(png_ptr);
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+#if defined(PNG_READ_EXPAND_SUPPORTED) && defined(PNG_READ_BACKGROUND_SUPPORTED)
+   /* Detect gray background and attempt to enable optimization for
+    * gray --> RGB case.
+    *
+    * Note:  if PNG_BACKGROUND_EXPAND is set and color_type is either RGB or
+    * RGB_ALPHA (in which case need_expand is superfluous anyway), the
+    * background color might actually be gray yet not be flagged as such.
+    * This is not a problem for the current code, which uses
+    * PNG_BACKGROUND_IS_GRAY only to decide when to do the
+    * png_do_gray_to_rgb() transformation.
+    *
+    * TODO: this code needs to be revised to avoid the complexity and
+    * interdependencies.  The color type of the background should be recorded in
+    * png_set_background, along with the bit depth, then the code has a record
+    * of exactly what color space the background is currently in.
+    */
+   if (png_ptr->transformations & PNG_BACKGROUND_EXPAND)
+   {
+      /* PNG_BACKGROUND_EXPAND: the background is in the file color space, so if
+       * the file was grayscale the background value is gray.
+       */
+      if (!(png_ptr->color_type & PNG_COLOR_MASK_COLOR))
+         png_ptr->mode |= PNG_BACKGROUND_IS_GRAY;
+   }
+
+   else if (png_ptr->transformations & PNG_COMPOSE)
+   {
+      /* PNG_COMPOSE: png_set_background was called with need_expand false,
+       * so the color is in the color space of the output or png_set_alpha_mode
+       * was called and the color is black.  Ignore RGB_TO_GRAY because that
+       * happens before GRAY_TO_RGB.
+       */
+      if (png_ptr->transformations & PNG_GRAY_TO_RGB)
+      {
+         if (png_ptr->background.red == png_ptr->background.green &&
+             png_ptr->background.red == png_ptr->background.blue)
+         {
+            png_ptr->mode |= PNG_BACKGROUND_IS_GRAY;
+            png_ptr->background.gray = png_ptr->background.red;
+         }
+      }
+   }
+#endif /* PNG_READ_EXPAND_SUPPORTED && PNG_READ_BACKGROUND_SUPPORTED */
+#endif /* PNG_READ_GRAY_TO_RGB_SUPPORTED */
+
+   /* For indexed PNG data (PNG_COLOR_TYPE_PALETTE) many of the transformations
+    * can be performed directly on the palette, and some (such as rgb to gray)
+    * can be optimized inside the palette.  This is particularly true of the
+    * composite (background and alpha) stuff, which can be pretty much all done
+    * in the palette even if the result is expanded to RGB or gray afterward.
+    *
+    * NOTE: this is Not Yet Implemented, the code behaves as in 1.5.1 and
+    * earlier and the palette stuff is actually handled on the first row.  This
+    * leads to the reported bug that the palette returned by png_get_PLTE is not
+    * updated.
+    */
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      png_init_palette_transformations(png_ptr);
+
+   else
+      png_init_rgb_transformations(png_ptr);
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) && \
+   defined(PNG_READ_EXPAND_16_SUPPORTED)
+   if ((png_ptr->transformations & PNG_EXPAND_16) &&
+      (png_ptr->transformations & PNG_COMPOSE) &&
+      !(png_ptr->transformations & PNG_BACKGROUND_EXPAND) &&
+      png_ptr->bit_depth != 16)
+   {
+      /* TODO: fix this.  Because the expand_16 operation is after the compose
+       * handling the background color must be 8, not 16, bits deep, but the
+       * application will supply a 16-bit value so reduce it here.
+       *
+       * The PNG_BACKGROUND_EXPAND code above does not expand to 16 bits at
+       * present, so that case is ok (until do_expand_16 is moved.)
+       *
+       * NOTE: this discards the low 16 bits of the user supplied background
+       * color, but until expand_16 works properly there is no choice!
+       */
+#     define CHOP(x) (x)=((png_uint_16)PNG_DIV257(x))
+      CHOP(png_ptr->background.red);
+      CHOP(png_ptr->background.green);
+      CHOP(png_ptr->background.blue);
+      CHOP(png_ptr->background.gray);
+#     undef CHOP
+   }
+#endif /* PNG_READ_BACKGROUND_SUPPORTED && PNG_READ_EXPAND_16_SUPPORTED */
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) && \
+   (defined(PNG_READ_SCALE_16_TO_8_SUPPORTED) || \
+   defined(PNG_READ_STRIP_16_TO_8_SUPPORTED))
+   if ((png_ptr->transformations & (PNG_16_TO_8|PNG_SCALE_16_TO_8)) &&
+      (png_ptr->transformations & PNG_COMPOSE) &&
+      !(png_ptr->transformations & PNG_BACKGROUND_EXPAND) &&
+      png_ptr->bit_depth == 16)
+   {
+      /* On the other hand, if a 16-bit file is to be reduced to 8-bits per
+       * component this will also happen after PNG_COMPOSE and so the background
+       * color must be pre-expanded here.
+       *
+       * TODO: fix this too.
+       */
+      png_ptr->background.red = (png_uint_16)(png_ptr->background.red * 257);
+      png_ptr->background.green =
+         (png_uint_16)(png_ptr->background.green * 257);
+      png_ptr->background.blue = (png_uint_16)(png_ptr->background.blue * 257);
+      png_ptr->background.gray = (png_uint_16)(png_ptr->background.gray * 257);
+   }
+#endif
+
+   /* NOTE: below 'PNG_READ_ALPHA_MODE_SUPPORTED' is presumed to also enable the
+    * background support (see the comments in scripts/pnglibconf.dfa), this
+    * allows pre-multiplication of the alpha channel to be implemented as
+    * compositing on black.  This is probably sub-optimal and has been done in
+    * 1.5.4 betas simply to enable external critique and testing (i.e. to
+    * implement the new API quickly, without lots of internal changes.)
+    */
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+#  ifdef PNG_READ_BACKGROUND_SUPPORTED
+      /* Includes ALPHA_MODE */
+      png_ptr->background_1 = png_ptr->background;
+#  endif
+
+   /* This needs to change - in the palette image case a whole set of tables are
+    * built when it would be quicker to just calculate the correct value for
+    * each palette entry directly.  Also, the test is too tricky - why check
+    * PNG_RGB_TO_GRAY if PNG_GAMMA is not set?  The answer seems to be that
+    * PNG_GAMMA is cancelled even if the gamma is known?  The test excludes the
+    * PNG_COMPOSE case, so apparently if there is no *overall* gamma correction
+    * the gamma tables will not be built even if composition is required on a
+    * gamma encoded value.
+    *
+    * In 1.5.4 this is addressed below by an additional check on the individual
+    * file gamma - if it is not 1.0 both RGB_TO_GRAY and COMPOSE need the
+    * tables.
+    */
+   if ((png_ptr->transformations & PNG_GAMMA)
+      || ((png_ptr->transformations & PNG_RGB_TO_GRAY)
+         && (png_gamma_significant(png_ptr->colorspace.gamma) ||
+            png_gamma_significant(png_ptr->screen_gamma)))
+      || ((png_ptr->transformations & PNG_COMPOSE)
+         && (png_gamma_significant(png_ptr->colorspace.gamma)
+            || png_gamma_significant(png_ptr->screen_gamma)
+#  ifdef PNG_READ_BACKGROUND_SUPPORTED
+            || (png_ptr->background_gamma_type == PNG_BACKGROUND_GAMMA_UNIQUE
+               && png_gamma_significant(png_ptr->background_gamma))
+#  endif
+      )) || ((png_ptr->transformations & PNG_ENCODE_ALPHA)
+         && png_gamma_significant(png_ptr->screen_gamma))
+      )
+   {
+      png_build_gamma_table(png_ptr, png_ptr->bit_depth);
+
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+      if (png_ptr->transformations & PNG_COMPOSE)
+      {
+         /* Issue a warning about this combination: because RGB_TO_GRAY is
+          * optimized to do the gamma transform if present yet do_background has
+          * to do the same thing if both options are set a
+          * double-gamma-correction happens.  This is true in all versions of
+          * libpng to date.
+          */
+         if (png_ptr->transformations & PNG_RGB_TO_GRAY)
+            png_warning(png_ptr,
+               "libpng does not support gamma+background+rgb_to_gray");
+
+         if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+         {
+            /* We don't get to here unless there is a tRNS chunk with non-opaque
+             * entries - see the checking code at the start of this function.
+             */
+            png_color back, back_1;
+            png_colorp palette = png_ptr->palette;
+            int num_palette = png_ptr->num_palette;
+            int i;
+            if (png_ptr->background_gamma_type == PNG_BACKGROUND_GAMMA_FILE)
+            {
+
+               back.red = png_ptr->gamma_table[png_ptr->background.red];
+               back.green = png_ptr->gamma_table[png_ptr->background.green];
+               back.blue = png_ptr->gamma_table[png_ptr->background.blue];
+
+               back_1.red = png_ptr->gamma_to_1[png_ptr->background.red];
+               back_1.green = png_ptr->gamma_to_1[png_ptr->background.green];
+               back_1.blue = png_ptr->gamma_to_1[png_ptr->background.blue];
+            }
+            else
+            {
+               png_fixed_point g, gs;
+
+               switch (png_ptr->background_gamma_type)
+               {
+                  case PNG_BACKGROUND_GAMMA_SCREEN:
+                     g = (png_ptr->screen_gamma);
+                     gs = PNG_FP_1;
+                     break;
+
+                  case PNG_BACKGROUND_GAMMA_FILE:
+                     g = png_reciprocal(png_ptr->colorspace.gamma);
+                     gs = png_reciprocal2(png_ptr->colorspace.gamma,
+                        png_ptr->screen_gamma);
+                     break;
+
+                  case PNG_BACKGROUND_GAMMA_UNIQUE:
+                     g = png_reciprocal(png_ptr->background_gamma);
+                     gs = png_reciprocal2(png_ptr->background_gamma,
+                        png_ptr->screen_gamma);
+                     break;
+                  default:
+                     g = PNG_FP_1;    /* back_1 */
+                     gs = PNG_FP_1;   /* back */
+                     break;
+               }
+
+               if (png_gamma_significant(gs))
+               {
+                  back.red = png_gamma_8bit_correct(png_ptr->background.red,
+                      gs);
+                  back.green = png_gamma_8bit_correct(png_ptr->background.green,
+                      gs);
+                  back.blue = png_gamma_8bit_correct(png_ptr->background.blue,
+                      gs);
+               }
+
+               else
+               {
+                  back.red   = (png_byte)png_ptr->background.red;
+                  back.green = (png_byte)png_ptr->background.green;
+                  back.blue  = (png_byte)png_ptr->background.blue;
+               }
+
+               if (png_gamma_significant(g))
+               {
+                  back_1.red = png_gamma_8bit_correct(png_ptr->background.red,
+                     g);
+                  back_1.green = png_gamma_8bit_correct(
+                     png_ptr->background.green, g);
+                  back_1.blue = png_gamma_8bit_correct(png_ptr->background.blue,
+                     g);
+               }
+
+               else
+               {
+                  back_1.red   = (png_byte)png_ptr->background.red;
+                  back_1.green = (png_byte)png_ptr->background.green;
+                  back_1.blue  = (png_byte)png_ptr->background.blue;
+               }
+            }
+
+            for (i = 0; i < num_palette; i++)
+            {
+               if (i < (int)png_ptr->num_trans &&
+                   png_ptr->trans_alpha[i] != 0xff)
+               {
+                  if (png_ptr->trans_alpha[i] == 0)
+                  {
+                     palette[i] = back;
+                  }
+                  else /* if (png_ptr->trans_alpha[i] != 0xff) */
+                  {
+                     png_byte v, w;
+
+                     v = png_ptr->gamma_to_1[palette[i].red];
+                     png_composite(w, v, png_ptr->trans_alpha[i], back_1.red);
+                     palette[i].red = png_ptr->gamma_from_1[w];
+
+                     v = png_ptr->gamma_to_1[palette[i].green];
+                     png_composite(w, v, png_ptr->trans_alpha[i], back_1.green);
+                     palette[i].green = png_ptr->gamma_from_1[w];
+
+                     v = png_ptr->gamma_to_1[palette[i].blue];
+                     png_composite(w, v, png_ptr->trans_alpha[i], back_1.blue);
+                     palette[i].blue = png_ptr->gamma_from_1[w];
+                  }
+               }
+               else
+               {
+                  palette[i].red = png_ptr->gamma_table[palette[i].red];
+                  palette[i].green = png_ptr->gamma_table[palette[i].green];
+                  palette[i].blue = png_ptr->gamma_table[palette[i].blue];
+               }
+            }
+
+            /* Prevent the transformations being done again.
+             *
+             * NOTE: this is highly dubious; it removes the transformations in
+             * place.  This seems inconsistent with the general treatment of the
+             * transformations elsewhere.
+             */
+            png_ptr->transformations &= ~(PNG_COMPOSE | PNG_GAMMA);
+         } /* color_type == PNG_COLOR_TYPE_PALETTE */
+
+         /* if (png_ptr->background_gamma_type!=PNG_BACKGROUND_GAMMA_UNKNOWN) */
+         else /* color_type != PNG_COLOR_TYPE_PALETTE */
+         {
+            int gs_sig, g_sig;
+            png_fixed_point g = PNG_FP_1;  /* Correction to linear */
+            png_fixed_point gs = PNG_FP_1; /* Correction to screen */
+
+            switch (png_ptr->background_gamma_type)
+            {
+               case PNG_BACKGROUND_GAMMA_SCREEN:
+                  g = png_ptr->screen_gamma;
+                  /* gs = PNG_FP_1; */
+                  break;
+
+               case PNG_BACKGROUND_GAMMA_FILE:
+                  g = png_reciprocal(png_ptr->colorspace.gamma);
+                  gs = png_reciprocal2(png_ptr->colorspace.gamma,
+                     png_ptr->screen_gamma);
+                  break;
+
+               case PNG_BACKGROUND_GAMMA_UNIQUE:
+                  g = png_reciprocal(png_ptr->background_gamma);
+                  gs = png_reciprocal2(png_ptr->background_gamma,
+                      png_ptr->screen_gamma);
+                  break;
+
+               default:
+                  png_error(png_ptr, "invalid background gamma type");
+            }
+
+            g_sig = png_gamma_significant(g);
+            gs_sig = png_gamma_significant(gs);
+
+            if (g_sig)
+               png_ptr->background_1.gray = png_gamma_correct(png_ptr,
+                   png_ptr->background.gray, g);
+
+            if (gs_sig)
+               png_ptr->background.gray = png_gamma_correct(png_ptr,
+                   png_ptr->background.gray, gs);
+
+            if ((png_ptr->background.red != png_ptr->background.green) ||
+                (png_ptr->background.red != png_ptr->background.blue) ||
+                (png_ptr->background.red != png_ptr->background.gray))
+            {
+               /* RGB or RGBA with color background */
+               if (g_sig)
+               {
+                  png_ptr->background_1.red = png_gamma_correct(png_ptr,
+                      png_ptr->background.red, g);
+
+                  png_ptr->background_1.green = png_gamma_correct(png_ptr,
+                      png_ptr->background.green, g);
+
+                  png_ptr->background_1.blue = png_gamma_correct(png_ptr,
+                      png_ptr->background.blue, g);
+               }
+
+               if (gs_sig)
+               {
+                  png_ptr->background.red = png_gamma_correct(png_ptr,
+                      png_ptr->background.red, gs);
+
+                  png_ptr->background.green = png_gamma_correct(png_ptr,
+                      png_ptr->background.green, gs);
+
+                  png_ptr->background.blue = png_gamma_correct(png_ptr,
+                      png_ptr->background.blue, gs);
+               }
+            }
+
+            else
+            {
+               /* GRAY, GRAY ALPHA, RGB, or RGBA with gray background */
+               png_ptr->background_1.red = png_ptr->background_1.green
+                   = png_ptr->background_1.blue = png_ptr->background_1.gray;
+
+               png_ptr->background.red = png_ptr->background.green
+                   = png_ptr->background.blue = png_ptr->background.gray;
+            }
+
+            /* The background is now in screen gamma: */
+            png_ptr->background_gamma_type = PNG_BACKGROUND_GAMMA_SCREEN;
+         } /* color_type != PNG_COLOR_TYPE_PALETTE */
+      }/* png_ptr->transformations & PNG_BACKGROUND */
+
+      else
+      /* Transformation does not include PNG_BACKGROUND */
+#endif /* PNG_READ_BACKGROUND_SUPPORTED */
+      if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+         /* RGB_TO_GRAY needs to have non-gamma-corrected values! */
+         && ((png_ptr->transformations & PNG_EXPAND) == 0 ||
+         (png_ptr->transformations & PNG_RGB_TO_GRAY) == 0)
+#endif
+         )
+      {
+         png_colorp palette = png_ptr->palette;
+         int num_palette = png_ptr->num_palette;
+         int i;
+
+         /* NOTE: there are other transformations that should probably be in
+          * here too.
+          */
+         for (i = 0; i < num_palette; i++)
+         {
+            palette[i].red = png_ptr->gamma_table[palette[i].red];
+            palette[i].green = png_ptr->gamma_table[palette[i].green];
+            palette[i].blue = png_ptr->gamma_table[palette[i].blue];
+         }
+
+         /* Done the gamma correction. */
+         png_ptr->transformations &= ~PNG_GAMMA;
+      } /* color_type == PALETTE && !PNG_BACKGROUND transformation */
+   }
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+   else
+#endif
+#endif /* PNG_READ_GAMMA_SUPPORTED */
+
+#ifdef PNG_READ_BACKGROUND_SUPPORTED
+   /* No GAMMA transformation (see the hanging else 4 lines above) */
+   if ((png_ptr->transformations & PNG_COMPOSE) &&
+       (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE))
+   {
+      int i;
+      int istop = (int)png_ptr->num_trans;
+      png_color back;
+      png_colorp palette = png_ptr->palette;
+
+      back.red   = (png_byte)png_ptr->background.red;
+      back.green = (png_byte)png_ptr->background.green;
+      back.blue  = (png_byte)png_ptr->background.blue;
+
+      for (i = 0; i < istop; i++)
+      {
+         if (png_ptr->trans_alpha[i] == 0)
+         {
+            palette[i] = back;
+         }
+
+         else if (png_ptr->trans_alpha[i] != 0xff)
+         {
+            /* The png_composite() macro is defined in png.h */
+            png_composite(palette[i].red, palette[i].red,
+                png_ptr->trans_alpha[i], back.red);
+
+            png_composite(palette[i].green, palette[i].green,
+                png_ptr->trans_alpha[i], back.green);
+
+            png_composite(palette[i].blue, palette[i].blue,
+                png_ptr->trans_alpha[i], back.blue);
+         }
+      }
+
+      png_ptr->transformations &= ~PNG_COMPOSE;
+   }
+#endif /* PNG_READ_BACKGROUND_SUPPORTED */
+
+#ifdef PNG_READ_SHIFT_SUPPORTED
+   if ((png_ptr->transformations & PNG_SHIFT) &&
+      !(png_ptr->transformations & PNG_EXPAND) &&
+       (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE))
+   {
+      int i;
+      int istop = png_ptr->num_palette;
+      int shift = 8 - png_ptr->sig_bit.red;
+
+      png_ptr->transformations &= ~PNG_SHIFT;
+
+      /* significant bits can be in the range 1 to 7 for a meaninful result, if
+       * the number of significant bits is 0 then no shift is done (this is an
+       * error condition which is silently ignored.)
+       */
+      if (shift > 0 && shift < 8)
+         for (i=0; i<istop; ++i)
+         {
+            int component = png_ptr->palette[i].red;
+
+            component >>= shift;
+            png_ptr->palette[i].red = (png_byte)component;
+         }
+
+      shift = 8 - png_ptr->sig_bit.green;
+      if (shift > 0 && shift < 8)
+         for (i=0; i<istop; ++i)
+         {
+            int component = png_ptr->palette[i].green;
+
+            component >>= shift;
+            png_ptr->palette[i].green = (png_byte)component;
+         }
+
+      shift = 8 - png_ptr->sig_bit.blue;
+      if (shift > 0 && shift < 8)
+         for (i=0; i<istop; ++i)
+         {
+            int component = png_ptr->palette[i].blue;
+
+            component >>= shift;
+            png_ptr->palette[i].blue = (png_byte)component;
+         }
+   }
+#endif  /* PNG_READ_SHIFT_SUPPORTED */
+}
+
+/* Modify the info structure to reflect the transformations.  The
+ * info should be updated so a PNG file could be written with it,
+ * assuming the transformations result in valid PNG data.
+ */
+void /* PRIVATE */
+png_read_transform_info(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_debug(1, "in png_read_transform_info");
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+   if (png_ptr->transformations & PNG_EXPAND)
+   {
+      if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         /* This check must match what actually happens in
+          * png_do_expand_palette; if it ever checks the tRNS chunk to see if
+          * it is all opaque we must do the same (at present it does not.)
+          */
+         if (png_ptr->num_trans > 0)
+            info_ptr->color_type = PNG_COLOR_TYPE_RGB_ALPHA;
+
+         else
+            info_ptr->color_type = PNG_COLOR_TYPE_RGB;
+
+         info_ptr->bit_depth = 8;
+         info_ptr->num_trans = 0;
+      }
+      else
+      {
+         if (png_ptr->num_trans)
+         {
+            if (png_ptr->transformations & PNG_EXPAND_tRNS)
+               info_ptr->color_type |= PNG_COLOR_MASK_ALPHA;
+         }
+         if (info_ptr->bit_depth < 8)
+            info_ptr->bit_depth = 8;
+
+         info_ptr->num_trans = 0;
+      }
+   }
+#endif
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+   /* The following is almost certainly wrong unless the background value is in
+    * the screen space!
+    */
+   if (png_ptr->transformations & PNG_COMPOSE)
+      info_ptr->background = png_ptr->background;
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   /* The following used to be conditional on PNG_GAMMA (prior to 1.5.4),
+    * however it seems that the code in png_init_read_transformations, which has
+    * been called before this from png_read_update_info->png_read_start_row
+    * sometimes does the gamma transform and cancels the flag.
+    *
+    * TODO: this looks wrong; the info_ptr should end up with a gamma equal to
+    * the screen_gamma value.  The following probably results in weirdness if
+    * the info_ptr is used by the app after the rows have been read.
+    */
+   info_ptr->colorspace.gamma = png_ptr->colorspace.gamma;
+#endif
+
+   if (info_ptr->bit_depth == 16)
+   {
+#  ifdef PNG_READ_16BIT_SUPPORTED
+#     ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+         if (png_ptr->transformations & PNG_SCALE_16_TO_8)
+            info_ptr->bit_depth = 8;
+#     endif
+
+#     ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+         if (png_ptr->transformations & PNG_16_TO_8)
+            info_ptr->bit_depth = 8;
+#     endif
+
+#  else
+      /* No 16 bit support: force chopping 16-bit input down to 8, in this case
+       * the app program can chose if both APIs are available by setting the
+       * correct scaling to use.
+       */
+#     ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+         /* For compatibility with previous versions use the strip method by
+          * default.  This code works because if PNG_SCALE_16_TO_8 is already
+          * set the code below will do that in preference to the chop.
+          */
+         png_ptr->transformations |= PNG_16_TO_8;
+         info_ptr->bit_depth = 8;
+#     else
+
+#        ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+            png_ptr->transformations |= PNG_SCALE_16_TO_8;
+            info_ptr->bit_depth = 8;
+#        else
+
+            CONFIGURATION ERROR: you must enable at least one 16 to 8 method
+#        endif
+#    endif
+#endif /* !READ_16BIT_SUPPORTED */
+   }
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+   if (png_ptr->transformations & PNG_GRAY_TO_RGB)
+      info_ptr->color_type = (png_byte)(info_ptr->color_type |
+         PNG_COLOR_MASK_COLOR);
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+   if (png_ptr->transformations & PNG_RGB_TO_GRAY)
+      info_ptr->color_type = (png_byte)(info_ptr->color_type &
+         ~PNG_COLOR_MASK_COLOR);
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+   if (png_ptr->transformations & PNG_QUANTIZE)
+   {
+      if (((info_ptr->color_type == PNG_COLOR_TYPE_RGB) ||
+          (info_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA)) &&
+          png_ptr->palette_lookup && info_ptr->bit_depth == 8)
+      {
+         info_ptr->color_type = PNG_COLOR_TYPE_PALETTE;
+      }
+   }
+#endif
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+   if (png_ptr->transformations & PNG_EXPAND_16 && info_ptr->bit_depth == 8 &&
+      info_ptr->color_type != PNG_COLOR_TYPE_PALETTE)
+   {
+      info_ptr->bit_depth = 16;
+   }
+#endif
+
+#ifdef PNG_READ_PACK_SUPPORTED
+   if ((png_ptr->transformations & PNG_PACK) && (info_ptr->bit_depth < 8))
+      info_ptr->bit_depth = 8;
+#endif
+
+   if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      info_ptr->channels = 1;
+
+   else if (info_ptr->color_type & PNG_COLOR_MASK_COLOR)
+      info_ptr->channels = 3;
+
+   else
+      info_ptr->channels = 1;
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+   if (png_ptr->transformations & PNG_STRIP_ALPHA)
+   {
+      info_ptr->color_type = (png_byte)(info_ptr->color_type &
+         ~PNG_COLOR_MASK_ALPHA);
+      info_ptr->num_trans = 0;
+   }
+#endif
+
+   if (info_ptr->color_type & PNG_COLOR_MASK_ALPHA)
+      info_ptr->channels++;
+
+#ifdef PNG_READ_FILLER_SUPPORTED
+   /* STRIP_ALPHA and FILLER allowed:  MASK_ALPHA bit stripped above */
+   if ((png_ptr->transformations & PNG_FILLER) &&
+       ((info_ptr->color_type == PNG_COLOR_TYPE_RGB) ||
+       (info_ptr->color_type == PNG_COLOR_TYPE_GRAY)))
+   {
+      info_ptr->channels++;
+      /* If adding a true alpha channel not just filler */
+      if (png_ptr->transformations & PNG_ADD_ALPHA)
+         info_ptr->color_type |= PNG_COLOR_MASK_ALPHA;
+   }
+#endif
+
+#if defined(PNG_USER_TRANSFORM_PTR_SUPPORTED) && \
+defined(PNG_READ_USER_TRANSFORM_SUPPORTED)
+   if (png_ptr->transformations & PNG_USER_TRANSFORM)
+   {
+      if (info_ptr->bit_depth < png_ptr->user_transform_depth)
+         info_ptr->bit_depth = png_ptr->user_transform_depth;
+
+      if (info_ptr->channels < png_ptr->user_transform_channels)
+         info_ptr->channels = png_ptr->user_transform_channels;
+   }
+#endif
+
+   info_ptr->pixel_depth = (png_byte)(info_ptr->channels *
+       info_ptr->bit_depth);
+
+   info_ptr->rowbytes = PNG_ROWBYTES(info_ptr->pixel_depth, info_ptr->width);
+
+   /* Adding in 1.5.4: cache the above value in png_struct so that we can later
+    * check in png_rowbytes that the user buffer won't get overwritten.  Note
+    * that the field is not always set - if png_read_update_info isn't called
+    * the application has to either not do any transforms or get the calculation
+    * right itself.
+    */
+   png_ptr->info_rowbytes = info_ptr->rowbytes;
+
+#ifndef PNG_READ_EXPAND_SUPPORTED
+   if (png_ptr)
+      return;
+#endif
+}
+
+/* Transform the row.  The order of transformations is significant,
+ * and is very touchy.  If you add a transformation, take care to
+ * decide how it fits in with the other transformations here.
+ */
+void /* PRIVATE */
+png_do_read_transformations(png_structrp png_ptr, png_row_infop row_info)
+{
+   png_debug(1, "in png_do_read_transformations");
+
+   if (png_ptr->row_buf == NULL)
+   {
+      /* Prior to 1.5.4 this output row/pass where the NULL pointer is, but this
+       * error is incredibly rare and incredibly easy to debug without this
+       * information.
+       */
+      png_error(png_ptr, "NULL row buffer");
+   }
+
+   /* The following is debugging; prior to 1.5.4 the code was never compiled in;
+    * in 1.5.4 PNG_FLAG_DETECT_UNINITIALIZED was added and the macro
+    * PNG_WARN_UNINITIALIZED_ROW removed.  In 1.6 the new flag is set only for
+    * all transformations, however in practice the ROW_INIT always gets done on
+    * demand, if necessary.
+    */
+   if ((png_ptr->flags & PNG_FLAG_DETECT_UNINITIALIZED) != 0 &&
+      !(png_ptr->flags & PNG_FLAG_ROW_INIT))
+   {
+      /* Application has failed to call either png_read_start_image() or
+       * png_read_update_info() after setting transforms that expand pixels.
+       * This check added to libpng-1.2.19 (but not enabled until 1.5.4).
+       */
+      png_error(png_ptr, "Uninitialized row");
+   }
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+   if (png_ptr->transformations & PNG_EXPAND)
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         png_do_expand_palette(row_info, png_ptr->row_buf + 1,
+             png_ptr->palette, png_ptr->trans_alpha, png_ptr->num_trans);
+      }
+
+      else
+      {
+         if (png_ptr->num_trans &&
+             (png_ptr->transformations & PNG_EXPAND_tRNS))
+            png_do_expand(row_info, png_ptr->row_buf + 1,
+                &(png_ptr->trans_color));
+
+         else
+            png_do_expand(row_info, png_ptr->row_buf + 1,
+                NULL);
+      }
+   }
+#endif
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+   if ((png_ptr->transformations & PNG_STRIP_ALPHA) &&
+      !(png_ptr->transformations & PNG_COMPOSE) &&
+      (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+      row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA))
+      png_do_strip_channel(row_info, png_ptr->row_buf + 1,
+         0 /* at_start == false, because SWAP_ALPHA happens later */);
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+   if (png_ptr->transformations & PNG_RGB_TO_GRAY)
+   {
+      int rgb_error =
+          png_do_rgb_to_gray(png_ptr, row_info,
+              png_ptr->row_buf + 1);
+
+      if (rgb_error)
+      {
+         png_ptr->rgb_to_gray_status=1;
+         if ((png_ptr->transformations & PNG_RGB_TO_GRAY) ==
+             PNG_RGB_TO_GRAY_WARN)
+            png_warning(png_ptr, "png_do_rgb_to_gray found nongray pixel");
+
+         if ((png_ptr->transformations & PNG_RGB_TO_GRAY) ==
+             PNG_RGB_TO_GRAY_ERR)
+            png_error(png_ptr, "png_do_rgb_to_gray found nongray pixel");
+      }
+   }
+#endif
+
+/* From Andreas Dilger e-mail to png-implement, 26 March 1998:
+ *
+ *   In most cases, the "simple transparency" should be done prior to doing
+ *   gray-to-RGB, or you will have to test 3x as many bytes to check if a
+ *   pixel is transparent.  You would also need to make sure that the
+ *   transparency information is upgraded to RGB.
+ *
+ *   To summarize, the current flow is:
+ *   - Gray + simple transparency -> compare 1 or 2 gray bytes and composite
+ *                                   with background "in place" if transparent,
+ *                                   convert to RGB if necessary
+ *   - Gray + alpha -> composite with gray background and remove alpha bytes,
+ *                                   convert to RGB if necessary
+ *
+ *   To support RGB backgrounds for gray images we need:
+ *   - Gray + simple transparency -> convert to RGB + simple transparency,
+ *                                   compare 3 or 6 bytes and composite with
+ *                                   background "in place" if transparent
+ *                                   (3x compare/pixel compared to doing
+ *                                   composite with gray bkgrnd)
+ *   - Gray + alpha -> convert to RGB + alpha, composite with background and
+ *                                   remove alpha bytes (3x float
+ *                                   operations/pixel compared with composite
+ *                                   on gray background)
+ *
+ *  Greg's change will do this.  The reason it wasn't done before is for
+ *  performance, as this increases the per-pixel operations.  If we would check
+ *  in advance if the background was gray or RGB, and position the gray-to-RGB
+ *  transform appropriately, then it would save a lot of work/time.
+ */
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+   /* If gray -> RGB, do so now only if background is non-gray; else do later
+    * for performance reasons
+    */
+   if ((png_ptr->transformations & PNG_GRAY_TO_RGB) &&
+       !(png_ptr->mode & PNG_BACKGROUND_IS_GRAY))
+      png_do_gray_to_rgb(row_info, png_ptr->row_buf + 1);
+#endif
+
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+   if (png_ptr->transformations & PNG_COMPOSE)
+      png_do_compose(row_info, png_ptr->row_buf + 1, png_ptr);
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   if ((png_ptr->transformations & PNG_GAMMA) &&
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+      /* Because RGB_TO_GRAY does the gamma transform. */
+      !(png_ptr->transformations & PNG_RGB_TO_GRAY) &&
+#endif
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+      /* Because PNG_COMPOSE does the gamma transform if there is something to
+       * do (if there is an alpha channel or transparency.)
+       */
+       !((png_ptr->transformations & PNG_COMPOSE) &&
+       ((png_ptr->num_trans != 0) ||
+       (png_ptr->color_type & PNG_COLOR_MASK_ALPHA))) &&
+#endif
+      /* Because png_init_read_transformations transforms the palette, unless
+       * RGB_TO_GRAY will do the transform.
+       */
+       (png_ptr->color_type != PNG_COLOR_TYPE_PALETTE))
+      png_do_gamma(row_info, png_ptr->row_buf + 1, png_ptr);
+#endif
+
+#ifdef PNG_READ_STRIP_ALPHA_SUPPORTED
+   if ((png_ptr->transformations & PNG_STRIP_ALPHA) &&
+      (png_ptr->transformations & PNG_COMPOSE) &&
+      (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA ||
+      row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA))
+      png_do_strip_channel(row_info, png_ptr->row_buf + 1,
+         0 /* at_start == false, because SWAP_ALPHA happens later */);
+#endif
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+   if ((png_ptr->transformations & PNG_ENCODE_ALPHA) &&
+      (row_info->color_type & PNG_COLOR_MASK_ALPHA))
+      png_do_encode_alpha(row_info, png_ptr->row_buf + 1, png_ptr);
+#endif
+
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+   if (png_ptr->transformations & PNG_SCALE_16_TO_8)
+      png_do_scale_16_to_8(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+   /* There is no harm in doing both of these because only one has any effect,
+    * by putting the 'scale' option first if the app asks for scale (either by
+    * calling the API or in a TRANSFORM flag) this is what happens.
+    */
+   if (png_ptr->transformations & PNG_16_TO_8)
+      png_do_chop(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+   if (png_ptr->transformations & PNG_QUANTIZE)
+   {
+      png_do_quantize(row_info, png_ptr->row_buf + 1,
+          png_ptr->palette_lookup, png_ptr->quantize_index);
+
+      if (row_info->rowbytes == 0)
+         png_error(png_ptr, "png_do_quantize returned rowbytes=0");
+   }
+#endif /* PNG_READ_QUANTIZE_SUPPORTED */
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+   /* Do the expansion now, after all the arithmetic has been done.  Notice
+    * that previous transformations can handle the PNG_EXPAND_16 flag if this
+    * is efficient (particularly true in the case of gamma correction, where
+    * better accuracy results faster!)
+    */
+   if (png_ptr->transformations & PNG_EXPAND_16)
+      png_do_expand_16(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+   /* NOTE: moved here in 1.5.4 (from much later in this list.) */
+   if ((png_ptr->transformations & PNG_GRAY_TO_RGB) &&
+       (png_ptr->mode & PNG_BACKGROUND_IS_GRAY))
+      png_do_gray_to_rgb(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_INVERT_SUPPORTED
+   if (png_ptr->transformations & PNG_INVERT_MONO)
+      png_do_invert(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_SHIFT_SUPPORTED
+   if (png_ptr->transformations & PNG_SHIFT)
+      png_do_unshift(row_info, png_ptr->row_buf + 1,
+          &(png_ptr->shift));
+#endif
+
+#ifdef PNG_READ_PACK_SUPPORTED
+   if (png_ptr->transformations & PNG_PACK)
+      png_do_unpack(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   /* Added at libpng-1.5.10 */
+   if (row_info->color_type == PNG_COLOR_TYPE_PALETTE &&
+       png_ptr->num_palette_max >= 0)
+      png_do_check_palette_indexes(png_ptr, row_info);
+#endif
+
+#ifdef PNG_READ_BGR_SUPPORTED
+   if (png_ptr->transformations & PNG_BGR)
+      png_do_bgr(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_PACKSWAP_SUPPORTED
+   if (png_ptr->transformations & PNG_PACKSWAP)
+      png_do_packswap(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_FILLER_SUPPORTED
+   if (png_ptr->transformations & PNG_FILLER)
+      png_do_read_filler(row_info, png_ptr->row_buf + 1,
+          (png_uint_32)png_ptr->filler, png_ptr->flags);
+#endif
+
+#ifdef PNG_READ_INVERT_ALPHA_SUPPORTED
+   if (png_ptr->transformations & PNG_INVERT_ALPHA)
+      png_do_read_invert_alpha(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_SWAP_ALPHA_SUPPORTED
+   if (png_ptr->transformations & PNG_SWAP_ALPHA)
+      png_do_read_swap_alpha(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+#ifdef PNG_READ_SWAP_SUPPORTED
+   if (png_ptr->transformations & PNG_SWAP_BYTES)
+      png_do_swap(row_info, png_ptr->row_buf + 1);
+#endif
+#endif
+
+#ifdef PNG_READ_USER_TRANSFORM_SUPPORTED
+   if (png_ptr->transformations & PNG_USER_TRANSFORM)
+    {
+      if (png_ptr->read_user_transform_fn != NULL)
+         (*(png_ptr->read_user_transform_fn)) /* User read transform function */
+             (png_ptr,     /* png_ptr */
+             row_info,     /* row_info: */
+                /*  png_uint_32 width;       width of row */
+                /*  png_size_t rowbytes;     number of bytes in row */
+                /*  png_byte color_type;     color type of pixels */
+                /*  png_byte bit_depth;      bit depth of samples */
+                /*  png_byte channels;       number of channels (1-4) */
+                /*  png_byte pixel_depth;    bits per pixel (depth*channels) */
+             png_ptr->row_buf + 1);    /* start of pixel data for row */
+#ifdef PNG_USER_TRANSFORM_PTR_SUPPORTED
+      if (png_ptr->user_transform_depth)
+         row_info->bit_depth = png_ptr->user_transform_depth;
+
+      if (png_ptr->user_transform_channels)
+         row_info->channels = png_ptr->user_transform_channels;
+#endif
+      row_info->pixel_depth = (png_byte)(row_info->bit_depth *
+          row_info->channels);
+
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_info->width);
+   }
+#endif
+}
+
+#ifdef PNG_READ_PACK_SUPPORTED
+/* Unpack pixels of 1, 2, or 4 bits per pixel into 1 byte per pixel,
+ * without changing the actual values.  Thus, if you had a row with
+ * a bit depth of 1, you would end up with bytes that only contained
+ * the numbers 0 or 1.  If you would rather they contain 0 and 255, use
+ * png_do_shift() after this.
+ */
+void /* PRIVATE */
+png_do_unpack(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_unpack");
+
+   if (row_info->bit_depth < 8)
+   {
+      png_uint_32 i;
+      png_uint_32 row_width=row_info->width;
+
+      switch (row_info->bit_depth)
+      {
+         case 1:
+         {
+            png_bytep sp = row + (png_size_t)((row_width - 1) >> 3);
+            png_bytep dp = row + (png_size_t)row_width - 1;
+            png_uint_32 shift = 7 - (int)((row_width + 7) & 0x07);
+            for (i = 0; i < row_width; i++)
+            {
+               *dp = (png_byte)((*sp >> shift) & 0x01);
+
+               if (shift == 7)
+               {
+                  shift = 0;
+                  sp--;
+               }
+
+               else
+                  shift++;
+
+               dp--;
+            }
+            break;
+         }
+
+         case 2:
+         {
+
+            png_bytep sp = row + (png_size_t)((row_width - 1) >> 2);
+            png_bytep dp = row + (png_size_t)row_width - 1;
+            png_uint_32 shift = (int)((3 - ((row_width + 3) & 0x03)) << 1);
+            for (i = 0; i < row_width; i++)
+            {
+               *dp = (png_byte)((*sp >> shift) & 0x03);
+
+               if (shift == 6)
+               {
+                  shift = 0;
+                  sp--;
+               }
+
+               else
+                  shift += 2;
+
+               dp--;
+            }
+            break;
+         }
+
+         case 4:
+         {
+            png_bytep sp = row + (png_size_t)((row_width - 1) >> 1);
+            png_bytep dp = row + (png_size_t)row_width - 1;
+            png_uint_32 shift = (int)((1 - ((row_width + 1) & 0x01)) << 2);
+            for (i = 0; i < row_width; i++)
+            {
+               *dp = (png_byte)((*sp >> shift) & 0x0f);
+
+               if (shift == 4)
+               {
+                  shift = 0;
+                  sp--;
+               }
+
+               else
+                  shift = 4;
+
+               dp--;
+            }
+            break;
+         }
+
+         default:
+            break;
+      }
+      row_info->bit_depth = 8;
+      row_info->pixel_depth = (png_byte)(8 * row_info->channels);
+      row_info->rowbytes = row_width * row_info->channels;
+   }
+}
+#endif
+
+#ifdef PNG_READ_SHIFT_SUPPORTED
+/* Reverse the effects of png_do_shift.  This routine merely shifts the
+ * pixels back to their significant bits values.  Thus, if you have
+ * a row of bit depth 8, but only 5 are significant, this will shift
+ * the values back to 0 through 31.
+ */
+void /* PRIVATE */
+png_do_unshift(png_row_infop row_info, png_bytep row,
+    png_const_color_8p sig_bits)
+{
+   int color_type;
+
+   png_debug(1, "in png_do_unshift");
+
+   /* The palette case has already been handled in the _init routine. */
+   color_type = row_info->color_type;
+
+   if (color_type != PNG_COLOR_TYPE_PALETTE)
+   {
+      int shift[4];
+      int channels = 0;
+      int bit_depth = row_info->bit_depth;
+
+      if (color_type & PNG_COLOR_MASK_COLOR)
+      {
+         shift[channels++] = bit_depth - sig_bits->red;
+         shift[channels++] = bit_depth - sig_bits->green;
+         shift[channels++] = bit_depth - sig_bits->blue;
+      }
+
+      else
+      {
+         shift[channels++] = bit_depth - sig_bits->gray;
+      }
+
+      if (color_type & PNG_COLOR_MASK_ALPHA)
+      {
+         shift[channels++] = bit_depth - sig_bits->alpha;
+      }
+
+      {
+         int c, have_shift;
+
+         for (c = have_shift = 0; c < channels; ++c)
+         {
+            /* A shift of more than the bit depth is an error condition but it
+             * gets ignored here.
+             */
+            if (shift[c] <= 0 || shift[c] >= bit_depth)
+               shift[c] = 0;
+
+            else
+               have_shift = 1;
+         }
+
+         if (!have_shift)
+            return;
+      }
+
+      switch (bit_depth)
+      {
+         default:
+         /* Must be 1bpp gray: should not be here! */
+            /* NOTREACHED */
+            break;
+
+         case 2:
+         /* Must be 2bpp gray */
+         /* assert(channels == 1 && shift[0] == 1) */
+         {
+            png_bytep bp = row;
+            png_bytep bp_end = bp + row_info->rowbytes;
+
+            while (bp < bp_end)
+            {
+               int b = (*bp >> 1) & 0x55;
+               *bp++ = (png_byte)b;
+            }
+            break;
+         }
+
+         case 4:
+         /* Must be 4bpp gray */
+         /* assert(channels == 1) */
+         {
+            png_bytep bp = row;
+            png_bytep bp_end = bp + row_info->rowbytes;
+            int gray_shift = shift[0];
+            int mask =  0xf >> gray_shift;
+
+            mask |= mask << 4;
+
+            while (bp < bp_end)
+            {
+               int b = (*bp >> gray_shift) & mask;
+               *bp++ = (png_byte)b;
+            }
+            break;
+         }
+
+         case 8:
+         /* Single byte components, G, GA, RGB, RGBA */
+         {
+            png_bytep bp = row;
+            png_bytep bp_end = bp + row_info->rowbytes;
+            int channel = 0;
+
+            while (bp < bp_end)
+            {
+               int b = *bp >> shift[channel];
+               if (++channel >= channels)
+                  channel = 0;
+               *bp++ = (png_byte)b;
+            }
+            break;
+         }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+         case 16:
+         /* Double byte components, G, GA, RGB, RGBA */
+         {
+            png_bytep bp = row;
+            png_bytep bp_end = bp + row_info->rowbytes;
+            int channel = 0;
+
+            while (bp < bp_end)
+            {
+               int value = (bp[0] << 8) + bp[1];
+
+               value >>= shift[channel];
+               if (++channel >= channels)
+                  channel = 0;
+               *bp++ = (png_byte)(value >> 8);
+               *bp++ = (png_byte)(value & 0xff);
+            }
+            break;
+         }
+#endif
+      }
+   }
+}
+#endif
+
+#ifdef PNG_READ_SCALE_16_TO_8_SUPPORTED
+/* Scale rows of bit depth 16 down to 8 accurately */
+void /* PRIVATE */
+png_do_scale_16_to_8(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_scale_16_to_8");
+
+   if (row_info->bit_depth == 16)
+   {
+      png_bytep sp = row; /* source */
+      png_bytep dp = row; /* destination */
+      png_bytep ep = sp + row_info->rowbytes; /* end+1 */
+
+      while (sp < ep)
+      {
+         /* The input is an array of 16 bit components, these must be scaled to
+          * 8 bits each.  For a 16 bit value V the required value (from the PNG
+          * specification) is:
+          *
+          *    (V * 255) / 65535
+          *
+          * This reduces to round(V / 257), or floor((V + 128.5)/257)
+          *
+          * Represent V as the two byte value vhi.vlo.  Make a guess that the
+          * result is the top byte of V, vhi, then the correction to this value
+          * is:
+          *
+          *    error = floor(((V-vhi.vhi) + 128.5) / 257)
+          *          = floor(((vlo-vhi) + 128.5) / 257)
+          *
+          * This can be approximated using integer arithmetic (and a signed
+          * shift):
+          *
+          *    error = (vlo-vhi+128) >> 8;
+          *
+          * The approximate differs from the exact answer only when (vlo-vhi) is
+          * 128; it then gives a correction of +1 when the exact correction is
+          * 0.  This gives 128 errors.  The exact answer (correct for all 16 bit
+          * input values) is:
+          *
+          *    error = (vlo-vhi+128)*65535 >> 24;
+          *
+          * An alternative arithmetic calculation which also gives no errors is:
+          *
+          *    (V * 255 + 32895) >> 16
+          */
+
+         png_int_32 tmp = *sp++; /* must be signed! */
+         tmp += (((int)*sp++ - tmp + 128) * 65535) >> 24;
+         *dp++ = (png_byte)tmp;
+      }
+
+      row_info->bit_depth = 8;
+      row_info->pixel_depth = (png_byte)(8 * row_info->channels);
+      row_info->rowbytes = row_info->width * row_info->channels;
+   }
+}
+#endif
+
+#ifdef PNG_READ_STRIP_16_TO_8_SUPPORTED
+void /* PRIVATE */
+/* Simply discard the low byte.  This was the default behavior prior
+ * to libpng-1.5.4.
+ */
+png_do_chop(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_chop");
+
+   if (row_info->bit_depth == 16)
+   {
+      png_bytep sp = row; /* source */
+      png_bytep dp = row; /* destination */
+      png_bytep ep = sp + row_info->rowbytes; /* end+1 */
+
+      while (sp < ep)
+      {
+         *dp++ = *sp;
+         sp += 2; /* skip low byte */
+      }
+
+      row_info->bit_depth = 8;
+      row_info->pixel_depth = (png_byte)(8 * row_info->channels);
+      row_info->rowbytes = row_info->width * row_info->channels;
+   }
+}
+#endif
+
+#ifdef PNG_READ_SWAP_ALPHA_SUPPORTED
+void /* PRIVATE */
+png_do_read_swap_alpha(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_read_swap_alpha");
+
+   {
+      png_uint_32 row_width = row_info->width;
+      if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+      {
+         /* This converts from RGBA to ARGB */
+         if (row_info->bit_depth == 8)
+         {
+            png_bytep sp = row + row_info->rowbytes;
+            png_bytep dp = sp;
+            png_byte save;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               save = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = save;
+            }
+         }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+         /* This converts from RRGGBBAA to AARRGGBB */
+         else
+         {
+            png_bytep sp = row + row_info->rowbytes;
+            png_bytep dp = sp;
+            png_byte save[2];
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               save[0] = *(--sp);
+               save[1] = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = save[0];
+               *(--dp) = save[1];
+            }
+         }
+#endif
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+      {
+         /* This converts from GA to AG */
+         if (row_info->bit_depth == 8)
+         {
+            png_bytep sp = row + row_info->rowbytes;
+            png_bytep dp = sp;
+            png_byte save;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               save = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = save;
+            }
+         }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+         /* This converts from GGAA to AAGG */
+         else
+         {
+            png_bytep sp = row + row_info->rowbytes;
+            png_bytep dp = sp;
+            png_byte save[2];
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               save[0] = *(--sp);
+               save[1] = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = save[0];
+               *(--dp) = save[1];
+            }
+         }
+#endif
+      }
+   }
+}
+#endif
+
+#ifdef PNG_READ_INVERT_ALPHA_SUPPORTED
+void /* PRIVATE */
+png_do_read_invert_alpha(png_row_infop row_info, png_bytep row)
+{
+   png_uint_32 row_width;
+   png_debug(1, "in png_do_read_invert_alpha");
+
+   row_width = row_info->width;
+   if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         /* This inverts the alpha channel in RGBA */
+         png_bytep sp = row + row_info->rowbytes;
+         png_bytep dp = sp;
+         png_uint_32 i;
+
+         for (i = 0; i < row_width; i++)
+         {
+            *(--dp) = (png_byte)(255 - *(--sp));
+
+/*          This does nothing:
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            We can replace it with:
+*/
+            sp-=3;
+            dp=sp;
+         }
+      }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+      /* This inverts the alpha channel in RRGGBBAA */
+      else
+      {
+         png_bytep sp = row + row_info->rowbytes;
+         png_bytep dp = sp;
+         png_uint_32 i;
+
+         for (i = 0; i < row_width; i++)
+         {
+            *(--dp) = (png_byte)(255 - *(--sp));
+            *(--dp) = (png_byte)(255 - *(--sp));
+
+/*          This does nothing:
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+            We can replace it with:
+*/
+            sp-=6;
+            dp=sp;
+         }
+      }
+#endif
+   }
+   else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         /* This inverts the alpha channel in GA */
+         png_bytep sp = row + row_info->rowbytes;
+         png_bytep dp = sp;
+         png_uint_32 i;
+
+         for (i = 0; i < row_width; i++)
+         {
+            *(--dp) = (png_byte)(255 - *(--sp));
+            *(--dp) = *(--sp);
+         }
+      }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+      else
+      {
+         /* This inverts the alpha channel in GGAA */
+         png_bytep sp  = row + row_info->rowbytes;
+         png_bytep dp = sp;
+         png_uint_32 i;
+
+         for (i = 0; i < row_width; i++)
+         {
+            *(--dp) = (png_byte)(255 - *(--sp));
+            *(--dp) = (png_byte)(255 - *(--sp));
+/*
+            *(--dp) = *(--sp);
+            *(--dp) = *(--sp);
+*/
+            sp-=2;
+            dp=sp;
+         }
+      }
+#endif
+   }
+}
+#endif
+
+#ifdef PNG_READ_FILLER_SUPPORTED
+/* Add filler channel if we have RGB color */
+void /* PRIVATE */
+png_do_read_filler(png_row_infop row_info, png_bytep row,
+    png_uint_32 filler, png_uint_32 flags)
+{
+   png_uint_32 i;
+   png_uint_32 row_width = row_info->width;
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+   png_byte hi_filler = (png_byte)((filler>>8) & 0xff);
+#endif
+   png_byte lo_filler = (png_byte)(filler & 0xff);
+
+   png_debug(1, "in png_do_read_filler");
+
+   if (
+       row_info->color_type == PNG_COLOR_TYPE_GRAY)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         if (flags & PNG_FLAG_FILLER_AFTER)
+         {
+            /* This changes the data from G to GX */
+            png_bytep sp = row + (png_size_t)row_width;
+            png_bytep dp =  sp + (png_size_t)row_width;
+            for (i = 1; i < row_width; i++)
+            {
+               *(--dp) = lo_filler;
+               *(--dp) = *(--sp);
+            }
+            *(--dp) = lo_filler;
+            row_info->channels = 2;
+            row_info->pixel_depth = 16;
+            row_info->rowbytes = row_width * 2;
+         }
+
+         else
+         {
+            /* This changes the data from G to XG */
+            png_bytep sp = row + (png_size_t)row_width;
+            png_bytep dp = sp  + (png_size_t)row_width;
+            for (i = 0; i < row_width; i++)
+            {
+               *(--dp) = *(--sp);
+               *(--dp) = lo_filler;
+            }
+            row_info->channels = 2;
+            row_info->pixel_depth = 16;
+            row_info->rowbytes = row_width * 2;
+         }
+      }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+      else if (row_info->bit_depth == 16)
+      {
+         if (flags & PNG_FLAG_FILLER_AFTER)
+         {
+            /* This changes the data from GG to GGXX */
+            png_bytep sp = row + (png_size_t)row_width * 2;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 1; i < row_width; i++)
+            {
+               *(--dp) = hi_filler;
+               *(--dp) = lo_filler;
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+            }
+            *(--dp) = hi_filler;
+            *(--dp) = lo_filler;
+            row_info->channels = 2;
+            row_info->pixel_depth = 32;
+            row_info->rowbytes = row_width * 4;
+         }
+
+         else
+         {
+            /* This changes the data from GG to XXGG */
+            png_bytep sp = row + (png_size_t)row_width * 2;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 0; i < row_width; i++)
+            {
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = hi_filler;
+               *(--dp) = lo_filler;
+            }
+            row_info->channels = 2;
+            row_info->pixel_depth = 32;
+            row_info->rowbytes = row_width * 4;
+         }
+      }
+#endif
+   } /* COLOR_TYPE == GRAY */
+   else if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         if (flags & PNG_FLAG_FILLER_AFTER)
+         {
+            /* This changes the data from RGB to RGBX */
+            png_bytep sp = row + (png_size_t)row_width * 3;
+            png_bytep dp = sp  + (png_size_t)row_width;
+            for (i = 1; i < row_width; i++)
+            {
+               *(--dp) = lo_filler;
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+            }
+            *(--dp) = lo_filler;
+            row_info->channels = 4;
+            row_info->pixel_depth = 32;
+            row_info->rowbytes = row_width * 4;
+         }
+
+         else
+         {
+            /* This changes the data from RGB to XRGB */
+            png_bytep sp = row + (png_size_t)row_width * 3;
+            png_bytep dp = sp + (png_size_t)row_width;
+            for (i = 0; i < row_width; i++)
+            {
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = lo_filler;
+            }
+            row_info->channels = 4;
+            row_info->pixel_depth = 32;
+            row_info->rowbytes = row_width * 4;
+         }
+      }
+
+#ifdef PNG_READ_16BIT_SUPPORTED
+      else if (row_info->bit_depth == 16)
+      {
+         if (flags & PNG_FLAG_FILLER_AFTER)
+         {
+            /* This changes the data from RRGGBB to RRGGBBXX */
+            png_bytep sp = row + (png_size_t)row_width * 6;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 1; i < row_width; i++)
+            {
+               *(--dp) = hi_filler;
+               *(--dp) = lo_filler;
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+            }
+            *(--dp) = hi_filler;
+            *(--dp) = lo_filler;
+            row_info->channels = 4;
+            row_info->pixel_depth = 64;
+            row_info->rowbytes = row_width * 8;
+         }
+
+         else
+         {
+            /* This changes the data from RRGGBB to XXRRGGBB */
+            png_bytep sp = row + (png_size_t)row_width * 6;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 0; i < row_width; i++)
+            {
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = *(--sp);
+               *(--dp) = hi_filler;
+               *(--dp) = lo_filler;
+            }
+
+            row_info->channels = 4;
+            row_info->pixel_depth = 64;
+            row_info->rowbytes = row_width * 8;
+         }
+      }
+#endif
+   } /* COLOR_TYPE == RGB */
+}
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+/* Expand grayscale files to RGB, with or without alpha */
+void /* PRIVATE */
+png_do_gray_to_rgb(png_row_infop row_info, png_bytep row)
+{
+   png_uint_32 i;
+   png_uint_32 row_width = row_info->width;
+
+   png_debug(1, "in png_do_gray_to_rgb");
+
+   if (row_info->bit_depth >= 8 &&
+       !(row_info->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_GRAY)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This changes G to RGB */
+            png_bytep sp = row + (png_size_t)row_width - 1;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 0; i < row_width; i++)
+            {
+               *(dp--) = *sp;
+               *(dp--) = *sp;
+               *(dp--) = *(sp--);
+            }
+         }
+
+         else
+         {
+            /* This changes GG to RRGGBB */
+            png_bytep sp = row + (png_size_t)row_width * 2 - 1;
+            png_bytep dp = sp  + (png_size_t)row_width * 4;
+            for (i = 0; i < row_width; i++)
+            {
+               *(dp--) = *sp;
+               *(dp--) = *(sp - 1);
+               *(dp--) = *sp;
+               *(dp--) = *(sp - 1);
+               *(dp--) = *(sp--);
+               *(dp--) = *(sp--);
+            }
+         }
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This changes GA to RGBA */
+            png_bytep sp = row + (png_size_t)row_width * 2 - 1;
+            png_bytep dp = sp  + (png_size_t)row_width * 2;
+            for (i = 0; i < row_width; i++)
+            {
+               *(dp--) = *(sp--);
+               *(dp--) = *sp;
+               *(dp--) = *sp;
+               *(dp--) = *(sp--);
+            }
+         }
+
+         else
+         {
+            /* This changes GGAA to RRGGBBAA */
+            png_bytep sp = row + (png_size_t)row_width * 4 - 1;
+            png_bytep dp = sp  + (png_size_t)row_width * 4;
+            for (i = 0; i < row_width; i++)
+            {
+               *(dp--) = *(sp--);
+               *(dp--) = *(sp--);
+               *(dp--) = *sp;
+               *(dp--) = *(sp - 1);
+               *(dp--) = *sp;
+               *(dp--) = *(sp - 1);
+               *(dp--) = *(sp--);
+               *(dp--) = *(sp--);
+            }
+         }
+      }
+      row_info->channels = (png_byte)(row_info->channels + 2);
+      row_info->color_type |= PNG_COLOR_MASK_COLOR;
+      row_info->pixel_depth = (png_byte)(row_info->channels *
+          row_info->bit_depth);
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_width);
+   }
+}
+#endif
+
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+/* Reduce RGB files to grayscale, with or without alpha
+ * using the equation given in Poynton's ColorFAQ of 1998-01-04 at
+ * <http://www.inforamp.net/~poynton/>  (THIS LINK IS DEAD June 2008 but
+ * versions dated 1998 through November 2002 have been archived at
+ * http://web.archive.org/web/20000816232553/http://www.inforamp.net/
+ * ~poynton/notes/colour_and_gamma/ColorFAQ.txt )
+ * Charles Poynton poynton at poynton.com
+ *
+ *     Y = 0.212671 * R + 0.715160 * G + 0.072169 * B
+ *
+ *  which can be expressed with integers as
+ *
+ *     Y = (6969 * R + 23434 * G + 2365 * B)/32768
+ *
+ * Poynton's current link (as of January 2003 through July 2011):
+ * <http://www.poynton.com/notes/colour_and_gamma/>
+ * has changed the numbers slightly:
+ *
+ *     Y = 0.2126*R + 0.7152*G + 0.0722*B
+ *
+ *  which can be expressed with integers as
+ *
+ *     Y = (6966 * R + 23436 * G + 2366 * B)/32768
+ *
+ *  Historically, however, libpng uses numbers derived from the ITU-R Rec 709
+ *  end point chromaticities and the D65 white point.  Depending on the
+ *  precision used for the D65 white point this produces a variety of different
+ *  numbers, however if the four decimal place value used in ITU-R Rec 709 is
+ *  used (0.3127,0.3290) the Y calculation would be:
+ *
+ *     Y = (6968 * R + 23435 * G + 2366 * B)/32768
+ *
+ *  While this is correct the rounding results in an overflow for white, because
+ *  the sum of the rounded coefficients is 32769, not 32768.  Consequently
+ *  libpng uses, instead, the closest non-overflowing approximation:
+ *
+ *     Y = (6968 * R + 23434 * G + 2366 * B)/32768
+ *
+ *  Starting with libpng-1.5.5, if the image being converted has a cHRM chunk
+ *  (including an sRGB chunk) then the chromaticities are used to calculate the
+ *  coefficients.  See the chunk handling in pngrutil.c for more information.
+ *
+ *  In all cases the calculation is to be done in a linear colorspace.  If no
+ *  gamma information is available to correct the encoding of the original RGB
+ *  values this results in an implicit assumption that the original PNG RGB
+ *  values were linear.
+ *
+ *  Other integer coefficents can be used via png_set_rgb_to_gray().  Because
+ *  the API takes just red and green coefficients the blue coefficient is
+ *  calculated to make the sum 32768.  This will result in different rounding
+ *  to that used above.
+ */
+int /* PRIVATE */
+png_do_rgb_to_gray(png_structrp png_ptr, png_row_infop row_info, png_bytep row)
+
+{
+   int rgb_error = 0;
+
+   png_debug(1, "in png_do_rgb_to_gray");
+
+   if (!(row_info->color_type & PNG_COLOR_MASK_PALETTE) &&
+       (row_info->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      PNG_CONST png_uint_32 rc = png_ptr->rgb_to_gray_red_coeff;
+      PNG_CONST png_uint_32 gc = png_ptr->rgb_to_gray_green_coeff;
+      PNG_CONST png_uint_32 bc = 32768 - rc - gc;
+      PNG_CONST png_uint_32 row_width = row_info->width;
+      PNG_CONST int have_alpha =
+         (row_info->color_type & PNG_COLOR_MASK_ALPHA) != 0;
+
+      if (row_info->bit_depth == 8)
+      {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+         /* Notice that gamma to/from 1 are not necessarily inverses (if
+          * there is an overall gamma correction).  Prior to 1.5.5 this code
+          * checked the linearized values for equality; this doesn't match
+          * the documentation, the original values must be checked.
+          */
+         if (png_ptr->gamma_from_1 != NULL && png_ptr->gamma_to_1 != NULL)
+         {
+            png_bytep sp = row;
+            png_bytep dp = row;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_byte red   = *(sp++);
+               png_byte green = *(sp++);
+               png_byte blue  = *(sp++);
+
+               if (red != green || red != blue)
+               {
+                  red = png_ptr->gamma_to_1[red];
+                  green = png_ptr->gamma_to_1[green];
+                  blue = png_ptr->gamma_to_1[blue];
+
+                  rgb_error |= 1;
+                  *(dp++) = png_ptr->gamma_from_1[
+                      (rc*red + gc*green + bc*blue + 16384)>>15];
+               }
+
+               else
+               {
+                  /* If there is no overall correction the table will not be
+                   * set.
+                   */
+                  if (png_ptr->gamma_table != NULL)
+                     red = png_ptr->gamma_table[red];
+
+                  *(dp++) = red;
+               }
+
+               if (have_alpha)
+                  *(dp++) = *(sp++);
+            }
+         }
+         else
+#endif
+         {
+            png_bytep sp = row;
+            png_bytep dp = row;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_byte red   = *(sp++);
+               png_byte green = *(sp++);
+               png_byte blue  = *(sp++);
+
+               if (red != green || red != blue)
+               {
+                  rgb_error |= 1;
+                  /* NOTE: this is the historical approach which simply
+                   * truncates the results.
+                   */
+                  *(dp++) = (png_byte)((rc*red + gc*green + bc*blue)>>15);
+               }
+
+               else
+                  *(dp++) = red;
+
+               if (have_alpha)
+                  *(dp++) = *(sp++);
+            }
+         }
+      }
+
+      else /* RGB bit_depth == 16 */
+      {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+         if (png_ptr->gamma_16_to_1 != NULL && png_ptr->gamma_16_from_1 != NULL)
+         {
+            png_bytep sp = row;
+            png_bytep dp = row;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_uint_16 red, green, blue, w;
+
+               red   = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+               green = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+               blue  = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+
+               if (red == green && red == blue)
+               {
+                  if (png_ptr->gamma_16_table != NULL)
+                     w = png_ptr->gamma_16_table[(red&0xff)
+                         >> png_ptr->gamma_shift][red>>8];
+
+                  else
+                     w = red;
+               }
+
+               else
+               {
+                  png_uint_16 red_1   = png_ptr->gamma_16_to_1[(red&0xff)
+                      >> png_ptr->gamma_shift][red>>8];
+                  png_uint_16 green_1 =
+                      png_ptr->gamma_16_to_1[(green&0xff) >>
+                      png_ptr->gamma_shift][green>>8];
+                  png_uint_16 blue_1  = png_ptr->gamma_16_to_1[(blue&0xff)
+                      >> png_ptr->gamma_shift][blue>>8];
+                  png_uint_16 gray16  = (png_uint_16)((rc*red_1 + gc*green_1
+                      + bc*blue_1 + 16384)>>15);
+                  w = png_ptr->gamma_16_from_1[(gray16&0xff) >>
+                      png_ptr->gamma_shift][gray16 >> 8];
+                  rgb_error |= 1;
+               }
+
+               *(dp++) = (png_byte)((w>>8) & 0xff);
+               *(dp++) = (png_byte)(w & 0xff);
+
+               if (have_alpha)
+               {
+                  *(dp++) = *(sp++);
+                  *(dp++) = *(sp++);
+               }
+            }
+         }
+         else
+#endif
+         {
+            png_bytep sp = row;
+            png_bytep dp = row;
+            png_uint_32 i;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_uint_16 red, green, blue, gray16;
+
+               red   = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+               green = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+               blue  = (png_uint_16)(((*(sp))<<8) | *(sp + 1)); sp += 2;
+
+               if (red != green || red != blue)
+                  rgb_error |= 1;
+
+               /* From 1.5.5 in the 16 bit case do the accurate conversion even
+                * in the 'fast' case - this is because this is where the code
+                * ends up when handling linear 16 bit data.
+                */
+               gray16  = (png_uint_16)((rc*red + gc*green + bc*blue + 16384) >>
+                  15);
+               *(dp++) = (png_byte)((gray16>>8) & 0xff);
+               *(dp++) = (png_byte)(gray16 & 0xff);
+
+               if (have_alpha)
+               {
+                  *(dp++) = *(sp++);
+                  *(dp++) = *(sp++);
+               }
+            }
+         }
+      }
+
+      row_info->channels = (png_byte)(row_info->channels - 2);
+      row_info->color_type = (png_byte)(row_info->color_type &
+          ~PNG_COLOR_MASK_COLOR);
+      row_info->pixel_depth = (png_byte)(row_info->channels *
+          row_info->bit_depth);
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_width);
+   }
+   return rgb_error;
+}
+#endif
+#endif /* PNG_READ_TRANSFORMS_SUPPORTED */
+
+#ifdef PNG_BUILD_GRAYSCALE_PALETTE_SUPPORTED
+/* Build a grayscale palette.  Palette is assumed to be 1 << bit_depth
+ * large of png_color.  This lets grayscale images be treated as
+ * paletted.  Most useful for gamma correction and simplification
+ * of code.  This API is not used internally.
+ */
+void PNGAPI
+png_build_grayscale_palette(int bit_depth, png_colorp palette)
+{
+   int num_palette;
+   int color_inc;
+   int i;
+   int v;
+
+   png_debug(1, "in png_do_build_grayscale_palette");
+
+   if (palette == NULL)
+      return;
+
+   switch (bit_depth)
+   {
+      case 1:
+         num_palette = 2;
+         color_inc = 0xff;
+         break;
+
+      case 2:
+         num_palette = 4;
+         color_inc = 0x55;
+         break;
+
+      case 4:
+         num_palette = 16;
+         color_inc = 0x11;
+         break;
+
+      case 8:
+         num_palette = 256;
+         color_inc = 1;
+         break;
+
+      default:
+         num_palette = 0;
+         color_inc = 0;
+         break;
+   }
+
+   for (i = 0, v = 0; i < num_palette; i++, v += color_inc)
+   {
+      palette[i].red = (png_byte)v;
+      palette[i].green = (png_byte)v;
+      palette[i].blue = (png_byte)v;
+   }
+}
+#endif
+
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+/* Replace any alpha or transparency with the supplied background color.
+ * "background" is already in the screen gamma, while "background_1" is
+ * at a gamma of 1.0.  Paletted files have already been taken care of.
+ */
+void /* PRIVATE */
+png_do_compose(png_row_infop row_info, png_bytep row, png_structrp png_ptr)
+{
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   png_const_bytep gamma_table = png_ptr->gamma_table;
+   png_const_bytep gamma_from_1 = png_ptr->gamma_from_1;
+   png_const_bytep gamma_to_1 = png_ptr->gamma_to_1;
+   png_const_uint_16pp gamma_16 = png_ptr->gamma_16_table;
+   png_const_uint_16pp gamma_16_from_1 = png_ptr->gamma_16_from_1;
+   png_const_uint_16pp gamma_16_to_1 = png_ptr->gamma_16_to_1;
+   int gamma_shift = png_ptr->gamma_shift;
+   int optimize = (png_ptr->flags & PNG_FLAG_OPTIMIZE_ALPHA) != 0;
+#endif
+
+   png_bytep sp;
+   png_uint_32 i;
+   png_uint_32 row_width = row_info->width;
+   int shift;
+
+   png_debug(1, "in png_do_compose");
+
+   {
+      switch (row_info->color_type)
+      {
+         case PNG_COLOR_TYPE_GRAY:
+         {
+            switch (row_info->bit_depth)
+            {
+               case 1:
+               {
+                  sp = row;
+                  shift = 7;
+                  for (i = 0; i < row_width; i++)
+                  {
+                     if ((png_uint_16)((*sp >> shift) & 0x01)
+                        == png_ptr->trans_color.gray)
+                     {
+                        unsigned int tmp = *sp & (0x7f7f >> (7 - shift));
+                        tmp |= png_ptr->background.gray << shift;
+                        *sp = (png_byte)(tmp & 0xff);
+                     }
+
+                     if (!shift)
+                     {
+                        shift = 7;
+                        sp++;
+                     }
+
+                     else
+                        shift--;
+                  }
+                  break;
+               }
+
+               case 2:
+               {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+                  if (gamma_table != NULL)
+                  {
+                     sp = row;
+                     shift = 6;
+                     for (i = 0; i < row_width; i++)
+                     {
+                        if ((png_uint_16)((*sp >> shift) & 0x03)
+                            == png_ptr->trans_color.gray)
+                        {
+                           unsigned int tmp = *sp & (0x3f3f >> (6 - shift));
+                           tmp |= png_ptr->background.gray << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        else
+                        {
+                           unsigned int p = (*sp >> shift) & 0x03;
+                           unsigned int g = (gamma_table [p | (p << 2) |
+                               (p << 4) | (p << 6)] >> 6) & 0x03;
+                           unsigned int tmp = *sp & (0x3f3f >> (6 - shift));
+                           tmp |= g << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        if (!shift)
+                        {
+                           shift = 6;
+                           sp++;
+                        }
+
+                        else
+                           shift -= 2;
+                     }
+                  }
+
+                  else
+#endif
+                  {
+                     sp = row;
+                     shift = 6;
+                     for (i = 0; i < row_width; i++)
+                     {
+                        if ((png_uint_16)((*sp >> shift) & 0x03)
+                            == png_ptr->trans_color.gray)
+                        {
+                           unsigned int tmp = *sp & (0x3f3f >> (6 - shift));
+                           tmp |= png_ptr->background.gray << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        if (!shift)
+                        {
+                           shift = 6;
+                           sp++;
+                        }
+
+                        else
+                           shift -= 2;
+                     }
+                  }
+                  break;
+               }
+
+               case 4:
+               {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+                  if (gamma_table != NULL)
+                  {
+                     sp = row;
+                     shift = 4;
+                     for (i = 0; i < row_width; i++)
+                     {
+                        if ((png_uint_16)((*sp >> shift) & 0x0f)
+                            == png_ptr->trans_color.gray)
+                        {
+                           unsigned int tmp = *sp & (0xf0f >> (4 - shift));
+                           tmp |= png_ptr->background.gray << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        else
+                        {
+                           unsigned int p = (*sp >> shift) & 0x0f;
+                           unsigned int g = (gamma_table[p | (p << 4)] >> 4) &
+                              0x0f;
+                           unsigned int tmp = *sp & (0xf0f >> (4 - shift));
+                           tmp |= g << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        if (!shift)
+                        {
+                           shift = 4;
+                           sp++;
+                        }
+
+                        else
+                           shift -= 4;
+                     }
+                  }
+
+                  else
+#endif
+                  {
+                     sp = row;
+                     shift = 4;
+                     for (i = 0; i < row_width; i++)
+                     {
+                        if ((png_uint_16)((*sp >> shift) & 0x0f)
+                            == png_ptr->trans_color.gray)
+                        {
+                           unsigned int tmp = *sp & (0xf0f >> (4 - shift));
+                           tmp |= png_ptr->background.gray << shift;
+                           *sp = (png_byte)(tmp & 0xff);
+                        }
+
+                        if (!shift)
+                        {
+                           shift = 4;
+                           sp++;
+                        }
+
+                        else
+                           shift -= 4;
+                     }
+                  }
+                  break;
+               }
+
+               case 8:
+               {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+                  if (gamma_table != NULL)
+                  {
+                     sp = row;
+                     for (i = 0; i < row_width; i++, sp++)
+                     {
+                        if (*sp == png_ptr->trans_color.gray)
+                           *sp = (png_byte)png_ptr->background.gray;
+
+                        else
+                           *sp = gamma_table[*sp];
+                     }
+                  }
+                  else
+#endif
+                  {
+                     sp = row;
+                     for (i = 0; i < row_width; i++, sp++)
+                     {
+                        if (*sp == png_ptr->trans_color.gray)
+                           *sp = (png_byte)png_ptr->background.gray;
+                     }
+                  }
+                  break;
+               }
+
+               case 16:
+               {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+                  if (gamma_16 != NULL)
+                  {
+                     sp = row;
+                     for (i = 0; i < row_width; i++, sp += 2)
+                     {
+                        png_uint_16 v;
+
+                        v = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+
+                        if (v == png_ptr->trans_color.gray)
+                        {
+                           /* Background is already in screen gamma */
+                           *sp = (png_byte)((png_ptr->background.gray >> 8)
+                                & 0xff);
+                           *(sp + 1) = (png_byte)(png_ptr->background.gray
+                                & 0xff);
+                        }
+
+                        else
+                        {
+                           v = gamma_16[*(sp + 1) >> gamma_shift][*sp];
+                           *sp = (png_byte)((v >> 8) & 0xff);
+                           *(sp + 1) = (png_byte)(v & 0xff);
+                        }
+                     }
+                  }
+                  else
+#endif
+                  {
+                     sp = row;
+                     for (i = 0; i < row_width; i++, sp += 2)
+                     {
+                        png_uint_16 v;
+
+                        v = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+
+                        if (v == png_ptr->trans_color.gray)
+                        {
+                           *sp = (png_byte)((png_ptr->background.gray >> 8)
+                                & 0xff);
+                           *(sp + 1) = (png_byte)(png_ptr->background.gray
+                                & 0xff);
+                        }
+                     }
+                  }
+                  break;
+               }
+
+               default:
+                  break;
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_RGB:
+         {
+            if (row_info->bit_depth == 8)
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_table != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 3)
+                  {
+                     if (*sp == png_ptr->trans_color.red &&
+                         *(sp + 1) == png_ptr->trans_color.green &&
+                         *(sp + 2) == png_ptr->trans_color.blue)
+                     {
+                        *sp = (png_byte)png_ptr->background.red;
+                        *(sp + 1) = (png_byte)png_ptr->background.green;
+                        *(sp + 2) = (png_byte)png_ptr->background.blue;
+                     }
+
+                     else
+                     {
+                        *sp = gamma_table[*sp];
+                        *(sp + 1) = gamma_table[*(sp + 1)];
+                        *(sp + 2) = gamma_table[*(sp + 2)];
+                     }
+                  }
+               }
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 3)
+                  {
+                     if (*sp == png_ptr->trans_color.red &&
+                         *(sp + 1) == png_ptr->trans_color.green &&
+                         *(sp + 2) == png_ptr->trans_color.blue)
+                     {
+                        *sp = (png_byte)png_ptr->background.red;
+                        *(sp + 1) = (png_byte)png_ptr->background.green;
+                        *(sp + 2) = (png_byte)png_ptr->background.blue;
+                     }
+                  }
+               }
+            }
+            else /* if (row_info->bit_depth == 16) */
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_16 != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 6)
+                  {
+                     png_uint_16 r = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+
+                     png_uint_16 g = (png_uint_16)(((*(sp + 2)) << 8)
+                         + *(sp + 3));
+
+                     png_uint_16 b = (png_uint_16)(((*(sp + 4)) << 8)
+                         + *(sp + 5));
+
+                     if (r == png_ptr->trans_color.red &&
+                         g == png_ptr->trans_color.green &&
+                         b == png_ptr->trans_color.blue)
+                     {
+                        /* Background is already in screen gamma */
+                        *sp = (png_byte)((png_ptr->background.red >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.red & 0xff);
+                        *(sp + 2) = (png_byte)((png_ptr->background.green >> 8)
+                                & 0xff);
+                        *(sp + 3) = (png_byte)(png_ptr->background.green
+                                & 0xff);
+                        *(sp + 4) = (png_byte)((png_ptr->background.blue >> 8)
+                                & 0xff);
+                        *(sp + 5) = (png_byte)(png_ptr->background.blue & 0xff);
+                     }
+
+                     else
+                     {
+                        png_uint_16 v = gamma_16[*(sp + 1) >> gamma_shift][*sp];
+                        *sp = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(v & 0xff);
+
+                        v = gamma_16[*(sp + 3) >> gamma_shift][*(sp + 2)];
+                        *(sp + 2) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 3) = (png_byte)(v & 0xff);
+
+                        v = gamma_16[*(sp + 5) >> gamma_shift][*(sp + 4)];
+                        *(sp + 4) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 5) = (png_byte)(v & 0xff);
+                     }
+                  }
+               }
+
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 6)
+                  {
+                     png_uint_16 r = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+
+                     png_uint_16 g = (png_uint_16)(((*(sp + 2)) << 8)
+                         + *(sp + 3));
+
+                     png_uint_16 b = (png_uint_16)(((*(sp + 4)) << 8)
+                         + *(sp + 5));
+
+                     if (r == png_ptr->trans_color.red &&
+                         g == png_ptr->trans_color.green &&
+                         b == png_ptr->trans_color.blue)
+                     {
+                        *sp = (png_byte)((png_ptr->background.red >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.red & 0xff);
+                        *(sp + 2) = (png_byte)((png_ptr->background.green >> 8)
+                                & 0xff);
+                        *(sp + 3) = (png_byte)(png_ptr->background.green
+                                & 0xff);
+                        *(sp + 4) = (png_byte)((png_ptr->background.blue >> 8)
+                                & 0xff);
+                        *(sp + 5) = (png_byte)(png_ptr->background.blue & 0xff);
+                     }
+                  }
+               }
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_GRAY_ALPHA:
+         {
+            if (row_info->bit_depth == 8)
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_to_1 != NULL && gamma_from_1 != NULL &&
+                   gamma_table != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 2)
+                  {
+                     png_uint_16 a = *(sp + 1);
+
+                     if (a == 0xff)
+                        *sp = gamma_table[*sp];
+
+                     else if (a == 0)
+                     {
+                        /* Background is already in screen gamma */
+                        *sp = (png_byte)png_ptr->background.gray;
+                     }
+
+                     else
+                     {
+                        png_byte v, w;
+
+                        v = gamma_to_1[*sp];
+                        png_composite(w, v, a, png_ptr->background_1.gray);
+                        if (!optimize)
+                           w = gamma_from_1[w];
+                        *sp = w;
+                     }
+                  }
+               }
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 2)
+                  {
+                     png_byte a = *(sp + 1);
+
+                     if (a == 0)
+                        *sp = (png_byte)png_ptr->background.gray;
+
+                     else if (a < 0xff)
+                        png_composite(*sp, *sp, a, png_ptr->background.gray);
+                  }
+               }
+            }
+            else /* if (png_ptr->bit_depth == 16) */
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_16 != NULL && gamma_16_from_1 != NULL &&
+                   gamma_16_to_1 != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 4)
+                  {
+                     png_uint_16 a = (png_uint_16)(((*(sp + 2)) << 8)
+                         + *(sp + 3));
+
+                     if (a == (png_uint_16)0xffff)
+                     {
+                        png_uint_16 v;
+
+                        v = gamma_16[*(sp + 1) >> gamma_shift][*sp];
+                        *sp = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(v & 0xff);
+                     }
+
+                     else if (a == 0)
+                     {
+                        /* Background is already in screen gamma */
+                        *sp = (png_byte)((png_ptr->background.gray >> 8)
+                                & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.gray & 0xff);
+                     }
+
+                     else
+                     {
+                        png_uint_16 g, v, w;
+
+                        g = gamma_16_to_1[*(sp + 1) >> gamma_shift][*sp];
+                        png_composite_16(v, g, a, png_ptr->background_1.gray);
+                        if (optimize)
+                           w = v;
+                        else
+                           w = gamma_16_from_1[(v&0xff) >> gamma_shift][v >> 8];
+                        *sp = (png_byte)((w >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(w & 0xff);
+                     }
+                  }
+               }
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 4)
+                  {
+                     png_uint_16 a = (png_uint_16)(((*(sp + 2)) << 8)
+                         + *(sp + 3));
+
+                     if (a == 0)
+                     {
+                        *sp = (png_byte)((png_ptr->background.gray >> 8)
+                                & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.gray & 0xff);
+                     }
+
+                     else if (a < 0xffff)
+                     {
+                        png_uint_16 g, v;
+
+                        g = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+                        png_composite_16(v, g, a, png_ptr->background.gray);
+                        *sp = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(v & 0xff);
+                     }
+                  }
+               }
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_RGB_ALPHA:
+         {
+            if (row_info->bit_depth == 8)
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_to_1 != NULL && gamma_from_1 != NULL &&
+                   gamma_table != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 4)
+                  {
+                     png_byte a = *(sp + 3);
+
+                     if (a == 0xff)
+                     {
+                        *sp = gamma_table[*sp];
+                        *(sp + 1) = gamma_table[*(sp + 1)];
+                        *(sp + 2) = gamma_table[*(sp + 2)];
+                     }
+
+                     else if (a == 0)
+                     {
+                        /* Background is already in screen gamma */
+                        *sp = (png_byte)png_ptr->background.red;
+                        *(sp + 1) = (png_byte)png_ptr->background.green;
+                        *(sp + 2) = (png_byte)png_ptr->background.blue;
+                     }
+
+                     else
+                     {
+                        png_byte v, w;
+
+                        v = gamma_to_1[*sp];
+                        png_composite(w, v, a, png_ptr->background_1.red);
+                        if (!optimize) w = gamma_from_1[w];
+                        *sp = w;
+
+                        v = gamma_to_1[*(sp + 1)];
+                        png_composite(w, v, a, png_ptr->background_1.green);
+                        if (!optimize) w = gamma_from_1[w];
+                        *(sp + 1) = w;
+
+                        v = gamma_to_1[*(sp + 2)];
+                        png_composite(w, v, a, png_ptr->background_1.blue);
+                        if (!optimize) w = gamma_from_1[w];
+                        *(sp + 2) = w;
+                     }
+                  }
+               }
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 4)
+                  {
+                     png_byte a = *(sp + 3);
+
+                     if (a == 0)
+                     {
+                        *sp = (png_byte)png_ptr->background.red;
+                        *(sp + 1) = (png_byte)png_ptr->background.green;
+                        *(sp + 2) = (png_byte)png_ptr->background.blue;
+                     }
+
+                     else if (a < 0xff)
+                     {
+                        png_composite(*sp, *sp, a, png_ptr->background.red);
+
+                        png_composite(*(sp + 1), *(sp + 1), a,
+                            png_ptr->background.green);
+
+                        png_composite(*(sp + 2), *(sp + 2), a,
+                            png_ptr->background.blue);
+                     }
+                  }
+               }
+            }
+            else /* if (row_info->bit_depth == 16) */
+            {
+#ifdef PNG_READ_GAMMA_SUPPORTED
+               if (gamma_16 != NULL && gamma_16_from_1 != NULL &&
+                   gamma_16_to_1 != NULL)
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 8)
+                  {
+                     png_uint_16 a = (png_uint_16)(((png_uint_16)(*(sp + 6))
+                         << 8) + (png_uint_16)(*(sp + 7)));
+
+                     if (a == (png_uint_16)0xffff)
+                     {
+                        png_uint_16 v;
+
+                        v = gamma_16[*(sp + 1) >> gamma_shift][*sp];
+                        *sp = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(v & 0xff);
+
+                        v = gamma_16[*(sp + 3) >> gamma_shift][*(sp + 2)];
+                        *(sp + 2) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 3) = (png_byte)(v & 0xff);
+
+                        v = gamma_16[*(sp + 5) >> gamma_shift][*(sp + 4)];
+                        *(sp + 4) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 5) = (png_byte)(v & 0xff);
+                     }
+
+                     else if (a == 0)
+                     {
+                        /* Background is already in screen gamma */
+                        *sp = (png_byte)((png_ptr->background.red >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.red & 0xff);
+                        *(sp + 2) = (png_byte)((png_ptr->background.green >> 8)
+                                & 0xff);
+                        *(sp + 3) = (png_byte)(png_ptr->background.green
+                                & 0xff);
+                        *(sp + 4) = (png_byte)((png_ptr->background.blue >> 8)
+                                & 0xff);
+                        *(sp + 5) = (png_byte)(png_ptr->background.blue & 0xff);
+                     }
+
+                     else
+                     {
+                        png_uint_16 v, w;
+
+                        v = gamma_16_to_1[*(sp + 1) >> gamma_shift][*sp];
+                        png_composite_16(w, v, a, png_ptr->background_1.red);
+                        if (!optimize)
+                           w = gamma_16_from_1[((w&0xff) >> gamma_shift)][w >>
+                                8];
+                        *sp = (png_byte)((w >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(w & 0xff);
+
+                        v = gamma_16_to_1[*(sp + 3) >> gamma_shift][*(sp + 2)];
+                        png_composite_16(w, v, a, png_ptr->background_1.green);
+                        if (!optimize)
+                           w = gamma_16_from_1[((w&0xff) >> gamma_shift)][w >>
+                                8];
+
+                        *(sp + 2) = (png_byte)((w >> 8) & 0xff);
+                        *(sp + 3) = (png_byte)(w & 0xff);
+
+                        v = gamma_16_to_1[*(sp + 5) >> gamma_shift][*(sp + 4)];
+                        png_composite_16(w, v, a, png_ptr->background_1.blue);
+                        if (!optimize)
+                           w = gamma_16_from_1[((w&0xff) >> gamma_shift)][w >>
+                                8];
+
+                        *(sp + 4) = (png_byte)((w >> 8) & 0xff);
+                        *(sp + 5) = (png_byte)(w & 0xff);
+                     }
+                  }
+               }
+
+               else
+#endif
+               {
+                  sp = row;
+                  for (i = 0; i < row_width; i++, sp += 8)
+                  {
+                     png_uint_16 a = (png_uint_16)(((png_uint_16)(*(sp + 6))
+                         << 8) + (png_uint_16)(*(sp + 7)));
+
+                     if (a == 0)
+                     {
+                        *sp = (png_byte)((png_ptr->background.red >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(png_ptr->background.red & 0xff);
+                        *(sp + 2) = (png_byte)((png_ptr->background.green >> 8)
+                                & 0xff);
+                        *(sp + 3) = (png_byte)(png_ptr->background.green
+                                & 0xff);
+                        *(sp + 4) = (png_byte)((png_ptr->background.blue >> 8)
+                                & 0xff);
+                        *(sp + 5) = (png_byte)(png_ptr->background.blue & 0xff);
+                     }
+
+                     else if (a < 0xffff)
+                     {
+                        png_uint_16 v;
+
+                        png_uint_16 r = (png_uint_16)(((*sp) << 8) + *(sp + 1));
+                        png_uint_16 g = (png_uint_16)(((*(sp + 2)) << 8)
+                            + *(sp + 3));
+                        png_uint_16 b = (png_uint_16)(((*(sp + 4)) << 8)
+                            + *(sp + 5));
+
+                        png_composite_16(v, r, a, png_ptr->background.red);
+                        *sp = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 1) = (png_byte)(v & 0xff);
+
+                        png_composite_16(v, g, a, png_ptr->background.green);
+                        *(sp + 2) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 3) = (png_byte)(v & 0xff);
+
+                        png_composite_16(v, b, a, png_ptr->background.blue);
+                        *(sp + 4) = (png_byte)((v >> 8) & 0xff);
+                        *(sp + 5) = (png_byte)(v & 0xff);
+                     }
+                  }
+               }
+            }
+            break;
+         }
+
+         default:
+            break;
+      }
+   }
+}
+#endif /* PNG_READ_BACKGROUND_SUPPORTED || PNG_READ_ALPHA_MODE_SUPPORTED */
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+/* Gamma correct the image, avoiding the alpha channel.  Make sure
+ * you do this after you deal with the transparency issue on grayscale
+ * or RGB images. If your bit depth is 8, use gamma_table, if it
+ * is 16, use gamma_16_table and gamma_shift.  Build these with
+ * build_gamma_table().
+ */
+void /* PRIVATE */
+png_do_gamma(png_row_infop row_info, png_bytep row, png_structrp png_ptr)
+{
+   png_const_bytep gamma_table = png_ptr->gamma_table;
+   png_const_uint_16pp gamma_16_table = png_ptr->gamma_16_table;
+   int gamma_shift = png_ptr->gamma_shift;
+
+   png_bytep sp;
+   png_uint_32 i;
+   png_uint_32 row_width=row_info->width;
+
+   png_debug(1, "in png_do_gamma");
+
+   if (((row_info->bit_depth <= 8 && gamma_table != NULL) ||
+       (row_info->bit_depth == 16 && gamma_16_table != NULL)))
+   {
+      switch (row_info->color_type)
+      {
+         case PNG_COLOR_TYPE_RGB:
+         {
+            if (row_info->bit_depth == 8)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  *sp = gamma_table[*sp];
+                  sp++;
+                  *sp = gamma_table[*sp];
+                  sp++;
+                  *sp = gamma_table[*sp];
+                  sp++;
+               }
+            }
+
+            else /* if (row_info->bit_depth == 16) */
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  png_uint_16 v;
+
+                  v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+
+                  v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+
+                  v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+               }
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_RGB_ALPHA:
+         {
+            if (row_info->bit_depth == 8)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  *sp = gamma_table[*sp];
+                  sp++;
+
+                  *sp = gamma_table[*sp];
+                  sp++;
+
+                  *sp = gamma_table[*sp];
+                  sp++;
+
+                  sp++;
+               }
+            }
+
+            else /* if (row_info->bit_depth == 16) */
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  png_uint_16 v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+
+                  v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+
+                  v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 4;
+               }
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_GRAY_ALPHA:
+         {
+            if (row_info->bit_depth == 8)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  *sp = gamma_table[*sp];
+                  sp += 2;
+               }
+            }
+
+            else /* if (row_info->bit_depth == 16) */
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  png_uint_16 v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 4;
+               }
+            }
+            break;
+         }
+
+         case PNG_COLOR_TYPE_GRAY:
+         {
+            if (row_info->bit_depth == 2)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i += 4)
+               {
+                  int a = *sp & 0xc0;
+                  int b = *sp & 0x30;
+                  int c = *sp & 0x0c;
+                  int d = *sp & 0x03;
+
+                  *sp = (png_byte)(
+                      ((((int)gamma_table[a|(a>>2)|(a>>4)|(a>>6)])   ) & 0xc0)|
+                      ((((int)gamma_table[(b<<2)|b|(b>>2)|(b>>4)])>>2) & 0x30)|
+                      ((((int)gamma_table[(c<<4)|(c<<2)|c|(c>>2)])>>4) & 0x0c)|
+                      ((((int)gamma_table[(d<<6)|(d<<4)|(d<<2)|d])>>6) ));
+                  sp++;
+               }
+            }
+
+            if (row_info->bit_depth == 4)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i += 2)
+               {
+                  int msb = *sp & 0xf0;
+                  int lsb = *sp & 0x0f;
+
+                  *sp = (png_byte)((((int)gamma_table[msb | (msb >> 4)]) & 0xf0)
+                      | (((int)gamma_table[(lsb << 4) | lsb]) >> 4));
+                  sp++;
+               }
+            }
+
+            else if (row_info->bit_depth == 8)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  *sp = gamma_table[*sp];
+                  sp++;
+               }
+            }
+
+            else if (row_info->bit_depth == 16)
+            {
+               sp = row;
+               for (i = 0; i < row_width; i++)
+               {
+                  png_uint_16 v = gamma_16_table[*(sp + 1) >> gamma_shift][*sp];
+                  *sp = (png_byte)((v >> 8) & 0xff);
+                  *(sp + 1) = (png_byte)(v & 0xff);
+                  sp += 2;
+               }
+            }
+            break;
+         }
+
+         default:
+            break;
+      }
+   }
+}
+#endif
+
+#ifdef PNG_READ_ALPHA_MODE_SUPPORTED
+/* Encode the alpha channel to the output gamma (the input channel is always
+ * linear.)  Called only with color types that have an alpha channel.  Needs the
+ * from_1 tables.
+ */
+void /* PRIVATE */
+png_do_encode_alpha(png_row_infop row_info, png_bytep row, png_structrp png_ptr)
+{
+   png_uint_32 row_width = row_info->width;
+
+   png_debug(1, "in png_do_encode_alpha");
+
+   if (row_info->color_type & PNG_COLOR_MASK_ALPHA)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         PNG_CONST png_bytep table = png_ptr->gamma_from_1;
+
+         if (table != NULL)
+         {
+            PNG_CONST int step =
+               (row_info->color_type & PNG_COLOR_MASK_COLOR) ? 4 : 2;
+
+            /* The alpha channel is the last component: */
+            row += step - 1;
+
+            for (; row_width > 0; --row_width, row += step)
+               *row = table[*row];
+
+            return;
+         }
+      }
+
+      else if (row_info->bit_depth == 16)
+      {
+         PNG_CONST png_uint_16pp table = png_ptr->gamma_16_from_1;
+         PNG_CONST int gamma_shift = png_ptr->gamma_shift;
+
+         if (table != NULL)
+         {
+            PNG_CONST int step =
+               (row_info->color_type & PNG_COLOR_MASK_COLOR) ? 8 : 4;
+
+            /* The alpha channel is the last component: */
+            row += step - 2;
+
+            for (; row_width > 0; --row_width, row += step)
+            {
+               png_uint_16 v;
+
+               v = table[*(row + 1) >> gamma_shift][*row];
+               *row = (png_byte)((v >> 8) & 0xff);
+               *(row + 1) = (png_byte)(v & 0xff);
+            }
+
+            return;
+         }
+      }
+   }
+
+   /* Only get to here if called with a weird row_info; no harm has been done,
+    * so just issue a warning.
+    */
+   png_warning(png_ptr, "png_do_encode_alpha: unexpected call");
+}
+#endif
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+/* Expands a palette row to an RGB or RGBA row depending
+ * upon whether you supply trans and num_trans.
+ */
+void /* PRIVATE */
+png_do_expand_palette(png_row_infop row_info, png_bytep row,
+   png_const_colorp palette, png_const_bytep trans_alpha, int num_trans)
+{
+   int shift, value;
+   png_bytep sp, dp;
+   png_uint_32 i;
+   png_uint_32 row_width=row_info->width;
+
+   png_debug(1, "in png_do_expand_palette");
+
+   if (row_info->color_type == PNG_COLOR_TYPE_PALETTE)
+   {
+      if (row_info->bit_depth < 8)
+      {
+         switch (row_info->bit_depth)
+         {
+            case 1:
+            {
+               sp = row + (png_size_t)((row_width - 1) >> 3);
+               dp = row + (png_size_t)row_width - 1;
+               shift = 7 - (int)((row_width + 7) & 0x07);
+               for (i = 0; i < row_width; i++)
+               {
+                  if ((*sp >> shift) & 0x01)
+                     *dp = 1;
+
+                  else
+                     *dp = 0;
+
+                  if (shift == 7)
+                  {
+                     shift = 0;
+                     sp--;
+                  }
+
+                  else
+                     shift++;
+
+                  dp--;
+               }
+               break;
+            }
+
+            case 2:
+            {
+               sp = row + (png_size_t)((row_width - 1) >> 2);
+               dp = row + (png_size_t)row_width - 1;
+               shift = (int)((3 - ((row_width + 3) & 0x03)) << 1);
+               for (i = 0; i < row_width; i++)
+               {
+                  value = (*sp >> shift) & 0x03;
+                  *dp = (png_byte)value;
+                  if (shift == 6)
+                  {
+                     shift = 0;
+                     sp--;
+                  }
+
+                  else
+                     shift += 2;
+
+                  dp--;
+               }
+               break;
+            }
+
+            case 4:
+            {
+               sp = row + (png_size_t)((row_width - 1) >> 1);
+               dp = row + (png_size_t)row_width - 1;
+               shift = (int)((row_width & 0x01) << 2);
+               for (i = 0; i < row_width; i++)
+               {
+                  value = (*sp >> shift) & 0x0f;
+                  *dp = (png_byte)value;
+                  if (shift == 4)
+                  {
+                     shift = 0;
+                     sp--;
+                  }
+
+                  else
+                     shift += 4;
+
+                  dp--;
+               }
+               break;
+            }
+
+            default:
+               break;
+         }
+         row_info->bit_depth = 8;
+         row_info->pixel_depth = 8;
+         row_info->rowbytes = row_width;
+      }
+
+      if (row_info->bit_depth == 8)
+      {
+         {
+            if (num_trans > 0)
+            {
+               sp = row + (png_size_t)row_width - 1;
+               dp = row + (png_size_t)(row_width << 2) - 1;
+
+               for (i = 0; i < row_width; i++)
+               {
+                  if ((int)(*sp) >= num_trans)
+                     *dp-- = 0xff;
+
+                  else
+                     *dp-- = trans_alpha[*sp];
+
+                  *dp-- = palette[*sp].blue;
+                  *dp-- = palette[*sp].green;
+                  *dp-- = palette[*sp].red;
+                  sp--;
+               }
+               row_info->bit_depth = 8;
+               row_info->pixel_depth = 32;
+               row_info->rowbytes = row_width * 4;
+               row_info->color_type = 6;
+               row_info->channels = 4;
+            }
+
+            else
+            {
+               sp = row + (png_size_t)row_width - 1;
+               dp = row + (png_size_t)(row_width * 3) - 1;
+
+               for (i = 0; i < row_width; i++)
+               {
+                  *dp-- = palette[*sp].blue;
+                  *dp-- = palette[*sp].green;
+                  *dp-- = palette[*sp].red;
+                  sp--;
+               }
+
+               row_info->bit_depth = 8;
+               row_info->pixel_depth = 24;
+               row_info->rowbytes = row_width * 3;
+               row_info->color_type = 2;
+               row_info->channels = 3;
+            }
+         }
+      }
+   }
+}
+
+/* If the bit depth < 8, it is expanded to 8.  Also, if the already
+ * expanded transparency value is supplied, an alpha channel is built.
+ */
+void /* PRIVATE */
+png_do_expand(png_row_infop row_info, png_bytep row,
+    png_const_color_16p trans_color)
+{
+   int shift, value;
+   png_bytep sp, dp;
+   png_uint_32 i;
+   png_uint_32 row_width=row_info->width;
+
+   png_debug(1, "in png_do_expand");
+
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_GRAY)
+      {
+         unsigned int gray = trans_color ? trans_color->gray : 0;
+
+         if (row_info->bit_depth < 8)
+         {
+            switch (row_info->bit_depth)
+            {
+               case 1:
+               {
+                  gray = (gray & 0x01) * 0xff;
+                  sp = row + (png_size_t)((row_width - 1) >> 3);
+                  dp = row + (png_size_t)row_width - 1;
+                  shift = 7 - (int)((row_width + 7) & 0x07);
+                  for (i = 0; i < row_width; i++)
+                  {
+                     if ((*sp >> shift) & 0x01)
+                        *dp = 0xff;
+
+                     else
+                        *dp = 0;
+
+                     if (shift == 7)
+                     {
+                        shift = 0;
+                        sp--;
+                     }
+
+                     else
+                        shift++;
+
+                     dp--;
+                  }
+                  break;
+               }
+
+               case 2:
+               {
+                  gray = (gray & 0x03) * 0x55;
+                  sp = row + (png_size_t)((row_width - 1) >> 2);
+                  dp = row + (png_size_t)row_width - 1;
+                  shift = (int)((3 - ((row_width + 3) & 0x03)) << 1);
+                  for (i = 0; i < row_width; i++)
+                  {
+                     value = (*sp >> shift) & 0x03;
+                     *dp = (png_byte)(value | (value << 2) | (value << 4) |
+                        (value << 6));
+                     if (shift == 6)
+                     {
+                        shift = 0;
+                        sp--;
+                     }
+
+                     else
+                        shift += 2;
+
+                     dp--;
+                  }
+                  break;
+               }
+
+               case 4:
+               {
+                  gray = (gray & 0x0f) * 0x11;
+                  sp = row + (png_size_t)((row_width - 1) >> 1);
+                  dp = row + (png_size_t)row_width - 1;
+                  shift = (int)((1 - ((row_width + 1) & 0x01)) << 2);
+                  for (i = 0; i < row_width; i++)
+                  {
+                     value = (*sp >> shift) & 0x0f;
+                     *dp = (png_byte)(value | (value << 4));
+                     if (shift == 4)
+                     {
+                        shift = 0;
+                        sp--;
+                     }
+
+                     else
+                        shift = 4;
+
+                     dp--;
+                  }
+                  break;
+               }
+
+               default:
+                  break;
+            }
+
+            row_info->bit_depth = 8;
+            row_info->pixel_depth = 8;
+            row_info->rowbytes = row_width;
+         }
+
+         if (trans_color != NULL)
+         {
+            if (row_info->bit_depth == 8)
+            {
+               gray = gray & 0xff;
+               sp = row + (png_size_t)row_width - 1;
+               dp = row + (png_size_t)(row_width << 1) - 1;
+
+               for (i = 0; i < row_width; i++)
+               {
+                  if (*sp == gray)
+                     *dp-- = 0;
+
+                  else
+                     *dp-- = 0xff;
+
+                  *dp-- = *sp--;
+               }
+            }
+
+            else if (row_info->bit_depth == 16)
+            {
+               unsigned int gray_high = (gray >> 8) & 0xff;
+               unsigned int gray_low = gray & 0xff;
+               sp = row + row_info->rowbytes - 1;
+               dp = row + (row_info->rowbytes << 1) - 1;
+               for (i = 0; i < row_width; i++)
+               {
+                  if (*(sp - 1) == gray_high && *(sp) == gray_low)
+                  {
+                     *dp-- = 0;
+                     *dp-- = 0;
+                  }
+
+                  else
+                  {
+                     *dp-- = 0xff;
+                     *dp-- = 0xff;
+                  }
+
+                  *dp-- = *sp--;
+                  *dp-- = *sp--;
+               }
+            }
+
+            row_info->color_type = PNG_COLOR_TYPE_GRAY_ALPHA;
+            row_info->channels = 2;
+            row_info->pixel_depth = (png_byte)(row_info->bit_depth << 1);
+            row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth,
+               row_width);
+         }
+      }
+      else if (row_info->color_type == PNG_COLOR_TYPE_RGB && trans_color)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            png_byte red = (png_byte)(trans_color->red & 0xff);
+            png_byte green = (png_byte)(trans_color->green & 0xff);
+            png_byte blue = (png_byte)(trans_color->blue & 0xff);
+            sp = row + (png_size_t)row_info->rowbytes - 1;
+            dp = row + (png_size_t)(row_width << 2) - 1;
+            for (i = 0; i < row_width; i++)
+            {
+               if (*(sp - 2) == red && *(sp - 1) == green && *(sp) == blue)
+                  *dp-- = 0;
+
+               else
+                  *dp-- = 0xff;
+
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+            }
+         }
+         else if (row_info->bit_depth == 16)
+         {
+            png_byte red_high = (png_byte)((trans_color->red >> 8) & 0xff);
+            png_byte green_high = (png_byte)((trans_color->green >> 8) & 0xff);
+            png_byte blue_high = (png_byte)((trans_color->blue >> 8) & 0xff);
+            png_byte red_low = (png_byte)(trans_color->red & 0xff);
+            png_byte green_low = (png_byte)(trans_color->green & 0xff);
+            png_byte blue_low = (png_byte)(trans_color->blue & 0xff);
+            sp = row + row_info->rowbytes - 1;
+            dp = row + (png_size_t)(row_width << 3) - 1;
+            for (i = 0; i < row_width; i++)
+            {
+               if (*(sp - 5) == red_high &&
+                   *(sp - 4) == red_low &&
+                   *(sp - 3) == green_high &&
+                   *(sp - 2) == green_low &&
+                   *(sp - 1) == blue_high &&
+                   *(sp    ) == blue_low)
+               {
+                  *dp-- = 0;
+                  *dp-- = 0;
+               }
+
+               else
+               {
+                  *dp-- = 0xff;
+                  *dp-- = 0xff;
+               }
+
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+               *dp-- = *sp--;
+            }
+         }
+         row_info->color_type = PNG_COLOR_TYPE_RGB_ALPHA;
+         row_info->channels = 4;
+         row_info->pixel_depth = (png_byte)(row_info->bit_depth << 2);
+         row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_width);
+      }
+   }
+}
+#endif
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+/* If the bit depth is 8 and the color type is not a palette type expand the
+ * whole row to 16 bits.  Has no effect otherwise.
+ */
+void /* PRIVATE */
+png_do_expand_16(png_row_infop row_info, png_bytep row)
+{
+   if (row_info->bit_depth == 8 &&
+      row_info->color_type != PNG_COLOR_TYPE_PALETTE)
+   {
+      /* The row have a sequence of bytes containing [0..255] and we need
+       * to turn it into another row containing [0..65535], to do this we
+       * calculate:
+       *
+       *  (input / 255) * 65535
+       *
+       *  Which happens to be exactly input * 257 and this can be achieved
+       *  simply by byte replication in place (copying backwards).
+       */
+      png_byte *sp = row + row_info->rowbytes; /* source, last byte + 1 */
+      png_byte *dp = sp + row_info->rowbytes;  /* destination, end + 1 */
+      while (dp > sp)
+         dp[-2] = dp[-1] = *--sp, dp -= 2;
+
+      row_info->rowbytes *= 2;
+      row_info->bit_depth = 16;
+      row_info->pixel_depth = (png_byte)(row_info->channels * 16);
+   }
+}
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+void /* PRIVATE */
+png_do_quantize(png_row_infop row_info, png_bytep row,
+    png_const_bytep palette_lookup, png_const_bytep quantize_lookup)
+{
+   png_bytep sp, dp;
+   png_uint_32 i;
+   png_uint_32 row_width=row_info->width;
+
+   png_debug(1, "in png_do_quantize");
+
+   if (row_info->bit_depth == 8)
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_RGB && palette_lookup)
+      {
+         int r, g, b, p;
+         sp = row;
+         dp = row;
+         for (i = 0; i < row_width; i++)
+         {
+            r = *sp++;
+            g = *sp++;
+            b = *sp++;
+
+            /* This looks real messy, but the compiler will reduce
+             * it down to a reasonable formula.  For example, with
+             * 5 bits per color, we get:
+             * p = (((r >> 3) & 0x1f) << 10) |
+             *    (((g >> 3) & 0x1f) << 5) |
+             *    ((b >> 3) & 0x1f);
+             */
+            p = (((r >> (8 - PNG_QUANTIZE_RED_BITS)) &
+                ((1 << PNG_QUANTIZE_RED_BITS) - 1)) <<
+                (PNG_QUANTIZE_GREEN_BITS + PNG_QUANTIZE_BLUE_BITS)) |
+                (((g >> (8 - PNG_QUANTIZE_GREEN_BITS)) &
+                ((1 << PNG_QUANTIZE_GREEN_BITS) - 1)) <<
+                (PNG_QUANTIZE_BLUE_BITS)) |
+                ((b >> (8 - PNG_QUANTIZE_BLUE_BITS)) &
+                ((1 << PNG_QUANTIZE_BLUE_BITS) - 1));
+
+            *dp++ = palette_lookup[p];
+         }
+
+         row_info->color_type = PNG_COLOR_TYPE_PALETTE;
+         row_info->channels = 1;
+         row_info->pixel_depth = row_info->bit_depth;
+         row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_width);
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA &&
+         palette_lookup != NULL)
+      {
+         int r, g, b, p;
+         sp = row;
+         dp = row;
+         for (i = 0; i < row_width; i++)
+         {
+            r = *sp++;
+            g = *sp++;
+            b = *sp++;
+            sp++;
+
+            p = (((r >> (8 - PNG_QUANTIZE_RED_BITS)) &
+                ((1 << PNG_QUANTIZE_RED_BITS) - 1)) <<
+                (PNG_QUANTIZE_GREEN_BITS + PNG_QUANTIZE_BLUE_BITS)) |
+                (((g >> (8 - PNG_QUANTIZE_GREEN_BITS)) &
+                ((1 << PNG_QUANTIZE_GREEN_BITS) - 1)) <<
+                (PNG_QUANTIZE_BLUE_BITS)) |
+                ((b >> (8 - PNG_QUANTIZE_BLUE_BITS)) &
+                ((1 << PNG_QUANTIZE_BLUE_BITS) - 1));
+
+            *dp++ = palette_lookup[p];
+         }
+
+         row_info->color_type = PNG_COLOR_TYPE_PALETTE;
+         row_info->channels = 1;
+         row_info->pixel_depth = row_info->bit_depth;
+         row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, row_width);
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_PALETTE &&
+         quantize_lookup)
+      {
+         sp = row;
+
+         for (i = 0; i < row_width; i++, sp++)
+         {
+            *sp = quantize_lookup[*sp];
+         }
+      }
+   }
+}
+#endif /* PNG_READ_QUANTIZE_SUPPORTED */
+#endif /* PNG_READ_TRANSFORMS_SUPPORTED */
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+/* Undoes intrapixel differencing  */
+void /* PRIVATE */
+png_do_read_intrapixel(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_read_intrapixel");
+
+   if (
+       (row_info->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      int bytes_per_pixel;
+      png_uint_32 row_width = row_info->width;
+
+      if (row_info->bit_depth == 8)
+      {
+         png_bytep rp;
+         png_uint_32 i;
+
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+            bytes_per_pixel = 3;
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+            bytes_per_pixel = 4;
+
+         else
+            return;
+
+         for (i = 0, rp = row; i < row_width; i++, rp += bytes_per_pixel)
+         {
+            *(rp) = (png_byte)((256 + *rp + *(rp + 1)) & 0xff);
+            *(rp+2) = (png_byte)((256 + *(rp + 2) + *(rp + 1)) & 0xff);
+         }
+      }
+      else if (row_info->bit_depth == 16)
+      {
+         png_bytep rp;
+         png_uint_32 i;
+
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+            bytes_per_pixel = 6;
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+            bytes_per_pixel = 8;
+
+         else
+            return;
+
+         for (i = 0, rp = row; i < row_width; i++, rp += bytes_per_pixel)
+         {
+            png_uint_32 s0   = (*(rp    ) << 8) | *(rp + 1);
+            png_uint_32 s1   = (*(rp + 2) << 8) | *(rp + 3);
+            png_uint_32 s2   = (*(rp + 4) << 8) | *(rp + 5);
+            png_uint_32 red  = (s0 + s1 + 65536) & 0xffff;
+            png_uint_32 blue = (s2 + s1 + 65536) & 0xffff;
+            *(rp    ) = (png_byte)((red >> 8) & 0xff);
+            *(rp + 1) = (png_byte)(red & 0xff);
+            *(rp + 4) = (png_byte)((blue >> 8) & 0xff);
+            *(rp + 5) = (png_byte)(blue & 0xff);
+         }
+      }
+   }
+}
+#endif /* PNG_MNG_FEATURES_SUPPORTED */
+#endif /* PNG_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngrutil.c b/ml/dlib/dlib/external/libpng/pngrutil.c
new file mode 100644
index 000000000..2438384dd
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngrutil.c
@@ -0,0 +1,4475 @@
+
+/* pngrutil.c - utilities to read a PNG file
+ *
+ * Last changed in libpng 1.6.7 [November 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file contains routines that are only called from within
+ * libpng itself during the course of reading an image.
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+png_uint_32 PNGAPI
+png_get_uint_31(png_const_structrp png_ptr, png_const_bytep buf)
+{
+   png_uint_32 uval = png_get_uint_32(buf);
+
+   if (uval > PNG_UINT_31_MAX)
+      png_error(png_ptr, "PNG unsigned integer out of range");
+
+   return (uval);
+}
+
+#if defined(PNG_READ_gAMA_SUPPORTED) || defined(PNG_READ_cHRM_SUPPORTED)
+/* The following is a variation on the above for use with the fixed
+ * point values used for gAMA and cHRM.  Instead of png_error it
+ * issues a warning and returns (-1) - an invalid value because both
+ * gAMA and cHRM use *unsigned* integers for fixed point values.
+ */
+#define PNG_FIXED_ERROR (-1)
+
+static png_fixed_point /* PRIVATE */
+png_get_fixed_point(png_structrp png_ptr, png_const_bytep buf)
+{
+   png_uint_32 uval = png_get_uint_32(buf);
+
+   if (uval <= PNG_UINT_31_MAX)
+      return (png_fixed_point)uval; /* known to be in range */
+
+   /* The caller can turn off the warning by passing NULL. */
+   if (png_ptr != NULL)
+      png_warning(png_ptr, "PNG fixed point integer out of range");
+
+   return PNG_FIXED_ERROR;
+}
+#endif
+
+#ifdef PNG_READ_INT_FUNCTIONS_SUPPORTED
+/* NOTE: the read macros will obscure these definitions, so that if
+ * PNG_USE_READ_MACROS is set the library will not use them internally,
+ * but the APIs will still be available externally.
+ *
+ * The parentheses around "PNGAPI function_name" in the following three
+ * functions are necessary because they allow the macros to co-exist with
+ * these (unused but exported) functions.
+ */
+
+/* Grab an unsigned 32-bit integer from a buffer in big-endian format. */
+png_uint_32 (PNGAPI
+png_get_uint_32)(png_const_bytep buf)
+{
+   png_uint_32 uval =
+       ((png_uint_32)(*(buf    )) << 24) +
+       ((png_uint_32)(*(buf + 1)) << 16) +
+       ((png_uint_32)(*(buf + 2)) <<  8) +
+       ((png_uint_32)(*(buf + 3))      ) ;
+
+   return uval;
+}
+
+/* Grab a signed 32-bit integer from a buffer in big-endian format.  The
+ * data is stored in the PNG file in two's complement format and there
+ * is no guarantee that a 'png_int_32' is exactly 32 bits, therefore
+ * the following code does a two's complement to native conversion.
+ */
+png_int_32 (PNGAPI
+png_get_int_32)(png_const_bytep buf)
+{
+   png_uint_32 uval = png_get_uint_32(buf);
+   if ((uval & 0x80000000) == 0) /* non-negative */
+      return uval;
+
+   uval = (uval ^ 0xffffffff) + 1;  /* 2's complement: -x = ~x+1 */
+   return -(png_int_32)uval;
+}
+
+/* Grab an unsigned 16-bit integer from a buffer in big-endian format. */
+png_uint_16 (PNGAPI
+png_get_uint_16)(png_const_bytep buf)
+{
+   /* ANSI-C requires an int value to accomodate at least 16 bits so this
+    * works and allows the compiler not to worry about possible narrowing
+    * on 32 bit systems.  (Pre-ANSI systems did not make integers smaller
+    * than 16 bits either.)
+    */
+   unsigned int val =
+       ((unsigned int)(*buf) << 8) +
+       ((unsigned int)(*(buf + 1)));
+
+   return (png_uint_16)val;
+}
+
+#endif /* PNG_READ_INT_FUNCTIONS_SUPPORTED */
+
+/* Read and check the PNG file signature */
+void /* PRIVATE */
+png_read_sig(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_size_t num_checked, num_to_check;
+
+   /* Exit if the user application does not expect a signature. */
+   if (png_ptr->sig_bytes >= 8)
+      return;
+
+   num_checked = png_ptr->sig_bytes;
+   num_to_check = 8 - num_checked;
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   png_ptr->io_state = PNG_IO_READING | PNG_IO_SIGNATURE;
+#endif
+
+   /* The signature must be serialized in a single I/O call. */
+   png_read_data(png_ptr, &(info_ptr->signature[num_checked]), num_to_check);
+   png_ptr->sig_bytes = 8;
+
+   if (png_sig_cmp(info_ptr->signature, num_checked, num_to_check))
+   {
+      if (num_checked < 4 &&
+          png_sig_cmp(info_ptr->signature, num_checked, num_to_check - 4))
+         png_error(png_ptr, "Not a PNG file");
+      else
+         png_error(png_ptr, "PNG file corrupted by ASCII conversion");
+   }
+   if (num_checked < 3)
+      png_ptr->mode |= PNG_HAVE_PNG_SIGNATURE;
+}
+
+/* Read the chunk header (length + type name).
+ * Put the type name into png_ptr->chunk_name, and return the length.
+ */
+png_uint_32 /* PRIVATE */
+png_read_chunk_header(png_structrp png_ptr)
+{
+   png_byte buf[8];
+   png_uint_32 length;
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   png_ptr->io_state = PNG_IO_READING | PNG_IO_CHUNK_HDR;
+#endif
+
+   /* Read the length and the chunk name.
+    * This must be performed in a single I/O call.
+    */
+   png_read_data(png_ptr, buf, 8);
+   length = png_get_uint_31(png_ptr, buf);
+
+   /* Put the chunk name into png_ptr->chunk_name. */
+   png_ptr->chunk_name = PNG_CHUNK_FROM_STRING(buf+4);
+
+   png_debug2(0, "Reading %lx chunk, length = %lu",
+       (unsigned long)png_ptr->chunk_name, (unsigned long)length);
+
+   /* Reset the crc and run it over the chunk name. */
+   png_reset_crc(png_ptr);
+   png_calculate_crc(png_ptr, buf + 4, 4);
+
+   /* Check to see if chunk name is valid. */
+   png_check_chunk_name(png_ptr, png_ptr->chunk_name);
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   png_ptr->io_state = PNG_IO_READING | PNG_IO_CHUNK_DATA;
+#endif
+
+   return length;
+}
+
+/* Read data, and (optionally) run it through the CRC. */
+void /* PRIVATE */
+png_crc_read(png_structrp png_ptr, png_bytep buf, png_uint_32 length)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_read_data(png_ptr, buf, length);
+   png_calculate_crc(png_ptr, buf, length);
+}
+
+/* Optionally skip data and then check the CRC.  Depending on whether we
+ * are reading an ancillary or critical chunk, and how the program has set
+ * things up, we may calculate the CRC on the data and print a message.
+ * Returns '1' if there was a CRC error, '0' otherwise.
+ */
+int /* PRIVATE */
+png_crc_finish(png_structrp png_ptr, png_uint_32 skip)
+{
+   /* The size of the local buffer for inflate is a good guess as to a
+    * reasonable size to use for buffering reads from the application.
+    */
+   while (skip > 0)
+   {
+      png_uint_32 len;
+      png_byte tmpbuf[PNG_INFLATE_BUF_SIZE];
+
+      len = (sizeof tmpbuf);
+      if (len > skip)
+         len = skip;
+      skip -= len;
+
+      png_crc_read(png_ptr, tmpbuf, len);
+   }
+
+   if (png_crc_error(png_ptr))
+   {
+      if (PNG_CHUNK_ANCILLARY(png_ptr->chunk_name) ?
+          !(png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_NOWARN) :
+          (png_ptr->flags & PNG_FLAG_CRC_CRITICAL_USE))
+      {
+         png_chunk_warning(png_ptr, "CRC error");
+      }
+
+      else
+      {
+         png_chunk_benign_error(png_ptr, "CRC error");
+         return (0);
+      }
+
+      return (1);
+   }
+
+   return (0);
+}
+
+/* Compare the CRC stored in the PNG file with that calculated by libpng from
+ * the data it has read thus far.
+ */
+int /* PRIVATE */
+png_crc_error(png_structrp png_ptr)
+{
+   png_byte crc_bytes[4];
+   png_uint_32 crc;
+   int need_crc = 1;
+
+   if (PNG_CHUNK_ANCILLARY(png_ptr->chunk_name))
+   {
+      if ((png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_MASK) ==
+          (PNG_FLAG_CRC_ANCILLARY_USE | PNG_FLAG_CRC_ANCILLARY_NOWARN))
+         need_crc = 0;
+   }
+
+   else /* critical */
+   {
+      if (png_ptr->flags & PNG_FLAG_CRC_CRITICAL_IGNORE)
+         need_crc = 0;
+   }
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   png_ptr->io_state = PNG_IO_READING | PNG_IO_CHUNK_CRC;
+#endif
+
+   /* The chunk CRC must be serialized in a single I/O call. */
+   png_read_data(png_ptr, crc_bytes, 4);
+
+   if (need_crc)
+   {
+      crc = png_get_uint_32(crc_bytes);
+      return ((int)(crc != png_ptr->crc));
+   }
+
+   else
+      return (0);
+}
+
+/* Manage the read buffer; this simply reallocates the buffer if it is not small
+ * enough (or if it is not allocated).  The routine returns a pointer to the
+ * buffer; if an error occurs and 'warn' is set the routine returns NULL, else
+ * it will call png_error (via png_malloc) on failure.  (warn == 2 means
+ * 'silent').
+ */
+static png_bytep
+png_read_buffer(png_structrp png_ptr, png_alloc_size_t new_size, int warn)
+{
+   png_bytep buffer = png_ptr->read_buffer;
+
+   if (buffer != NULL && new_size > png_ptr->read_buffer_size)
+   {
+      png_ptr->read_buffer = NULL;
+      png_ptr->read_buffer = NULL;
+      png_ptr->read_buffer_size = 0;
+      png_free(png_ptr, buffer);
+      buffer = NULL;
+   }
+
+   if (buffer == NULL)
+   {
+      buffer = png_voidcast(png_bytep, png_malloc_base(png_ptr, new_size));
+
+      if (buffer != NULL)
+      {
+         png_ptr->read_buffer = buffer;
+         png_ptr->read_buffer_size = new_size;
+      }
+
+      else if (warn < 2) /* else silent */
+      {
+#ifdef PNG_WARNINGS_SUPPORTED
+         if (warn)
+             png_chunk_warning(png_ptr, "insufficient memory to read chunk");
+         else
+#endif
+         {
+#ifdef PNG_ERROR_TEXT_SUPPORTED
+             png_chunk_error(png_ptr, "insufficient memory to read chunk");
+#endif
+         }
+      }
+   }
+
+   return buffer;
+}
+
+/* png_inflate_claim: claim the zstream for some nefarious purpose that involves
+ * decompression.  Returns Z_OK on success, else a zlib error code.  It checks
+ * the owner but, in final release builds, just issues a warning if some other
+ * chunk apparently owns the stream.  Prior to release it does a png_error.
+ */
+static int
+png_inflate_claim(png_structrp png_ptr, png_uint_32 owner)
+{
+   if (png_ptr->zowner != 0)
+   {
+      char msg[64];
+
+      PNG_STRING_FROM_CHUNK(msg, png_ptr->zowner);
+      /* So the message that results is "<chunk> using zstream"; this is an
+       * internal error, but is very useful for debugging.  i18n requirements
+       * are minimal.
+       */
+      (void)png_safecat(msg, (sizeof msg), 4, " using zstream");
+#     if PNG_LIBPNG_BUILD_BASE_TYPE >= PNG_LIBPNG_BUILD_RC
+         png_chunk_warning(png_ptr, msg);
+         png_ptr->zowner = 0;
+#     else
+         png_chunk_error(png_ptr, msg);
+#     endif
+   }
+
+   /* Implementation note: unlike 'png_deflate_claim' this internal function
+    * does not take the size of the data as an argument.  Some efficiency could
+    * be gained by using this when it is known *if* the zlib stream itself does
+    * not record the number; however, this is an illusion: the original writer
+    * of the PNG may have selected a lower window size, and we really must
+    * follow that because, for systems with with limited capabilities, we
+    * would otherwise reject the application's attempts to use a smaller window
+    * size (zlib doesn't have an interface to say "this or lower"!).
+    *
+    * inflateReset2 was added to zlib 1.2.4; before this the window could not be
+    * reset, therefore it is necessary to always allocate the maximum window
+    * size with earlier zlibs just in case later compressed chunks need it.
+    */
+   {
+      int ret; /* zlib return code */
+#     if PNG_ZLIB_VERNUM >= 0x1240
+
+#        if defined(PNG_SET_OPTION_SUPPORTED) && \
+            defined(PNG_MAXIMUM_INFLATE_WINDOW)
+            int window_bits;
+
+            if (((png_ptr->options >> PNG_MAXIMUM_INFLATE_WINDOW) & 3) ==
+               PNG_OPTION_ON)
+               window_bits = 15;
+
+            else
+               window_bits = 0;
+#        else
+#           define window_bits 0
+#        endif
+#     endif
+
+      /* Set this for safety, just in case the previous owner left pointers to
+       * memory allocations.
+       */
+      png_ptr->zstream.next_in = NULL;
+      png_ptr->zstream.avail_in = 0;
+      png_ptr->zstream.next_out = NULL;
+      png_ptr->zstream.avail_out = 0;
+
+      if (png_ptr->flags & PNG_FLAG_ZSTREAM_INITIALIZED)
+      {
+#        if PNG_ZLIB_VERNUM < 0x1240
+            ret = inflateReset(&png_ptr->zstream);
+#        else
+            ret = inflateReset2(&png_ptr->zstream, window_bits);
+#        endif
+      }
+
+      else
+      {
+#        if PNG_ZLIB_VERNUM < 0x1240
+            ret = inflateInit(&png_ptr->zstream);
+#        else
+            ret = inflateInit2(&png_ptr->zstream, window_bits);
+#        endif
+
+         if (ret == Z_OK)
+            png_ptr->flags |= PNG_FLAG_ZSTREAM_INITIALIZED;
+      }
+
+      if (ret == Z_OK)
+         png_ptr->zowner = owner;
+
+      else
+         png_zstream_error(png_ptr, ret);
+
+      return ret;
+   }
+
+#  ifdef window_bits
+#     undef window_bits
+#  endif
+}
+
+#ifdef PNG_READ_COMPRESSED_TEXT_SUPPORTED
+/* png_inflate now returns zlib error codes including Z_OK and Z_STREAM_END to
+ * allow the caller to do multiple calls if required.  If the 'finish' flag is
+ * set Z_FINISH will be passed to the final inflate() call and Z_STREAM_END must
+ * be returned or there has been a problem, otherwise Z_SYNC_FLUSH is used and
+ * Z_OK or Z_STREAM_END will be returned on success.
+ *
+ * The input and output sizes are updated to the actual amounts of data consumed
+ * or written, not the amount available (as in a z_stream).  The data pointers
+ * are not changed, so the next input is (data+input_size) and the next
+ * available output is (output+output_size).
+ */
+static int
+png_inflate(png_structrp png_ptr, png_uint_32 owner, int finish,
+    /* INPUT: */ png_const_bytep input, png_uint_32p input_size_ptr,
+    /* OUTPUT: */ png_bytep output, png_alloc_size_t *output_size_ptr)
+{
+   if (png_ptr->zowner == owner) /* Else not claimed */
+   {
+      int ret;
+      png_alloc_size_t avail_out = *output_size_ptr;
+      png_uint_32 avail_in = *input_size_ptr;
+
+      /* zlib can't necessarily handle more than 65535 bytes at once (i.e. it
+       * can't even necessarily handle 65536 bytes) because the type uInt is
+       * "16 bits or more".  Consequently it is necessary to chunk the input to
+       * zlib.  This code uses ZLIB_IO_MAX, from pngpriv.h, as the maximum (the
+       * maximum value that can be stored in a uInt.)  It is possible to set
+       * ZLIB_IO_MAX to a lower value in pngpriv.h and this may sometimes have
+       * a performance advantage, because it reduces the amount of data accessed
+       * at each step and that may give the OS more time to page it in.
+       */
+      png_ptr->zstream.next_in = PNGZ_INPUT_CAST(input);
+      /* avail_in and avail_out are set below from 'size' */
+      png_ptr->zstream.avail_in = 0;
+      png_ptr->zstream.avail_out = 0;
+
+      /* Read directly into the output if it is available (this is set to
+       * a local buffer below if output is NULL).
+       */
+      if (output != NULL)
+         png_ptr->zstream.next_out = output;
+
+      do
+      {
+         uInt avail;
+         Byte local_buffer[PNG_INFLATE_BUF_SIZE];
+
+         /* zlib INPUT BUFFER */
+         /* The setting of 'avail_in' used to be outside the loop; by setting it
+          * inside it is possible to chunk the input to zlib and simply rely on
+          * zlib to advance the 'next_in' pointer.  This allows arbitrary
+          * amounts of data to be passed through zlib at the unavoidable cost of
+          * requiring a window save (memcpy of up to 32768 output bytes)
+          * every ZLIB_IO_MAX input bytes.
+          */
+         avail_in += png_ptr->zstream.avail_in; /* not consumed last time */
+
+         avail = ZLIB_IO_MAX;
+
+         if (avail_in < avail)
+            avail = (uInt)avail_in; /* safe: < than ZLIB_IO_MAX */
+
+         avail_in -= avail;
+         png_ptr->zstream.avail_in = avail;
+
+         /* zlib OUTPUT BUFFER */
+         avail_out += png_ptr->zstream.avail_out; /* not written last time */
+
+         avail = ZLIB_IO_MAX; /* maximum zlib can process */
+
+         if (output == NULL)
+         {
+            /* Reset the output buffer each time round if output is NULL and
+             * make available the full buffer, up to 'remaining_space'
+             */
+            png_ptr->zstream.next_out = local_buffer;
+            if ((sizeof local_buffer) < avail)
+               avail = (sizeof local_buffer);
+         }
+
+         if (avail_out < avail)
+            avail = (uInt)avail_out; /* safe: < ZLIB_IO_MAX */
+
+         png_ptr->zstream.avail_out = avail;
+         avail_out -= avail;
+
+         /* zlib inflate call */
+         /* In fact 'avail_out' may be 0 at this point, that happens at the end
+          * of the read when the final LZ end code was not passed at the end of
+          * the previous chunk of input data.  Tell zlib if we have reached the
+          * end of the output buffer.
+          */
+         ret = inflate(&png_ptr->zstream, avail_out > 0 ? Z_NO_FLUSH :
+            (finish ? Z_FINISH : Z_SYNC_FLUSH));
+      } while (ret == Z_OK);
+
+      /* For safety kill the local buffer pointer now */
+      if (output == NULL)
+         png_ptr->zstream.next_out = NULL;
+
+      /* Claw back the 'size' and 'remaining_space' byte counts. */
+      avail_in += png_ptr->zstream.avail_in;
+      avail_out += png_ptr->zstream.avail_out;
+
+      /* Update the input and output sizes; the updated values are the amount
+       * consumed or written, effectively the inverse of what zlib uses.
+       */
+      if (avail_out > 0)
+         *output_size_ptr -= avail_out;
+
+      if (avail_in > 0)
+         *input_size_ptr -= avail_in;
+
+      /* Ensure png_ptr->zstream.msg is set (even in the success case!) */
+      png_zstream_error(png_ptr, ret);
+      return ret;
+   }
+
+   else
+   {
+      /* This is a bad internal error.  The recovery assigns to the zstream msg
+       * pointer, which is not owned by the caller, but this is safe; it's only
+       * used on errors!
+       */
+      png_ptr->zstream.msg = PNGZ_MSG_CAST("zstream unclaimed");
+      return Z_STREAM_ERROR;
+   }
+}
+
+/*
+ * Decompress trailing data in a chunk.  The assumption is that read_buffer
+ * points at an allocated area holding the contents of a chunk with a
+ * trailing compressed part.  What we get back is an allocated area
+ * holding the original prefix part and an uncompressed version of the
+ * trailing part (the malloc area passed in is freed).
+ */
+static int
+png_decompress_chunk(png_structrp png_ptr,
+   png_uint_32 chunklength, png_uint_32 prefix_size,
+   png_alloc_size_t *newlength /* must be initialized to the maximum! */,
+   int terminate /*add a '\0' to the end of the uncompressed data*/)
+{
+   /* TODO: implement different limits for different types of chunk.
+    *
+    * The caller supplies *newlength set to the maximum length of the
+    * uncompressed data, but this routine allocates space for the prefix and
+    * maybe a '\0' terminator too.  We have to assume that 'prefix_size' is
+    * limited only by the maximum chunk size.
+    */
+   png_alloc_size_t limit = PNG_SIZE_MAX;
+
+#  ifdef PNG_SET_CHUNK_MALLOC_LIMIT_SUPPORTED
+      if (png_ptr->user_chunk_malloc_max > 0 &&
+         png_ptr->user_chunk_malloc_max < limit)
+         limit = png_ptr->user_chunk_malloc_max;
+#  elif PNG_USER_CHUNK_MALLOC_MAX > 0
+      if (PNG_USER_CHUNK_MALLOC_MAX < limit)
+         limit = PNG_USER_CHUNK_MALLOC_MAX;
+#  endif
+
+   if (limit >= prefix_size + (terminate != 0))
+   {
+      int ret;
+
+      limit -= prefix_size + (terminate != 0);
+
+      if (limit < *newlength)
+         *newlength = limit;
+
+      /* Now try to claim the stream. */
+      ret = png_inflate_claim(png_ptr, png_ptr->chunk_name);
+
+      if (ret == Z_OK)
+      {
+         png_uint_32 lzsize = chunklength - prefix_size;
+
+         ret = png_inflate(png_ptr, png_ptr->chunk_name, 1/*finish*/,
+            /* input: */ png_ptr->read_buffer + prefix_size, &lzsize,
+            /* output: */ NULL, newlength);
+
+         if (ret == Z_STREAM_END)
+         {
+            /* Use 'inflateReset' here, not 'inflateReset2' because this
+             * preserves the previously decided window size (otherwise it would
+             * be necessary to store the previous window size.)  In practice
+             * this doesn't matter anyway, because png_inflate will call inflate
+             * with Z_FINISH in almost all cases, so the window will not be
+             * maintained.
+             */
+            if (inflateReset(&png_ptr->zstream) == Z_OK)
+            {
+               /* Because of the limit checks above we know that the new,
+                * expanded, size will fit in a size_t (let alone an
+                * png_alloc_size_t).  Use png_malloc_base here to avoid an
+                * extra OOM message.
+                */
+               png_alloc_size_t new_size = *newlength;
+               png_alloc_size_t buffer_size = prefix_size + new_size +
+                  (terminate != 0);
+               png_bytep text = png_voidcast(png_bytep, png_malloc_base(png_ptr,
+                  buffer_size));
+
+               if (text != NULL)
+               {
+                  ret = png_inflate(png_ptr, png_ptr->chunk_name, 1/*finish*/,
+                     png_ptr->read_buffer + prefix_size, &lzsize,
+                     text + prefix_size, newlength);
+
+                  if (ret == Z_STREAM_END)
+                  {
+                     if (new_size == *newlength)
+                     {
+                        if (terminate)
+                           text[prefix_size + *newlength] = 0;
+
+                        if (prefix_size > 0)
+                           memcpy(text, png_ptr->read_buffer, prefix_size);
+
+                        {
+                           png_bytep old_ptr = png_ptr->read_buffer;
+
+                           png_ptr->read_buffer = text;
+                           png_ptr->read_buffer_size = buffer_size;
+                           text = old_ptr; /* freed below */
+                        }
+                     }
+
+                     else
+                     {
+                        /* The size changed on the second read, there can be no
+                         * guarantee that anything is correct at this point.
+                         * The 'msg' pointer has been set to "unexpected end of
+                         * LZ stream", which is fine, but return an error code
+                         * that the caller won't accept.
+                         */
+                        ret = PNG_UNEXPECTED_ZLIB_RETURN;
+                     }
+                  }
+
+                  else if (ret == Z_OK)
+                     ret = PNG_UNEXPECTED_ZLIB_RETURN; /* for safety */
+
+                  /* Free the text pointer (this is the old read_buffer on
+                   * success)
+                   */
+                  png_free(png_ptr, text);
+
+                  /* This really is very benign, but it's still an error because
+                   * the extra space may otherwise be used as a Trojan Horse.
+                   */
+                  if (ret == Z_STREAM_END &&
+                     chunklength - prefix_size != lzsize)
+                     png_chunk_benign_error(png_ptr, "extra compressed data");
+               }
+
+               else
+               {
+                  /* Out of memory allocating the buffer */
+                  ret = Z_MEM_ERROR;
+                  png_zstream_error(png_ptr, Z_MEM_ERROR);
+               }
+            }
+
+            else
+            {
+               /* inflateReset failed, store the error message */
+               png_zstream_error(png_ptr, ret);
+
+               if (ret == Z_STREAM_END)
+                  ret = PNG_UNEXPECTED_ZLIB_RETURN;
+            }
+         }
+
+         else if (ret == Z_OK)
+            ret = PNG_UNEXPECTED_ZLIB_RETURN;
+
+         /* Release the claimed stream */
+         png_ptr->zowner = 0;
+      }
+
+      else /* the claim failed */ if (ret == Z_STREAM_END) /* impossible! */
+         ret = PNG_UNEXPECTED_ZLIB_RETURN;
+
+      return ret;
+   }
+
+   else
+   {
+      /* Application/configuration limits exceeded */
+      png_zstream_error(png_ptr, Z_MEM_ERROR);
+      return Z_MEM_ERROR;
+   }
+}
+#endif /* PNG_READ_COMPRESSED_TEXT_SUPPORTED */
+
+#ifdef PNG_READ_iCCP_SUPPORTED
+/* Perform a partial read and decompress, producing 'avail_out' bytes and
+ * reading from the current chunk as required.
+ */
+static int
+png_inflate_read(png_structrp png_ptr, png_bytep read_buffer, uInt read_size,
+   png_uint_32p chunk_bytes, png_bytep next_out, png_alloc_size_t *out_size,
+   int finish)
+{
+   if (png_ptr->zowner == png_ptr->chunk_name)
+   {
+      int ret;
+
+      /* next_in and avail_in must have been initialized by the caller. */
+      png_ptr->zstream.next_out = next_out;
+      png_ptr->zstream.avail_out = 0; /* set in the loop */
+
+      do
+      {
+         if (png_ptr->zstream.avail_in == 0)
+         {
+            if (read_size > *chunk_bytes)
+               read_size = (uInt)*chunk_bytes;
+            *chunk_bytes -= read_size;
+
+            if (read_size > 0)
+               png_crc_read(png_ptr, read_buffer, read_size);
+
+            png_ptr->zstream.next_in = read_buffer;
+            png_ptr->zstream.avail_in = read_size;
+         }
+
+         if (png_ptr->zstream.avail_out == 0)
+         {
+            uInt avail = ZLIB_IO_MAX;
+            if (avail > *out_size)
+               avail = (uInt)*out_size;
+            *out_size -= avail;
+
+            png_ptr->zstream.avail_out = avail;
+         }
+
+         /* Use Z_SYNC_FLUSH when there is no more chunk data to ensure that all
+          * the available output is produced; this allows reading of truncated
+          * streams.
+          */
+         ret = inflate(&png_ptr->zstream,
+            *chunk_bytes > 0 ? Z_NO_FLUSH : (finish ? Z_FINISH : Z_SYNC_FLUSH));
+      }
+      while (ret == Z_OK && (*out_size > 0 || png_ptr->zstream.avail_out > 0));
+
+      *out_size += png_ptr->zstream.avail_out;
+      png_ptr->zstream.avail_out = 0; /* Should not be required, but is safe */
+
+      /* Ensure the error message pointer is always set: */
+      png_zstream_error(png_ptr, ret);
+      return ret;
+   }
+
+   else
+   {
+      png_ptr->zstream.msg = PNGZ_MSG_CAST("zstream unclaimed");
+      return Z_STREAM_ERROR;
+   }
+}
+#endif
+
+/* Read and check the IDHR chunk */
+void /* PRIVATE */
+png_handle_IHDR(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte buf[13];
+   png_uint_32 width, height;
+   int bit_depth, color_type, compression_type, filter_type;
+   int interlace_type;
+
+   png_debug(1, "in png_handle_IHDR");
+
+   if (png_ptr->mode & PNG_HAVE_IHDR)
+      png_chunk_error(png_ptr, "out of place");
+
+   /* Check the length */
+   if (length != 13)
+      png_chunk_error(png_ptr, "invalid");
+
+   png_ptr->mode |= PNG_HAVE_IHDR;
+
+   png_crc_read(png_ptr, buf, 13);
+   png_crc_finish(png_ptr, 0);
+
+   width = png_get_uint_31(png_ptr, buf);
+   height = png_get_uint_31(png_ptr, buf + 4);
+   bit_depth = buf[8];
+   color_type = buf[9];
+   compression_type = buf[10];
+   filter_type = buf[11];
+   interlace_type = buf[12];
+
+   /* Set internal variables */
+   png_ptr->width = width;
+   png_ptr->height = height;
+   png_ptr->bit_depth = (png_byte)bit_depth;
+   png_ptr->interlaced = (png_byte)interlace_type;
+   png_ptr->color_type = (png_byte)color_type;
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   png_ptr->filter_type = (png_byte)filter_type;
+#endif
+   png_ptr->compression_type = (png_byte)compression_type;
+
+   /* Find number of channels */
+   switch (png_ptr->color_type)
+   {
+      default: /* invalid, png_set_IHDR calls png_error */
+      case PNG_COLOR_TYPE_GRAY:
+      case PNG_COLOR_TYPE_PALETTE:
+         png_ptr->channels = 1;
+         break;
+
+      case PNG_COLOR_TYPE_RGB:
+         png_ptr->channels = 3;
+         break;
+
+      case PNG_COLOR_TYPE_GRAY_ALPHA:
+         png_ptr->channels = 2;
+         break;
+
+      case PNG_COLOR_TYPE_RGB_ALPHA:
+         png_ptr->channels = 4;
+         break;
+   }
+
+   /* Set up other useful info */
+   png_ptr->pixel_depth = (png_byte)(png_ptr->bit_depth *
+   png_ptr->channels);
+   png_ptr->rowbytes = PNG_ROWBYTES(png_ptr->pixel_depth, png_ptr->width);
+   png_debug1(3, "bit_depth = %d", png_ptr->bit_depth);
+   png_debug1(3, "channels = %d", png_ptr->channels);
+   png_debug1(3, "rowbytes = %lu", (unsigned long)png_ptr->rowbytes);
+   png_set_IHDR(png_ptr, info_ptr, width, height, bit_depth,
+       color_type, interlace_type, compression_type, filter_type);
+}
+
+/* Read and check the palette */
+void /* PRIVATE */
+png_handle_PLTE(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_color palette[PNG_MAX_PALETTE_LENGTH];
+   int num, i;
+#ifdef PNG_POINTER_INDEXING_SUPPORTED
+   png_colorp pal_ptr;
+#endif
+
+   png_debug(1, "in png_handle_PLTE");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   /* Moved to before the 'after IDAT' check below because otherwise duplicate
+    * PLTE chunks are potentially ignored (the spec says there shall not be more
+    * than one PLTE, the error is not treated as benign, so this check trumps
+    * the requirement that PLTE appears before IDAT.)
+    */
+   else if (png_ptr->mode & PNG_HAVE_PLTE)
+      png_chunk_error(png_ptr, "duplicate");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      /* This is benign because the non-benign error happened before, when an
+       * IDAT was encountered in a color-mapped image with no PLTE.
+       */
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   png_ptr->mode |= PNG_HAVE_PLTE;
+
+   if (!(png_ptr->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "ignored in grayscale PNG");
+      return;
+   }
+
+#ifndef PNG_READ_OPT_PLTE_SUPPORTED
+   if (png_ptr->color_type != PNG_COLOR_TYPE_PALETTE)
+   {
+      png_crc_finish(png_ptr, length);
+      return;
+   }
+#endif
+
+   if (length > 3*PNG_MAX_PALETTE_LENGTH || length % 3)
+   {
+      png_crc_finish(png_ptr, length);
+
+      if (png_ptr->color_type != PNG_COLOR_TYPE_PALETTE)
+         png_chunk_benign_error(png_ptr, "invalid");
+
+      else
+         png_chunk_error(png_ptr, "invalid");
+
+      return;
+   }
+
+   /* The cast is safe because 'length' is less than 3*PNG_MAX_PALETTE_LENGTH */
+   num = (int)length / 3;
+
+#ifdef PNG_POINTER_INDEXING_SUPPORTED
+   for (i = 0, pal_ptr = palette; i < num; i++, pal_ptr++)
+   {
+      png_byte buf[3];
+
+      png_crc_read(png_ptr, buf, 3);
+      pal_ptr->red = buf[0];
+      pal_ptr->green = buf[1];
+      pal_ptr->blue = buf[2];
+   }
+#else
+   for (i = 0; i < num; i++)
+   {
+      png_byte buf[3];
+
+      png_crc_read(png_ptr, buf, 3);
+      /* Don't depend upon png_color being any order */
+      palette[i].red = buf[0];
+      palette[i].green = buf[1];
+      palette[i].blue = buf[2];
+   }
+#endif
+
+   /* If we actually need the PLTE chunk (ie for a paletted image), we do
+    * whatever the normal CRC configuration tells us.  However, if we
+    * have an RGB image, the PLTE can be considered ancillary, so
+    * we will act as though it is.
+    */
+#ifndef PNG_READ_OPT_PLTE_SUPPORTED
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+#endif
+   {
+      png_crc_finish(png_ptr, 0);
+   }
+
+#ifndef PNG_READ_OPT_PLTE_SUPPORTED
+   else if (png_crc_error(png_ptr))  /* Only if we have a CRC error */
+   {
+      /* If we don't want to use the data from an ancillary chunk,
+       * we have two options: an error abort, or a warning and we
+       * ignore the data in this chunk (which should be OK, since
+       * it's considered ancillary for a RGB or RGBA image).
+       *
+       * IMPLEMENTATION NOTE: this is only here because png_crc_finish uses the
+       * chunk type to determine whether to check the ancillary or the critical
+       * flags.
+       */
+      if (!(png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_USE))
+      {
+         if (png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_NOWARN)
+         {
+            png_chunk_benign_error(png_ptr, "CRC error");
+         }
+
+         else
+         {
+            png_chunk_warning(png_ptr, "CRC error");
+            return;
+         }
+      }
+
+      /* Otherwise, we (optionally) emit a warning and use the chunk. */
+      else if (!(png_ptr->flags & PNG_FLAG_CRC_ANCILLARY_NOWARN))
+      {
+         png_chunk_warning(png_ptr, "CRC error");
+      }
+   }
+#endif
+
+   /* TODO: png_set_PLTE has the side effect of setting png_ptr->palette to its
+    * own copy of the palette.  This has the side effect that when png_start_row
+    * is called (this happens after any call to png_read_update_info) the
+    * info_ptr palette gets changed.  This is extremely unexpected and
+    * confusing.
+    *
+    * Fix this by not sharing the palette in this way.
+    */
+   png_set_PLTE(png_ptr, info_ptr, palette, num);
+
+   /* The three chunks, bKGD, hIST and tRNS *must* appear after PLTE and before
+    * IDAT.  Prior to 1.6.0 this was not checked; instead the code merely
+    * checked the apparent validity of a tRNS chunk inserted before PLTE on a
+    * palette PNG.  1.6.0 attempts to rigorously follow the standard and
+    * therefore does a benign error if the erroneous condition is detected *and*
+    * cancels the tRNS if the benign error returns.  The alternative is to
+    * amend the standard since it would be rather hypocritical of the standards
+    * maintainers to ignore it.
+    */
+#ifdef PNG_READ_tRNS_SUPPORTED
+   if (png_ptr->num_trans > 0 ||
+      (info_ptr != NULL && (info_ptr->valid & PNG_INFO_tRNS) != 0))
+   {
+      /* Cancel this because otherwise it would be used if the transforms
+       * require it.  Don't cancel the 'valid' flag because this would prevent
+       * detection of duplicate chunks.
+       */
+      png_ptr->num_trans = 0;
+
+      if (info_ptr != NULL)
+         info_ptr->num_trans = 0;
+
+      png_chunk_benign_error(png_ptr, "tRNS must be after");
+   }
+#endif
+
+#ifdef PNG_READ_hIST_SUPPORTED
+   if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_hIST) != 0)
+      png_chunk_benign_error(png_ptr, "hIST must be after");
+#endif
+
+#ifdef PNG_READ_bKGD_SUPPORTED
+   if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_bKGD) != 0)
+      png_chunk_benign_error(png_ptr, "bKGD must be after");
+#endif
+}
+
+void /* PRIVATE */
+png_handle_IEND(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_debug(1, "in png_handle_IEND");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR) || !(png_ptr->mode & PNG_HAVE_IDAT))
+      png_chunk_error(png_ptr, "out of place");
+
+   png_ptr->mode |= (PNG_AFTER_IDAT | PNG_HAVE_IEND);
+
+   png_crc_finish(png_ptr, length);
+
+   if (length != 0)
+      png_chunk_benign_error(png_ptr, "invalid");
+
+   PNG_UNUSED(info_ptr)
+}
+
+#ifdef PNG_READ_gAMA_SUPPORTED
+void /* PRIVATE */
+png_handle_gAMA(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_fixed_point igamma;
+   png_byte buf[4];
+
+   png_debug(1, "in png_handle_gAMA");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & (PNG_HAVE_IDAT|PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   if (length != 4)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, 4);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   igamma = png_get_fixed_point(NULL, buf);
+
+   png_colorspace_set_gamma(png_ptr, &png_ptr->colorspace, igamma);
+   png_colorspace_sync(png_ptr, info_ptr);
+}
+#endif
+
+#ifdef PNG_READ_sBIT_SUPPORTED
+void /* PRIVATE */
+png_handle_sBIT(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   unsigned int truelen;
+   png_byte buf[4];
+
+   png_debug(1, "in png_handle_sBIT");
+
+   buf[0] = buf[1] = buf[2] = buf[3] = 0;
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & (PNG_HAVE_IDAT|PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_sBIT))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      truelen = 3;
+
+   else
+      truelen = png_ptr->channels;
+
+   if (length != truelen || length > 4)
+   {
+      png_chunk_benign_error(png_ptr, "invalid");
+      png_crc_finish(png_ptr, length);
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, truelen);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   if (png_ptr->color_type & PNG_COLOR_MASK_COLOR)
+   {
+      png_ptr->sig_bit.red = buf[0];
+      png_ptr->sig_bit.green = buf[1];
+      png_ptr->sig_bit.blue = buf[2];
+      png_ptr->sig_bit.alpha = buf[3];
+   }
+
+   else
+   {
+      png_ptr->sig_bit.gray = buf[0];
+      png_ptr->sig_bit.red = buf[0];
+      png_ptr->sig_bit.green = buf[0];
+      png_ptr->sig_bit.blue = buf[0];
+      png_ptr->sig_bit.alpha = buf[1];
+   }
+
+   png_set_sBIT(png_ptr, info_ptr, &(png_ptr->sig_bit));
+}
+#endif
+
+#ifdef PNG_READ_cHRM_SUPPORTED
+void /* PRIVATE */
+png_handle_cHRM(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte buf[32];
+   png_xy xy;
+
+   png_debug(1, "in png_handle_cHRM");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & (PNG_HAVE_IDAT|PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   if (length != 32)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, 32);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   xy.whitex = png_get_fixed_point(NULL, buf);
+   xy.whitey = png_get_fixed_point(NULL, buf + 4);
+   xy.redx   = png_get_fixed_point(NULL, buf + 8);
+   xy.redy   = png_get_fixed_point(NULL, buf + 12);
+   xy.greenx = png_get_fixed_point(NULL, buf + 16);
+   xy.greeny = png_get_fixed_point(NULL, buf + 20);
+   xy.bluex  = png_get_fixed_point(NULL, buf + 24);
+   xy.bluey  = png_get_fixed_point(NULL, buf + 28);
+
+   if (xy.whitex == PNG_FIXED_ERROR ||
+       xy.whitey == PNG_FIXED_ERROR ||
+       xy.redx   == PNG_FIXED_ERROR ||
+       xy.redy   == PNG_FIXED_ERROR ||
+       xy.greenx == PNG_FIXED_ERROR ||
+       xy.greeny == PNG_FIXED_ERROR ||
+       xy.bluex  == PNG_FIXED_ERROR ||
+       xy.bluey  == PNG_FIXED_ERROR)
+   {
+      png_chunk_benign_error(png_ptr, "invalid values");
+      return;
+   }
+
+   /* If a colorspace error has already been output skip this chunk */
+   if (png_ptr->colorspace.flags & PNG_COLORSPACE_INVALID)
+      return;
+
+   if (png_ptr->colorspace.flags & PNG_COLORSPACE_FROM_cHRM)
+   {
+      png_ptr->colorspace.flags |= PNG_COLORSPACE_INVALID;
+      png_colorspace_sync(png_ptr, info_ptr);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   png_ptr->colorspace.flags |= PNG_COLORSPACE_FROM_cHRM;
+   (void)png_colorspace_set_chromaticities(png_ptr, &png_ptr->colorspace, &xy,
+      1/*prefer cHRM values*/);
+   png_colorspace_sync(png_ptr, info_ptr);
+}
+#endif
+
+#ifdef PNG_READ_sRGB_SUPPORTED
+void /* PRIVATE */
+png_handle_sRGB(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte intent;
+
+   png_debug(1, "in png_handle_sRGB");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & (PNG_HAVE_IDAT|PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   if (length != 1)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, &intent, 1);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   /* If a colorspace error has already been output skip this chunk */
+   if (png_ptr->colorspace.flags & PNG_COLORSPACE_INVALID)
+      return;
+
+   /* Only one sRGB or iCCP chunk is allowed, use the HAVE_INTENT flag to detect
+    * this.
+    */
+   if (png_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_INTENT)
+   {
+      png_ptr->colorspace.flags |= PNG_COLORSPACE_INVALID;
+      png_colorspace_sync(png_ptr, info_ptr);
+      png_chunk_benign_error(png_ptr, "too many profiles");
+      return;
+   }
+
+   (void)png_colorspace_set_sRGB(png_ptr, &png_ptr->colorspace, intent);
+   png_colorspace_sync(png_ptr, info_ptr);
+}
+#endif /* PNG_READ_sRGB_SUPPORTED */
+
+#ifdef PNG_READ_iCCP_SUPPORTED
+void /* PRIVATE */
+png_handle_iCCP(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+/* Note: this does not properly handle profiles that are > 64K under DOS */
+{
+   png_const_charp errmsg = NULL; /* error message output, or no error */
+   int finished = 0; /* crc checked */
+
+   png_debug(1, "in png_handle_iCCP");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & (PNG_HAVE_IDAT|PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   /* Consistent with all the above colorspace handling an obviously *invalid*
+    * chunk is just ignored, so does not invalidate the color space.  An
+    * alternative is to set the 'invalid' flags at the start of this routine
+    * and only clear them in they were not set before and all the tests pass.
+    * The minimum 'deflate' stream is assumed to be just the 2 byte header and 4
+    * byte checksum.  The keyword must be one character and there is a
+    * terminator (0) byte and the compression method.
+    */
+   if (length < 9)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "too short");
+      return;
+   }
+
+   /* If a colorspace error has already been output skip this chunk */
+   if (png_ptr->colorspace.flags & PNG_COLORSPACE_INVALID)
+   {
+      png_crc_finish(png_ptr, length);
+      return;
+   }
+
+   /* Only one sRGB or iCCP chunk is allowed, use the HAVE_INTENT flag to detect
+    * this.
+    */
+   if ((png_ptr->colorspace.flags & PNG_COLORSPACE_HAVE_INTENT) == 0)
+   {
+      uInt read_length, keyword_length;
+      char keyword[81];
+
+      /* Find the keyword; the keyword plus separator and compression method
+       * bytes can be at most 81 characters long.
+       */
+      read_length = 81; /* maximum */
+      if (read_length > length)
+         read_length = (uInt)length;
+
+      png_crc_read(png_ptr, (png_bytep)keyword, read_length);
+      length -= read_length;
+
+      keyword_length = 0;
+      while (keyword_length < 80 && keyword_length < read_length &&
+         keyword[keyword_length] != 0)
+         ++keyword_length;
+
+      /* TODO: make the keyword checking common */
+      if (keyword_length >= 1 && keyword_length <= 79)
+      {
+         /* We only understand '0' compression - deflate - so if we get a
+          * different value we can't safely decode the chunk.
+          */
+         if (keyword_length+1 < read_length &&
+            keyword[keyword_length+1] == PNG_COMPRESSION_TYPE_BASE)
+         {
+            read_length -= keyword_length+2;
+
+            if (png_inflate_claim(png_ptr, png_iCCP) == Z_OK)
+            {
+               Byte profile_header[132];
+               Byte local_buffer[PNG_INFLATE_BUF_SIZE];
+               png_alloc_size_t size = (sizeof profile_header);
+
+               png_ptr->zstream.next_in = (Bytef*)keyword + (keyword_length+2);
+               png_ptr->zstream.avail_in = read_length;
+               (void)png_inflate_read(png_ptr, local_buffer,
+                  (sizeof local_buffer), &length, profile_header, &size,
+                  0/*finish: don't, because the output is too small*/);
+
+               if (size == 0)
+               {
+                  /* We have the ICC profile header; do the basic header checks.
+                   */
+                  const png_uint_32 profile_length =
+                     png_get_uint_32(profile_header);
+
+                  if (png_icc_check_length(png_ptr, &png_ptr->colorspace,
+                     keyword, profile_length))
+                  {
+                     /* The length is apparently ok, so we can check the 132
+                      * byte header.
+                      */
+                     if (png_icc_check_header(png_ptr, &png_ptr->colorspace,
+                        keyword, profile_length, profile_header,
+                        png_ptr->color_type))
+                     {
+                        /* Now read the tag table; a variable size buffer is
+                         * needed at this point, allocate one for the whole
+                         * profile.  The header check has already validated
+                         * that none of these stuff will overflow.
+                         */
+                        const png_uint_32 tag_count = png_get_uint_32(
+                           profile_header+128);
+                        png_bytep profile = png_read_buffer(png_ptr,
+                           profile_length, 2/*silent*/);
+
+                        if (profile != NULL)
+                        {
+                           memcpy(profile, profile_header,
+                              (sizeof profile_header));
+
+                           size = 12 * tag_count;
+
+                           (void)png_inflate_read(png_ptr, local_buffer,
+                              (sizeof local_buffer), &length,
+                              profile + (sizeof profile_header), &size, 0);
+
+                           /* Still expect a a buffer error because we expect
+                            * there to be some tag data!
+                            */
+                           if (size == 0)
+                           {
+                              if (png_icc_check_tag_table(png_ptr,
+                                 &png_ptr->colorspace, keyword, profile_length,
+                                 profile))
+                              {
+                                 /* The profile has been validated for basic
+                                  * security issues, so read the whole thing in.
+                                  */
+                                 size = profile_length - (sizeof profile_header)
+                                    - 12 * tag_count;
+
+                                 (void)png_inflate_read(png_ptr, local_buffer,
+                                    (sizeof local_buffer), &length,
+                                    profile + (sizeof profile_header) +
+                                    12 * tag_count, &size, 1/*finish*/);
+
+                                 if (length > 0 && !(png_ptr->flags &
+                                       PNG_FLAG_BENIGN_ERRORS_WARN))
+                                    errmsg = "extra compressed data";
+
+                                 /* But otherwise allow extra data: */
+                                 else if (size == 0)
+                                 {
+                                    if (length > 0)
+                                    {
+                                       /* This can be handled completely, so
+                                        * keep going.
+                                        */
+                                       png_chunk_warning(png_ptr,
+                                          "extra compressed data");
+                                    }
+
+                                    png_crc_finish(png_ptr, length);
+                                    finished = 1;
+
+#                                   ifdef PNG_sRGB_SUPPORTED
+                                       /* Check for a match against sRGB */
+                                       png_icc_set_sRGB(png_ptr,
+                                          &png_ptr->colorspace, profile,
+                                          png_ptr->zstream.adler);
+#                                   endif
+
+                                    /* Steal the profile for info_ptr. */
+                                    if (info_ptr != NULL)
+                                    {
+                                       png_free_data(png_ptr, info_ptr,
+                                          PNG_FREE_ICCP, 0);
+
+                                       info_ptr->iccp_name = png_voidcast(char*,
+                                          png_malloc_base(png_ptr,
+                                          keyword_length+1));
+                                       if (info_ptr->iccp_name != NULL)
+                                       {
+                                          memcpy(info_ptr->iccp_name, keyword,
+                                             keyword_length+1);
+                                          info_ptr->iccp_proflen =
+                                             profile_length;
+                                          info_ptr->iccp_profile = profile;
+                                          png_ptr->read_buffer = NULL; /*steal*/
+                                          info_ptr->free_me |= PNG_FREE_ICCP;
+                                          info_ptr->valid |= PNG_INFO_iCCP;
+                                       }
+
+                                       else
+                                       {
+                                          png_ptr->colorspace.flags |=
+                                             PNG_COLORSPACE_INVALID;
+                                          errmsg = "out of memory";
+                                       }
+                                    }
+
+                                    /* else the profile remains in the read
+                                     * buffer which gets reused for subsequent
+                                     * chunks.
+                                     */
+
+                                    if (info_ptr != NULL)
+                                       png_colorspace_sync(png_ptr, info_ptr);
+
+                                    if (errmsg == NULL)
+                                    {
+                                       png_ptr->zowner = 0;
+                                       return;
+                                    }
+                                 }
+
+                                 else if (size > 0)
+                                    errmsg = "truncated";
+
+                                 else
+                                    errmsg = png_ptr->zstream.msg;
+                              }
+
+                              /* else png_icc_check_tag_table output an error */
+                           }
+
+                           else /* profile truncated */
+                              errmsg = png_ptr->zstream.msg;
+                        }
+
+                        else
+                           errmsg = "out of memory";
+                     }
+
+                     /* else png_icc_check_header output an error */
+                  }
+
+                  /* else png_icc_check_length output an error */
+               }
+
+               else /* profile truncated */
+                  errmsg = png_ptr->zstream.msg;
+
+               /* Release the stream */
+               png_ptr->zowner = 0;
+            }
+
+            else /* png_inflate_claim failed */
+               errmsg = png_ptr->zstream.msg;
+         }
+
+         else
+            errmsg = "bad compression method"; /* or missing */
+      }
+
+      else
+         errmsg = "bad keyword";
+   }
+
+   else
+      errmsg = "too many profiles";
+
+   /* Failure: the reason is in 'errmsg' */
+   if (!finished)
+      png_crc_finish(png_ptr, length);
+
+   png_ptr->colorspace.flags |= PNG_COLORSPACE_INVALID;
+   png_colorspace_sync(png_ptr, info_ptr);
+   if (errmsg != NULL) /* else already output */
+      png_chunk_benign_error(png_ptr, errmsg);
+}
+#endif /* PNG_READ_iCCP_SUPPORTED */
+
+#ifdef PNG_READ_sPLT_SUPPORTED
+void /* PRIVATE */
+png_handle_sPLT(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+/* Note: this does not properly handle chunks that are > 64K under DOS */
+{
+   png_bytep entry_start, buffer;
+   png_sPLT_t new_palette;
+   png_sPLT_entryp pp;
+   png_uint_32 data_length;
+   int entry_size, i;
+   png_uint_32 skip = 0;
+   png_uint_32 dl;
+   png_size_t max_dl;
+
+   png_debug(1, "in png_handle_sPLT");
+
+#ifdef PNG_USER_LIMITS_SUPPORTED
+   if (png_ptr->user_chunk_cache_max != 0)
+   {
+      if (png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         return;
+      }
+
+      if (--png_ptr->user_chunk_cache_max == 1)
+      {
+         png_warning(png_ptr, "No space in chunk cache for sPLT");
+         png_crc_finish(png_ptr, length);
+         return;
+      }
+   }
+#endif
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+#ifdef PNG_MAX_MALLOC_64K
+   if (length > 65535U)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "too large to fit in memory");
+      return;
+   }
+#endif
+
+   buffer = png_read_buffer(png_ptr, length+1, 2/*silent*/);
+   if (buffer == NULL)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of memory");
+      return;
+   }
+
+
+   /* WARNING: this may break if size_t is less than 32 bits; it is assumed
+    * that the PNG_MAX_MALLOC_64K test is enabled in this case, but this is a
+    * potential breakage point if the types in pngconf.h aren't exactly right.
+    */
+   png_crc_read(png_ptr, buffer, length);
+
+   if (png_crc_finish(png_ptr, skip))
+      return;
+
+   buffer[length] = 0;
+
+   for (entry_start = buffer; *entry_start; entry_start++)
+      /* Empty loop to find end of name */ ;
+
+   ++entry_start;
+
+   /* A sample depth should follow the separator, and we should be on it  */
+   if (entry_start > buffer + length - 2)
+   {
+      png_warning(png_ptr, "malformed sPLT chunk");
+      return;
+   }
+
+   new_palette.depth = *entry_start++;
+   entry_size = (new_palette.depth == 8 ? 6 : 10);
+   /* This must fit in a png_uint_32 because it is derived from the original
+    * chunk data length.
+    */
+   data_length = length - (png_uint_32)(entry_start - buffer);
+
+   /* Integrity-check the data length */
+   if (data_length % entry_size)
+   {
+      png_warning(png_ptr, "sPLT chunk has bad length");
+      return;
+   }
+
+   dl = (png_int_32)(data_length / entry_size);
+   max_dl = PNG_SIZE_MAX / (sizeof (png_sPLT_entry));
+
+   if (dl > max_dl)
+   {
+       png_warning(png_ptr, "sPLT chunk too long");
+       return;
+   }
+
+   new_palette.nentries = (png_int_32)(data_length / entry_size);
+
+   new_palette.entries = (png_sPLT_entryp)png_malloc_warn(
+       png_ptr, new_palette.nentries * (sizeof (png_sPLT_entry)));
+
+   if (new_palette.entries == NULL)
+   {
+       png_warning(png_ptr, "sPLT chunk requires too much memory");
+       return;
+   }
+
+#ifdef PNG_POINTER_INDEXING_SUPPORTED
+   for (i = 0; i < new_palette.nentries; i++)
+   {
+      pp = new_palette.entries + i;
+
+      if (new_palette.depth == 8)
+      {
+         pp->red = *entry_start++;
+         pp->green = *entry_start++;
+         pp->blue = *entry_start++;
+         pp->alpha = *entry_start++;
+      }
+
+      else
+      {
+         pp->red   = png_get_uint_16(entry_start); entry_start += 2;
+         pp->green = png_get_uint_16(entry_start); entry_start += 2;
+         pp->blue  = png_get_uint_16(entry_start); entry_start += 2;
+         pp->alpha = png_get_uint_16(entry_start); entry_start += 2;
+      }
+
+      pp->frequency = png_get_uint_16(entry_start); entry_start += 2;
+   }
+#else
+   pp = new_palette.entries;
+
+   for (i = 0; i < new_palette.nentries; i++)
+   {
+
+      if (new_palette.depth == 8)
+      {
+         pp[i].red   = *entry_start++;
+         pp[i].green = *entry_start++;
+         pp[i].blue  = *entry_start++;
+         pp[i].alpha = *entry_start++;
+      }
+
+      else
+      {
+         pp[i].red   = png_get_uint_16(entry_start); entry_start += 2;
+         pp[i].green = png_get_uint_16(entry_start); entry_start += 2;
+         pp[i].blue  = png_get_uint_16(entry_start); entry_start += 2;
+         pp[i].alpha = png_get_uint_16(entry_start); entry_start += 2;
+      }
+
+      pp[i].frequency = png_get_uint_16(entry_start); entry_start += 2;
+   }
+#endif
+
+   /* Discard all chunk data except the name and stash that */
+   new_palette.name = (png_charp)buffer;
+
+   png_set_sPLT(png_ptr, info_ptr, &new_palette, 1);
+
+   png_free(png_ptr, new_palette.entries);
+}
+#endif /* PNG_READ_sPLT_SUPPORTED */
+
+#ifdef PNG_READ_tRNS_SUPPORTED
+void /* PRIVATE */
+png_handle_tRNS(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte readbuf[PNG_MAX_PALETTE_LENGTH];
+
+   png_debug(1, "in png_handle_tRNS");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_tRNS))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (png_ptr->color_type == PNG_COLOR_TYPE_GRAY)
+   {
+      png_byte buf[2];
+
+      if (length != 2)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "invalid");
+         return;
+      }
+
+      png_crc_read(png_ptr, buf, 2);
+      png_ptr->num_trans = 1;
+      png_ptr->trans_color.gray = png_get_uint_16(buf);
+   }
+
+   else if (png_ptr->color_type == PNG_COLOR_TYPE_RGB)
+   {
+      png_byte buf[6];
+
+      if (length != 6)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "invalid");
+         return;
+      }
+
+      png_crc_read(png_ptr, buf, length);
+      png_ptr->num_trans = 1;
+      png_ptr->trans_color.red = png_get_uint_16(buf);
+      png_ptr->trans_color.green = png_get_uint_16(buf + 2);
+      png_ptr->trans_color.blue = png_get_uint_16(buf + 4);
+   }
+
+   else if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+   {
+      if (!(png_ptr->mode & PNG_HAVE_PLTE))
+      {
+         /* TODO: is this actually an error in the ISO spec? */
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "out of place");
+         return;
+      }
+
+      if (length > png_ptr->num_palette || length > PNG_MAX_PALETTE_LENGTH ||
+         length == 0)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "invalid");
+         return;
+      }
+
+      png_crc_read(png_ptr, readbuf, length);
+      png_ptr->num_trans = (png_uint_16)length;
+   }
+
+   else
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid with alpha channel");
+      return;
+   }
+
+   if (png_crc_finish(png_ptr, 0))
+   {
+      png_ptr->num_trans = 0;
+      return;
+   }
+
+   /* TODO: this is a horrible side effect in the palette case because the
+    * png_struct ends up with a pointer to the tRNS buffer owned by the
+    * png_info.  Fix this.
+    */
+   png_set_tRNS(png_ptr, info_ptr, readbuf, png_ptr->num_trans,
+       &(png_ptr->trans_color));
+}
+#endif
+
+#ifdef PNG_READ_bKGD_SUPPORTED
+void /* PRIVATE */
+png_handle_bKGD(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   unsigned int truelen;
+   png_byte buf[6];
+   png_color_16 background;
+
+   png_debug(1, "in png_handle_bKGD");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if ((png_ptr->mode & PNG_HAVE_IDAT) ||
+      (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE &&
+       !(png_ptr->mode & PNG_HAVE_PLTE)))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_bKGD))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      truelen = 1;
+
+   else if (png_ptr->color_type & PNG_COLOR_MASK_COLOR)
+      truelen = 6;
+
+   else
+      truelen = 2;
+
+   if (length != truelen)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, truelen);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   /* We convert the index value into RGB components so that we can allow
+    * arbitrary RGB values for background when we have transparency, and
+    * so it is easy to determine the RGB values of the background color
+    * from the info_ptr struct.
+    */
+   if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+   {
+      background.index = buf[0];
+
+      if (info_ptr && info_ptr->num_palette)
+      {
+         if (buf[0] >= info_ptr->num_palette)
+         {
+            png_chunk_benign_error(png_ptr, "invalid index");
+            return;
+         }
+
+         background.red = (png_uint_16)png_ptr->palette[buf[0]].red;
+         background.green = (png_uint_16)png_ptr->palette[buf[0]].green;
+         background.blue = (png_uint_16)png_ptr->palette[buf[0]].blue;
+      }
+
+      else
+         background.red = background.green = background.blue = 0;
+
+      background.gray = 0;
+   }
+
+   else if (!(png_ptr->color_type & PNG_COLOR_MASK_COLOR)) /* GRAY */
+   {
+      background.index = 0;
+      background.red =
+      background.green =
+      background.blue =
+      background.gray = png_get_uint_16(buf);
+   }
+
+   else
+   {
+      background.index = 0;
+      background.red = png_get_uint_16(buf);
+      background.green = png_get_uint_16(buf + 2);
+      background.blue = png_get_uint_16(buf + 4);
+      background.gray = 0;
+   }
+
+   png_set_bKGD(png_ptr, info_ptr, &background);
+}
+#endif
+
+#ifdef PNG_READ_hIST_SUPPORTED
+void /* PRIVATE */
+png_handle_hIST(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   unsigned int num, i;
+   png_uint_16 readbuf[PNG_MAX_PALETTE_LENGTH];
+
+   png_debug(1, "in png_handle_hIST");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if ((png_ptr->mode & PNG_HAVE_IDAT) || !(png_ptr->mode & PNG_HAVE_PLTE))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_hIST))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   num = length / 2 ;
+
+   if (num != png_ptr->num_palette || num > PNG_MAX_PALETTE_LENGTH)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   for (i = 0; i < num; i++)
+   {
+      png_byte buf[2];
+
+      png_crc_read(png_ptr, buf, 2);
+      readbuf[i] = png_get_uint_16(buf);
+   }
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   png_set_hIST(png_ptr, info_ptr, readbuf);
+}
+#endif
+
+#ifdef PNG_READ_pHYs_SUPPORTED
+void /* PRIVATE */
+png_handle_pHYs(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte buf[9];
+   png_uint_32 res_x, res_y;
+   int unit_type;
+
+   png_debug(1, "in png_handle_pHYs");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_pHYs))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (length != 9)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, 9);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   res_x = png_get_uint_32(buf);
+   res_y = png_get_uint_32(buf + 4);
+   unit_type = buf[8];
+   png_set_pHYs(png_ptr, info_ptr, res_x, res_y, unit_type);
+}
+#endif
+
+#ifdef PNG_READ_oFFs_SUPPORTED
+void /* PRIVATE */
+png_handle_oFFs(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte buf[9];
+   png_int_32 offset_x, offset_y;
+   int unit_type;
+
+   png_debug(1, "in png_handle_oFFs");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_oFFs))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (length != 9)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, 9);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   offset_x = png_get_int_32(buf);
+   offset_y = png_get_int_32(buf + 4);
+   unit_type = buf[8];
+   png_set_oFFs(png_ptr, info_ptr, offset_x, offset_y, unit_type);
+}
+#endif
+
+#ifdef PNG_READ_pCAL_SUPPORTED
+/* Read the pCAL chunk (described in the PNG Extensions document) */
+void /* PRIVATE */
+png_handle_pCAL(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_int_32 X0, X1;
+   png_byte type, nparams;
+   png_bytep buffer, buf, units, endptr;
+   png_charpp params;
+   int i;
+
+   png_debug(1, "in png_handle_pCAL");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_pCAL))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   png_debug1(2, "Allocating and reading pCAL chunk data (%u bytes)",
+       length + 1);
+
+   buffer = png_read_buffer(png_ptr, length+1, 2/*silent*/);
+
+   if (buffer == NULL)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of memory");
+      return;
+   }
+
+   png_crc_read(png_ptr, buffer, length);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   buffer[length] = 0; /* Null terminate the last string */
+
+   png_debug(3, "Finding end of pCAL purpose string");
+   for (buf = buffer; *buf; buf++)
+      /* Empty loop */ ;
+
+   endptr = buffer + length;
+
+   /* We need to have at least 12 bytes after the purpose string
+    * in order to get the parameter information.
+    */
+   if (endptr <= buf + 12)
+   {
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_debug(3, "Reading pCAL X0, X1, type, nparams, and units");
+   X0 = png_get_int_32((png_bytep)buf+1);
+   X1 = png_get_int_32((png_bytep)buf+5);
+   type = buf[9];
+   nparams = buf[10];
+   units = buf + 11;
+
+   png_debug(3, "Checking pCAL equation type and number of parameters");
+   /* Check that we have the right number of parameters for known
+    * equation types.
+    */
+   if ((type == PNG_EQUATION_LINEAR && nparams != 2) ||
+       (type == PNG_EQUATION_BASE_E && nparams != 3) ||
+       (type == PNG_EQUATION_ARBITRARY && nparams != 3) ||
+       (type == PNG_EQUATION_HYPERBOLIC && nparams != 4))
+   {
+      png_chunk_benign_error(png_ptr, "invalid parameter count");
+      return;
+   }
+
+   else if (type >= PNG_EQUATION_LAST)
+   {
+      png_chunk_benign_error(png_ptr, "unrecognized equation type");
+   }
+
+   for (buf = units; *buf; buf++)
+      /* Empty loop to move past the units string. */ ;
+
+   png_debug(3, "Allocating pCAL parameters array");
+
+   params = png_voidcast(png_charpp, png_malloc_warn(png_ptr,
+       nparams * (sizeof (png_charp))));
+
+   if (params == NULL)
+   {
+      png_chunk_benign_error(png_ptr, "out of memory");
+      return;
+   }
+
+   /* Get pointers to the start of each parameter string. */
+   for (i = 0; i < nparams; i++)
+   {
+      buf++; /* Skip the null string terminator from previous parameter. */
+
+      png_debug1(3, "Reading pCAL parameter %d", i);
+
+      for (params[i] = (png_charp)buf; buf <= endptr && *buf != 0; buf++)
+         /* Empty loop to move past each parameter string */ ;
+
+      /* Make sure we haven't run out of data yet */
+      if (buf > endptr)
+      {
+         png_free(png_ptr, params);
+         png_chunk_benign_error(png_ptr, "invalid data");
+         return;
+      }
+   }
+
+   png_set_pCAL(png_ptr, info_ptr, (png_charp)buffer, X0, X1, type, nparams,
+      (png_charp)units, params);
+
+   png_free(png_ptr, params);
+}
+#endif
+
+#ifdef PNG_READ_sCAL_SUPPORTED
+/* Read the sCAL chunk */
+void /* PRIVATE */
+png_handle_sCAL(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_bytep buffer;
+   png_size_t i;
+   int state;
+
+   png_debug(1, "in png_handle_sCAL");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (png_ptr->mode & PNG_HAVE_IDAT)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of place");
+      return;
+   }
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_sCAL))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   /* Need unit type, width, \0, height: minimum 4 bytes */
+   else if (length < 4)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_debug1(2, "Allocating and reading sCAL chunk data (%u bytes)",
+      length + 1);
+
+   buffer = png_read_buffer(png_ptr, length+1, 2/*silent*/);
+
+   if (buffer == NULL)
+   {
+      png_chunk_benign_error(png_ptr, "out of memory");
+      png_crc_finish(png_ptr, length);
+      return;
+   }
+
+   png_crc_read(png_ptr, buffer, length);
+   buffer[length] = 0; /* Null terminate the last string */
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   /* Validate the unit. */
+   if (buffer[0] != 1 && buffer[0] != 2)
+   {
+      png_chunk_benign_error(png_ptr, "invalid unit");
+      return;
+   }
+
+   /* Validate the ASCII numbers, need two ASCII numbers separated by
+    * a '\0' and they need to fit exactly in the chunk data.
+    */
+   i = 1;
+   state = 0;
+
+   if (!png_check_fp_number((png_const_charp)buffer, length, &state, &i) ||
+       i >= length || buffer[i++] != 0)
+      png_chunk_benign_error(png_ptr, "bad width format");
+
+   else if (!PNG_FP_IS_POSITIVE(state))
+      png_chunk_benign_error(png_ptr, "non-positive width");
+
+   else
+   {
+      png_size_t heighti = i;
+
+      state = 0;
+      if (!png_check_fp_number((png_const_charp)buffer, length, &state, &i) ||
+         i != length)
+         png_chunk_benign_error(png_ptr, "bad height format");
+
+      else if (!PNG_FP_IS_POSITIVE(state))
+         png_chunk_benign_error(png_ptr, "non-positive height");
+
+      else
+         /* This is the (only) success case. */
+         png_set_sCAL_s(png_ptr, info_ptr, buffer[0],
+            (png_charp)buffer+1, (png_charp)buffer+heighti);
+   }
+}
+#endif
+
+#ifdef PNG_READ_tIME_SUPPORTED
+void /* PRIVATE */
+png_handle_tIME(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_byte buf[7];
+   png_time mod_time;
+
+   png_debug(1, "in png_handle_tIME");
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   else if (info_ptr != NULL && (info_ptr->valid & PNG_INFO_tIME))
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "duplicate");
+      return;
+   }
+
+   if (png_ptr->mode & PNG_HAVE_IDAT)
+      png_ptr->mode |= PNG_AFTER_IDAT;
+
+   if (length != 7)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "invalid");
+      return;
+   }
+
+   png_crc_read(png_ptr, buf, 7);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   mod_time.second = buf[6];
+   mod_time.minute = buf[5];
+   mod_time.hour = buf[4];
+   mod_time.day = buf[3];
+   mod_time.month = buf[2];
+   mod_time.year = png_get_uint_16(buf);
+
+   png_set_tIME(png_ptr, info_ptr, &mod_time);
+}
+#endif
+
+#ifdef PNG_READ_tEXt_SUPPORTED
+/* Note: this does not properly handle chunks that are > 64K under DOS */
+void /* PRIVATE */
+png_handle_tEXt(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_text  text_info;
+   png_bytep buffer;
+   png_charp key;
+   png_charp text;
+   png_uint_32 skip = 0;
+
+   png_debug(1, "in png_handle_tEXt");
+
+#ifdef PNG_USER_LIMITS_SUPPORTED
+   if (png_ptr->user_chunk_cache_max != 0)
+   {
+      if (png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         return;
+      }
+
+      if (--png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "no space in chunk cache");
+         return;
+      }
+   }
+#endif
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   if (png_ptr->mode & PNG_HAVE_IDAT)
+      png_ptr->mode |= PNG_AFTER_IDAT;
+
+#ifdef PNG_MAX_MALLOC_64K
+   if (length > 65535U)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "too large to fit in memory");
+      return;
+   }
+#endif
+
+   buffer = png_read_buffer(png_ptr, length+1, 1/*warn*/);
+
+   if (buffer == NULL)
+   {
+     png_chunk_benign_error(png_ptr, "out of memory");
+     return;
+   }
+
+   png_crc_read(png_ptr, buffer, length);
+
+   if (png_crc_finish(png_ptr, skip))
+      return;
+
+   key = (png_charp)buffer;
+   key[length] = 0;
+
+   for (text = key; *text; text++)
+      /* Empty loop to find end of key */ ;
+
+   if (text != key + length)
+      text++;
+
+   text_info.compression = PNG_TEXT_COMPRESSION_NONE;
+   text_info.key = key;
+   text_info.lang = NULL;
+   text_info.lang_key = NULL;
+   text_info.itxt_length = 0;
+   text_info.text = text;
+   text_info.text_length = strlen(text);
+
+   if (png_set_text_2(png_ptr, info_ptr, &text_info, 1))
+      png_warning(png_ptr, "Insufficient memory to process text chunk");
+}
+#endif
+
+#ifdef PNG_READ_zTXt_SUPPORTED
+/* Note: this does not correctly handle chunks that are > 64K under DOS */
+void /* PRIVATE */
+png_handle_zTXt(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_const_charp errmsg = NULL;
+   png_bytep       buffer;
+   png_uint_32     keyword_length;
+
+   png_debug(1, "in png_handle_zTXt");
+
+#ifdef PNG_USER_LIMITS_SUPPORTED
+   if (png_ptr->user_chunk_cache_max != 0)
+   {
+      if (png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         return;
+      }
+
+      if (--png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "no space in chunk cache");
+         return;
+      }
+   }
+#endif
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   if (png_ptr->mode & PNG_HAVE_IDAT)
+      png_ptr->mode |= PNG_AFTER_IDAT;
+
+   buffer = png_read_buffer(png_ptr, length, 2/*silent*/);
+
+   if (buffer == NULL)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of memory");
+      return;
+   }
+
+   png_crc_read(png_ptr, buffer, length);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   /* TODO: also check that the keyword contents match the spec! */
+   for (keyword_length = 0;
+      keyword_length < length && buffer[keyword_length] != 0;
+      ++keyword_length)
+      /* Empty loop to find end of name */ ;
+
+   if (keyword_length > 79 || keyword_length < 1)
+      errmsg = "bad keyword";
+
+   /* zTXt must have some LZ data after the keyword, although it may expand to
+    * zero bytes; we need a '\0' at the end of the keyword, the compression type
+    * then the LZ data:
+    */
+   else if (keyword_length + 3 > length)
+      errmsg = "truncated";
+
+   else if (buffer[keyword_length+1] != PNG_COMPRESSION_TYPE_BASE)
+      errmsg = "unknown compression type";
+
+   else
+   {
+      png_alloc_size_t uncompressed_length = PNG_SIZE_MAX;
+
+      /* TODO: at present png_decompress_chunk imposes a single application
+       * level memory limit, this should be split to different values for iCCP
+       * and text chunks.
+       */
+      if (png_decompress_chunk(png_ptr, length, keyword_length+2,
+         &uncompressed_length, 1/*terminate*/) == Z_STREAM_END)
+      {
+         png_text text;
+
+         /* It worked; png_ptr->read_buffer now looks like a tEXt chunk except
+          * for the extra compression type byte and the fact that it isn't
+          * necessarily '\0' terminated.
+          */
+         buffer = png_ptr->read_buffer;
+         buffer[uncompressed_length+(keyword_length+2)] = 0;
+
+         text.compression = PNG_TEXT_COMPRESSION_zTXt;
+         text.key = (png_charp)buffer;
+         text.text = (png_charp)(buffer + keyword_length+2);
+         text.text_length = uncompressed_length;
+         text.itxt_length = 0;
+         text.lang = NULL;
+         text.lang_key = NULL;
+
+         if (png_set_text_2(png_ptr, info_ptr, &text, 1))
+            errmsg = "insufficient memory";
+      }
+
+      else
+         errmsg = png_ptr->zstream.msg;
+   }
+
+   if (errmsg != NULL)
+      png_chunk_benign_error(png_ptr, errmsg);
+}
+#endif
+
+#ifdef PNG_READ_iTXt_SUPPORTED
+/* Note: this does not correctly handle chunks that are > 64K under DOS */
+void /* PRIVATE */
+png_handle_iTXt(png_structrp png_ptr, png_inforp info_ptr, png_uint_32 length)
+{
+   png_const_charp errmsg = NULL;
+   png_bytep buffer;
+   png_uint_32 prefix_length;
+
+   png_debug(1, "in png_handle_iTXt");
+
+#ifdef PNG_USER_LIMITS_SUPPORTED
+   if (png_ptr->user_chunk_cache_max != 0)
+   {
+      if (png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         return;
+      }
+
+      if (--png_ptr->user_chunk_cache_max == 1)
+      {
+         png_crc_finish(png_ptr, length);
+         png_chunk_benign_error(png_ptr, "no space in chunk cache");
+         return;
+      }
+   }
+#endif
+
+   if (!(png_ptr->mode & PNG_HAVE_IHDR))
+      png_chunk_error(png_ptr, "missing IHDR");
+
+   if (png_ptr->mode & PNG_HAVE_IDAT)
+      png_ptr->mode |= PNG_AFTER_IDAT;
+
+   buffer = png_read_buffer(png_ptr, length+1, 1/*warn*/);
+
+   if (buffer == NULL)
+   {
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "out of memory");
+      return;
+   }
+
+   png_crc_read(png_ptr, buffer, length);
+
+   if (png_crc_finish(png_ptr, 0))
+      return;
+
+   /* First the keyword. */
+   for (prefix_length=0;
+      prefix_length < length && buffer[prefix_length] != 0;
+      ++prefix_length)
+      /* Empty loop */ ;
+
+   /* Perform a basic check on the keyword length here. */
+   if (prefix_length > 79 || prefix_length < 1)
+      errmsg = "bad keyword";
+
+   /* Expect keyword, compression flag, compression type, language, translated
+    * keyword (both may be empty but are 0 terminated) then the text, which may
+    * be empty.
+    */
+   else if (prefix_length + 5 > length)
+      errmsg = "truncated";
+
+   else if (buffer[prefix_length+1] == 0 ||
+      (buffer[prefix_length+1] == 1 &&
+      buffer[prefix_length+2] == PNG_COMPRESSION_TYPE_BASE))
+   {
+      int compressed = buffer[prefix_length+1] != 0;
+      png_uint_32 language_offset, translated_keyword_offset;
+      png_alloc_size_t uncompressed_length = 0;
+
+      /* Now the language tag */
+      prefix_length += 3;
+      language_offset = prefix_length;
+
+      for (; prefix_length < length && buffer[prefix_length] != 0;
+         ++prefix_length)
+         /* Empty loop */ ;
+
+      /* WARNING: the length may be invalid here, this is checked below. */
+      translated_keyword_offset = ++prefix_length;
+
+      for (; prefix_length < length && buffer[prefix_length] != 0;
+         ++prefix_length)
+         /* Empty loop */ ;
+
+      /* prefix_length should now be at the trailing '\0' of the translated
+       * keyword, but it may already be over the end.  None of this arithmetic
+       * can overflow because chunks are at most 2^31 bytes long, but on 16-bit
+       * systems the available allocaton may overflow.
+       */
+      ++prefix_length;
+
+      if (!compressed && prefix_length <= length)
+         uncompressed_length = length - prefix_length;
+
+      else if (compressed && prefix_length < length)
+      {
+         uncompressed_length = PNG_SIZE_MAX;
+
+         /* TODO: at present png_decompress_chunk imposes a single application
+          * level memory limit, this should be split to different values for
+          * iCCP and text chunks.
+          */
+         if (png_decompress_chunk(png_ptr, length, prefix_length,
+            &uncompressed_length, 1/*terminate*/) == Z_STREAM_END)
+            buffer = png_ptr->read_buffer;
+
+         else
+            errmsg = png_ptr->zstream.msg;
+      }
+
+      else
+         errmsg = "truncated";
+
+      if (errmsg == NULL)
+      {
+         png_text text;
+
+         buffer[uncompressed_length+prefix_length] = 0;
+
+         if (compressed)
+            text.compression = PNG_ITXT_COMPRESSION_NONE;
+
+         else
+            text.compression = PNG_ITXT_COMPRESSION_zTXt;
+
+         text.key = (png_charp)buffer;
+         text.lang = (png_charp)buffer + language_offset;
+         text.lang_key = (png_charp)buffer + translated_keyword_offset;
+         text.text = (png_charp)buffer + prefix_length;
+         text.text_length = 0;
+         text.itxt_length = uncompressed_length;
+
+         if (png_set_text_2(png_ptr, info_ptr, &text, 1))
+            errmsg = "insufficient memory";
+      }
+   }
+
+   else
+      errmsg = "bad compression info";
+
+   if (errmsg != NULL)
+      png_chunk_benign_error(png_ptr, errmsg);
+}
+#endif
+
+#ifdef PNG_READ_UNKNOWN_CHUNKS_SUPPORTED
+/* Utility function for png_handle_unknown; set up png_ptr::unknown_chunk */
+static int
+png_cache_unknown_chunk(png_structrp png_ptr, png_uint_32 length)
+{
+   png_alloc_size_t limit = PNG_SIZE_MAX;
+
+   if (png_ptr->unknown_chunk.data != NULL)
+   {
+      png_free(png_ptr, png_ptr->unknown_chunk.data);
+      png_ptr->unknown_chunk.data = NULL;
+   }
+
+#  ifdef PNG_SET_CHUNK_MALLOC_LIMIT_SUPPORTED
+      if (png_ptr->user_chunk_malloc_max > 0 &&
+         png_ptr->user_chunk_malloc_max < limit)
+         limit = png_ptr->user_chunk_malloc_max;
+
+#  elif PNG_USER_CHUNK_MALLOC_MAX > 0
+      if (PNG_USER_CHUNK_MALLOC_MAX < limit)
+         limit = PNG_USER_CHUNK_MALLOC_MAX;
+#  endif
+
+   if (length <= limit)
+   {
+      PNG_CSTRING_FROM_CHUNK(png_ptr->unknown_chunk.name, png_ptr->chunk_name);
+      /* The following is safe because of the PNG_SIZE_MAX init above */
+      png_ptr->unknown_chunk.size = (png_size_t)length/*SAFE*/;
+      /* 'mode' is a flag array, only the bottom four bits matter here */
+      png_ptr->unknown_chunk.location = (png_byte)png_ptr->mode/*SAFE*/;
+
+      if (length == 0)
+         png_ptr->unknown_chunk.data = NULL;
+
+      else
+      {
+         /* Do a 'warn' here - it is handled below. */
+         png_ptr->unknown_chunk.data = png_voidcast(png_bytep,
+            png_malloc_warn(png_ptr, length));
+      }
+   }
+
+   if (png_ptr->unknown_chunk.data == NULL && length > 0)
+   {
+      /* This is benign because we clean up correctly */
+      png_crc_finish(png_ptr, length);
+      png_chunk_benign_error(png_ptr, "unknown chunk exceeds memory limits");
+      return 0;
+   }
+
+   else
+   {
+      if (length > 0)
+         png_crc_read(png_ptr, png_ptr->unknown_chunk.data, length);
+      png_crc_finish(png_ptr, 0);
+      return 1;
+   }
+}
+#endif /* PNG_READ_UNKNOWN_CHUNKS_SUPPORTED */
+
+/* Handle an unknown, or known but disabled, chunk */
+void /* PRIVATE */
+png_handle_unknown(png_structrp png_ptr, png_inforp info_ptr,
+   png_uint_32 length, int keep)
+{
+   int handled = 0; /* the chunk was handled */
+
+   png_debug(1, "in png_handle_unknown");
+
+#ifdef PNG_READ_UNKNOWN_CHUNKS_SUPPORTED
+   /* NOTE: this code is based on the code in libpng-1.4.12 except for fixing
+    * the bug which meant that setting a non-default behavior for a specific
+    * chunk would be ignored (the default was always used unless a user
+    * callback was installed).
+    *
+    * 'keep' is the value from the png_chunk_unknown_handling, the setting for
+    * this specific chunk_name, if PNG_HANDLE_AS_UNKNOWN_SUPPORTED, if not it
+    * will always be PNG_HANDLE_CHUNK_AS_DEFAULT and it needs to be set here.
+    * This is just an optimization to avoid multiple calls to the lookup
+    * function.
+    */
+#  ifndef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+#     ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+         keep = png_chunk_unknown_handling(png_ptr, png_ptr->chunk_name);
+#     endif
+#  endif
+
+   /* One of the following methods will read the chunk or skip it (at least one
+    * of these is always defined because this is the only way to switch on
+    * PNG_READ_UNKNOWN_CHUNKS_SUPPORTED)
+    */
+#  ifdef PNG_READ_USER_CHUNKS_SUPPORTED
+      /* The user callback takes precedence over the chunk keep value, but the
+       * keep value is still required to validate a save of a critical chunk.
+       */
+      if (png_ptr->read_user_chunk_fn != NULL)
+      {
+         if (png_cache_unknown_chunk(png_ptr, length))
+         {
+            /* Callback to user unknown chunk handler */
+            int ret = (*(png_ptr->read_user_chunk_fn))(png_ptr,
+               &png_ptr->unknown_chunk);
+
+            /* ret is:
+             * negative: An error occured, png_chunk_error will be called.
+             *     zero: The chunk was not handled, the chunk will be discarded
+             *           unless png_set_keep_unknown_chunks has been used to set
+             *           a 'keep' behavior for this particular chunk, in which
+             *           case that will be used.  A critical chunk will cause an
+             *           error at this point unless it is to be saved.
+             * positive: The chunk was handled, libpng will ignore/discard it.
+             */
+            if (ret < 0)
+               png_chunk_error(png_ptr, "error in user chunk");
+
+            else if (ret == 0)
+            {
+               /* If the keep value is 'default' or 'never' override it, but
+                * still error out on critical chunks unless the keep value is
+                * 'always'  While this is weird it is the behavior in 1.4.12.
+                * A possible improvement would be to obey the value set for the
+                * chunk, but this would be an API change that would probably
+                * damage some applications.
+                *
+                * The png_app_warning below catches the case that matters, where
+                * the application has not set specific save or ignore for this
+                * chunk or global save or ignore.
+                */
+               if (keep < PNG_HANDLE_CHUNK_IF_SAFE)
+               {
+#                 ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+                     if (png_ptr->unknown_default < PNG_HANDLE_CHUNK_IF_SAFE)
+                     {
+                        png_chunk_warning(png_ptr, "Saving unknown chunk:");
+                        png_app_warning(png_ptr,
+                           "forcing save of an unhandled chunk;"
+                           " please call png_set_keep_unknown_chunks");
+                           /* with keep = PNG_HANDLE_CHUNK_IF_SAFE */
+                     }
+#                 endif
+                  keep = PNG_HANDLE_CHUNK_IF_SAFE;
+               }
+            }
+
+            else /* chunk was handled */
+            {
+               handled = 1;
+               /* Critical chunks can be safely discarded at this point. */
+               keep = PNG_HANDLE_CHUNK_NEVER;
+            }
+         }
+
+         else
+            keep = PNG_HANDLE_CHUNK_NEVER; /* insufficient memory */
+      }
+
+      else
+         /* Use the SAVE_UNKNOWN_CHUNKS code or skip the chunk */
+#  endif /* PNG_READ_USER_CHUNKS_SUPPORTED */
+
+#  ifdef PNG_SAVE_UNKNOWN_CHUNKS_SUPPORTED
+      {
+         /* keep is currently just the per-chunk setting, if there was no
+          * setting change it to the global default now (not that this may
+          * still be AS_DEFAULT) then obtain the cache of the chunk if required,
+          * if not simply skip the chunk.
+          */
+         if (keep == PNG_HANDLE_CHUNK_AS_DEFAULT)
+            keep = png_ptr->unknown_default;
+
+         if (keep == PNG_HANDLE_CHUNK_ALWAYS ||
+            (keep == PNG_HANDLE_CHUNK_IF_SAFE &&
+             PNG_CHUNK_ANCILLARY(png_ptr->chunk_name)))
+         {
+            if (!png_cache_unknown_chunk(png_ptr, length))
+               keep = PNG_HANDLE_CHUNK_NEVER;
+         }
+
+         else
+            png_crc_finish(png_ptr, length);
+      }
+#  else
+#     ifndef PNG_READ_USER_CHUNKS_SUPPORTED
+#        error no method to support READ_UNKNOWN_CHUNKS
+#     endif
+
+      {
+         /* If here there is no read callback pointer set and no support is
+          * compiled in to just save the unknown chunks, so simply skip this
+          * chunk.  If 'keep' is something other than AS_DEFAULT or NEVER then
+          * the app has erroneously asked for unknown chunk saving when there
+          * is no support.
+          */
+         if (keep > PNG_HANDLE_CHUNK_NEVER)
+            png_app_error(png_ptr, "no unknown chunk support available");
+
+         png_crc_finish(png_ptr, length);
+      }
+#  endif
+
+#  ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+      /* Now store the chunk in the chunk list if appropriate, and if the limits
+       * permit it.
+       */
+      if (keep == PNG_HANDLE_CHUNK_ALWAYS ||
+         (keep == PNG_HANDLE_CHUNK_IF_SAFE &&
+          PNG_CHUNK_ANCILLARY(png_ptr->chunk_name)))
+      {
+#     ifdef PNG_USER_LIMITS_SUPPORTED
+         switch (png_ptr->user_chunk_cache_max)
+         {
+            case 2:
+               png_ptr->user_chunk_cache_max = 1;
+               png_chunk_benign_error(png_ptr, "no space in chunk cache");
+               /* FALL THROUGH */
+            case 1:
+               /* NOTE: prior to 1.6.0 this case resulted in an unknown critical
+                * chunk being skipped, now there will be a hard error below.
+                */
+               break;
+
+            default: /* not at limit */
+               --(png_ptr->user_chunk_cache_max);
+               /* FALL THROUGH */
+            case 0: /* no limit */
+#     endif /* PNG_USER_LIMITS_SUPPORTED */
+               /* Here when the limit isn't reached or when limits are compiled
+                * out; store the chunk.
+                */
+               png_set_unknown_chunks(png_ptr, info_ptr,
+                  &png_ptr->unknown_chunk, 1);
+               handled = 1;
+#     ifdef PNG_USER_LIMITS_SUPPORTED
+               break;
+         }
+#     endif
+      }
+#  else /* no store support: the chunk must be handled by the user callback */
+      PNG_UNUSED(info_ptr)
+#  endif
+
+   /* Regardless of the error handling below the cached data (if any) can be
+    * freed now.  Notice that the data is not freed if there is a png_error, but
+    * it will be freed by destroy_read_struct.
+    */
+   if (png_ptr->unknown_chunk.data != NULL)
+      png_free(png_ptr, png_ptr->unknown_chunk.data);
+   png_ptr->unknown_chunk.data = NULL;
+
+#else /* !PNG_READ_UNKNOWN_CHUNKS_SUPPORTED */
+   /* There is no support to read an unknown chunk, so just skip it. */
+   png_crc_finish(png_ptr, length);
+   PNG_UNUSED(info_ptr)
+   PNG_UNUSED(keep)
+#endif /* !PNG_READ_UNKNOWN_CHUNKS_SUPPORTED */
+
+   /* Check for unhandled critical chunks */
+   if (!handled && PNG_CHUNK_CRITICAL(png_ptr->chunk_name))
+      png_chunk_error(png_ptr, "unhandled critical chunk");
+}
+
+/* This function is called to verify that a chunk name is valid.
+ * This function can't have the "critical chunk check" incorporated
+ * into it, since in the future we will need to be able to call user
+ * functions to handle unknown critical chunks after we check that
+ * the chunk name itself is valid.
+ */
+
+/* Bit hacking: the test for an invalid byte in the 4 byte chunk name is:
+ *
+ * ((c) < 65 || (c) > 122 || ((c) > 90 && (c) < 97))
+ */
+
+void /* PRIVATE */
+png_check_chunk_name(png_structrp png_ptr, png_uint_32 chunk_name)
+{
+   int i;
+
+   png_debug(1, "in png_check_chunk_name");
+
+   for (i=1; i<=4; ++i)
+   {
+      int c = chunk_name & 0xff;
+
+      if (c < 65 || c > 122 || (c > 90 && c < 97))
+         png_chunk_error(png_ptr, "invalid chunk type");
+
+      chunk_name >>= 8;
+   }
+}
+
+/* Combines the row recently read in with the existing pixels in the row.  This
+ * routine takes care of alpha and transparency if requested.  This routine also
+ * handles the two methods of progressive display of interlaced images,
+ * depending on the 'display' value; if 'display' is true then the whole row
+ * (dp) is filled from the start by replicating the available pixels.  If
+ * 'display' is false only those pixels present in the pass are filled in.
+ */
+void /* PRIVATE */
+png_combine_row(png_const_structrp png_ptr, png_bytep dp, int display)
+{
+   unsigned int pixel_depth = png_ptr->transformed_pixel_depth;
+   png_const_bytep sp = png_ptr->row_buf + 1;
+   png_uint_32 row_width = png_ptr->width;
+   unsigned int pass = png_ptr->pass;
+   png_bytep end_ptr = 0;
+   png_byte end_byte = 0;
+   unsigned int end_mask;
+
+   png_debug(1, "in png_combine_row");
+
+   /* Added in 1.5.6: it should not be possible to enter this routine until at
+    * least one row has been read from the PNG data and transformed.
+    */
+   if (pixel_depth == 0)
+      png_error(png_ptr, "internal row logic error");
+
+   /* Added in 1.5.4: the pixel depth should match the information returned by
+    * any call to png_read_update_info at this point.  Do not continue if we got
+    * this wrong.
+    */
+   if (png_ptr->info_rowbytes != 0 && png_ptr->info_rowbytes !=
+          PNG_ROWBYTES(pixel_depth, row_width))
+      png_error(png_ptr, "internal row size calculation error");
+
+   /* Don't expect this to ever happen: */
+   if (row_width == 0)
+      png_error(png_ptr, "internal row width error");
+
+   /* Preserve the last byte in cases where only part of it will be overwritten,
+    * the multiply below may overflow, we don't care because ANSI-C guarantees
+    * we get the low bits.
+    */
+   end_mask = (pixel_depth * row_width) & 7;
+   if (end_mask != 0)
+   {
+      /* end_ptr == NULL is a flag to say do nothing */
+      end_ptr = dp + PNG_ROWBYTES(pixel_depth, row_width) - 1;
+      end_byte = *end_ptr;
+#     ifdef PNG_READ_PACKSWAP_SUPPORTED
+         if (png_ptr->transformations & PNG_PACKSWAP) /* little-endian byte */
+            end_mask = 0xff << end_mask;
+
+         else /* big-endian byte */
+#     endif
+         end_mask = 0xff >> end_mask;
+      /* end_mask is now the bits to *keep* from the destination row */
+   }
+
+   /* For non-interlaced images this reduces to a memcpy(). A memcpy()
+    * will also happen if interlacing isn't supported or if the application
+    * does not call png_set_interlace_handling().  In the latter cases the
+    * caller just gets a sequence of the unexpanded rows from each interlace
+    * pass.
+    */
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   if (png_ptr->interlaced && (png_ptr->transformations & PNG_INTERLACE) &&
+      pass < 6 && (display == 0 ||
+      /* The following copies everything for 'display' on passes 0, 2 and 4. */
+      (display == 1 && (pass & 1) != 0)))
+   {
+      /* Narrow images may have no bits in a pass; the caller should handle
+       * this, but this test is cheap:
+       */
+      if (row_width <= PNG_PASS_START_COL(pass))
+         return;
+
+      if (pixel_depth < 8)
+      {
+         /* For pixel depths up to 4 bpp the 8-pixel mask can be expanded to fit
+          * into 32 bits, then a single loop over the bytes using the four byte
+          * values in the 32-bit mask can be used.  For the 'display' option the
+          * expanded mask may also not require any masking within a byte.  To
+          * make this work the PACKSWAP option must be taken into account - it
+          * simply requires the pixels to be reversed in each byte.
+          *
+          * The 'regular' case requires a mask for each of the first 6 passes,
+          * the 'display' case does a copy for the even passes in the range
+          * 0..6.  This has already been handled in the test above.
+          *
+          * The masks are arranged as four bytes with the first byte to use in
+          * the lowest bits (little-endian) regardless of the order (PACKSWAP or
+          * not) of the pixels in each byte.
+          *
+          * NOTE: the whole of this logic depends on the caller of this function
+          * only calling it on rows appropriate to the pass.  This function only
+          * understands the 'x' logic; the 'y' logic is handled by the caller.
+          *
+          * The following defines allow generation of compile time constant bit
+          * masks for each pixel depth and each possibility of swapped or not
+          * swapped bytes.  Pass 'p' is in the range 0..6; 'x', a pixel index,
+          * is in the range 0..7; and the result is 1 if the pixel is to be
+          * copied in the pass, 0 if not.  'S' is for the sparkle method, 'B'
+          * for the block method.
+          *
+          * With some compilers a compile time expression of the general form:
+          *
+          *    (shift >= 32) ? (a >> (shift-32)) : (b >> shift)
+          *
+          * Produces warnings with values of 'shift' in the range 33 to 63
+          * because the right hand side of the ?: expression is evaluated by
+          * the compiler even though it isn't used.  Microsoft Visual C (various
+          * versions) and the Intel C compiler are known to do this.  To avoid
+          * this the following macros are used in 1.5.6.  This is a temporary
+          * solution to avoid destabilizing the code during the release process.
+          */
+#        if PNG_USE_COMPILE_TIME_MASKS
+#           define PNG_LSR(x,s) ((x)>>((s) & 0x1f))
+#           define PNG_LSL(x,s) ((x)<<((s) & 0x1f))
+#        else
+#           define PNG_LSR(x,s) ((x)>>(s))
+#           define PNG_LSL(x,s) ((x)<<(s))
+#        endif
+#        define S_COPY(p,x) (((p)<4 ? PNG_LSR(0x80088822,(3-(p))*8+(7-(x))) :\
+           PNG_LSR(0xaa55ff00,(7-(p))*8+(7-(x)))) & 1)
+#        define B_COPY(p,x) (((p)<4 ? PNG_LSR(0xff0fff33,(3-(p))*8+(7-(x))) :\
+           PNG_LSR(0xff55ff00,(7-(p))*8+(7-(x)))) & 1)
+
+         /* Return a mask for pass 'p' pixel 'x' at depth 'd'.  The mask is
+          * little endian - the first pixel is at bit 0 - however the extra
+          * parameter 's' can be set to cause the mask position to be swapped
+          * within each byte, to match the PNG format.  This is done by XOR of
+          * the shift with 7, 6 or 4 for bit depths 1, 2 and 4.
+          */
+#        define PIXEL_MASK(p,x,d,s) \
+            (PNG_LSL(((PNG_LSL(1U,(d)))-1),(((x)*(d))^((s)?8-(d):0))))
+
+         /* Hence generate the appropriate 'block' or 'sparkle' pixel copy mask.
+          */
+#        define S_MASKx(p,x,d,s) (S_COPY(p,x)?PIXEL_MASK(p,x,d,s):0)
+#        define B_MASKx(p,x,d,s) (B_COPY(p,x)?PIXEL_MASK(p,x,d,s):0)
+
+         /* Combine 8 of these to get the full mask.  For the 1-bpp and 2-bpp
+          * cases the result needs replicating, for the 4-bpp case the above
+          * generates a full 32 bits.
+          */
+#        define MASK_EXPAND(m,d) ((m)*((d)==1?0x01010101:((d)==2?0x00010001:1)))
+
+#        define S_MASK(p,d,s) MASK_EXPAND(S_MASKx(p,0,d,s) + S_MASKx(p,1,d,s) +\
+            S_MASKx(p,2,d,s) + S_MASKx(p,3,d,s) + S_MASKx(p,4,d,s) +\
+            S_MASKx(p,5,d,s) + S_MASKx(p,6,d,s) + S_MASKx(p,7,d,s), d)
+
+#        define B_MASK(p,d,s) MASK_EXPAND(B_MASKx(p,0,d,s) + B_MASKx(p,1,d,s) +\
+            B_MASKx(p,2,d,s) + B_MASKx(p,3,d,s) + B_MASKx(p,4,d,s) +\
+            B_MASKx(p,5,d,s) + B_MASKx(p,6,d,s) + B_MASKx(p,7,d,s), d)
+
+#if PNG_USE_COMPILE_TIME_MASKS
+         /* Utility macros to construct all the masks for a depth/swap
+          * combination.  The 's' parameter says whether the format is PNG
+          * (big endian bytes) or not.  Only the three odd-numbered passes are
+          * required for the display/block algorithm.
+          */
+#        define S_MASKS(d,s) { S_MASK(0,d,s), S_MASK(1,d,s), S_MASK(2,d,s),\
+            S_MASK(3,d,s), S_MASK(4,d,s), S_MASK(5,d,s) }
+
+#        define B_MASKS(d,s) { B_MASK(1,d,s), S_MASK(3,d,s), S_MASK(5,d,s) }
+
+#        define DEPTH_INDEX(d) ((d)==1?0:((d)==2?1:2))
+
+         /* Hence the pre-compiled masks indexed by PACKSWAP (or not), depth and
+          * then pass:
+          */
+         static PNG_CONST png_uint_32 row_mask[2/*PACKSWAP*/][3/*depth*/][6] =
+         {
+            /* Little-endian byte masks for PACKSWAP */
+            { S_MASKS(1,0), S_MASKS(2,0), S_MASKS(4,0) },
+            /* Normal (big-endian byte) masks - PNG format */
+            { S_MASKS(1,1), S_MASKS(2,1), S_MASKS(4,1) }
+         };
+
+         /* display_mask has only three entries for the odd passes, so index by
+          * pass>>1.
+          */
+         static PNG_CONST png_uint_32 display_mask[2][3][3] =
+         {
+            /* Little-endian byte masks for PACKSWAP */
+            { B_MASKS(1,0), B_MASKS(2,0), B_MASKS(4,0) },
+            /* Normal (big-endian byte) masks - PNG format */
+            { B_MASKS(1,1), B_MASKS(2,1), B_MASKS(4,1) }
+         };
+
+#        define MASK(pass,depth,display,png)\
+            ((display)?display_mask[png][DEPTH_INDEX(depth)][pass>>1]:\
+               row_mask[png][DEPTH_INDEX(depth)][pass])
+
+#else /* !PNG_USE_COMPILE_TIME_MASKS */
+         /* This is the runtime alternative: it seems unlikely that this will
+          * ever be either smaller or faster than the compile time approach.
+          */
+#        define MASK(pass,depth,display,png)\
+            ((display)?B_MASK(pass,depth,png):S_MASK(pass,depth,png))
+#endif /* !PNG_USE_COMPILE_TIME_MASKS */
+
+         /* Use the appropriate mask to copy the required bits.  In some cases
+          * the byte mask will be 0 or 0xff, optimize these cases.  row_width is
+          * the number of pixels, but the code copies bytes, so it is necessary
+          * to special case the end.
+          */
+         png_uint_32 pixels_per_byte = 8 / pixel_depth;
+         png_uint_32 mask;
+
+#        ifdef PNG_READ_PACKSWAP_SUPPORTED
+            if (png_ptr->transformations & PNG_PACKSWAP)
+               mask = MASK(pass, pixel_depth, display, 0);
+
+            else
+#        endif
+            mask = MASK(pass, pixel_depth, display, 1);
+
+         for (;;)
+         {
+            png_uint_32 m;
+
+            /* It doesn't matter in the following if png_uint_32 has more than
+             * 32 bits because the high bits always match those in m<<24; it is,
+             * however, essential to use OR here, not +, because of this.
+             */
+            m = mask;
+            mask = (m >> 8) | (m << 24); /* rotate right to good compilers */
+            m &= 0xff;
+
+            if (m != 0) /* something to copy */
+            {
+               if (m != 0xff)
+                  *dp = (png_byte)((*dp & ~m) | (*sp & m));
+               else
+                  *dp = *sp;
+            }
+
+            /* NOTE: this may overwrite the last byte with garbage if the image
+             * is not an exact number of bytes wide; libpng has always done
+             * this.
+             */
+            if (row_width <= pixels_per_byte)
+               break; /* May need to restore part of the last byte */
+
+            row_width -= pixels_per_byte;
+            ++dp;
+            ++sp;
+         }
+      }
+
+      else /* pixel_depth >= 8 */
+      {
+         unsigned int bytes_to_copy, bytes_to_jump;
+
+         /* Validate the depth - it must be a multiple of 8 */
+         if (pixel_depth & 7)
+            png_error(png_ptr, "invalid user transform pixel depth");
+
+         pixel_depth >>= 3; /* now in bytes */
+         row_width *= pixel_depth;
+
+         /* Regardless of pass number the Adam 7 interlace always results in a
+          * fixed number of pixels to copy then to skip.  There may be a
+          * different number of pixels to skip at the start though.
+          */
+         {
+            unsigned int offset = PNG_PASS_START_COL(pass) * pixel_depth;
+
+            row_width -= offset;
+            dp += offset;
+            sp += offset;
+         }
+
+         /* Work out the bytes to copy. */
+         if (display)
+         {
+            /* When doing the 'block' algorithm the pixel in the pass gets
+             * replicated to adjacent pixels.  This is why the even (0,2,4,6)
+             * passes are skipped above - the entire expanded row is copied.
+             */
+            bytes_to_copy = (1<<((6-pass)>>1)) * pixel_depth;
+
+            /* But don't allow this number to exceed the actual row width. */
+            if (bytes_to_copy > row_width)
+               bytes_to_copy = row_width;
+         }
+
+         else /* normal row; Adam7 only ever gives us one pixel to copy. */
+            bytes_to_copy = pixel_depth;
+
+         /* In Adam7 there is a constant offset between where the pixels go. */
+         bytes_to_jump = PNG_PASS_COL_OFFSET(pass) * pixel_depth;
+
+         /* And simply copy these bytes.  Some optimization is possible here,
+          * depending on the value of 'bytes_to_copy'.  Special case the low
+          * byte counts, which we know to be frequent.
+          *
+          * Notice that these cases all 'return' rather than 'break' - this
+          * avoids an unnecessary test on whether to restore the last byte
+          * below.
+          */
+         switch (bytes_to_copy)
+         {
+            case 1:
+               for (;;)
+               {
+                  *dp = *sp;
+
+                  if (row_width <= bytes_to_jump)
+                     return;
+
+                  dp += bytes_to_jump;
+                  sp += bytes_to_jump;
+                  row_width -= bytes_to_jump;
+               }
+
+            case 2:
+               /* There is a possibility of a partial copy at the end here; this
+                * slows the code down somewhat.
+                */
+               do
+               {
+                  dp[0] = sp[0], dp[1] = sp[1];
+
+                  if (row_width <= bytes_to_jump)
+                     return;
+
+                  sp += bytes_to_jump;
+                  dp += bytes_to_jump;
+                  row_width -= bytes_to_jump;
+               }
+               while (row_width > 1);
+
+               /* And there can only be one byte left at this point: */
+               *dp = *sp;
+               return;
+
+            case 3:
+               /* This can only be the RGB case, so each copy is exactly one
+                * pixel and it is not necessary to check for a partial copy.
+                */
+               for(;;)
+               {
+                  dp[0] = sp[0], dp[1] = sp[1], dp[2] = sp[2];
+
+                  if (row_width <= bytes_to_jump)
+                     return;
+
+                  sp += bytes_to_jump;
+                  dp += bytes_to_jump;
+                  row_width -= bytes_to_jump;
+               }
+
+            default:
+#if PNG_ALIGN_TYPE != PNG_ALIGN_NONE
+               /* Check for double byte alignment and, if possible, use a
+                * 16-bit copy.  Don't attempt this for narrow images - ones that
+                * are less than an interlace panel wide.  Don't attempt it for
+                * wide bytes_to_copy either - use the memcpy there.
+                */
+               if (bytes_to_copy < 16 /*else use memcpy*/ &&
+                  png_isaligned(dp, png_uint_16) &&
+                  png_isaligned(sp, png_uint_16) &&
+                  bytes_to_copy % (sizeof (png_uint_16)) == 0 &&
+                  bytes_to_jump % (sizeof (png_uint_16)) == 0)
+               {
+                  /* Everything is aligned for png_uint_16 copies, but try for
+                   * png_uint_32 first.
+                   */
+                  if (png_isaligned(dp, png_uint_32) &&
+                     png_isaligned(sp, png_uint_32) &&
+                     bytes_to_copy % (sizeof (png_uint_32)) == 0 &&
+                     bytes_to_jump % (sizeof (png_uint_32)) == 0)
+                  {
+                     png_uint_32p dp32 = png_aligncast(png_uint_32p,dp);
+                     png_const_uint_32p sp32 = png_aligncastconst(
+                        png_const_uint_32p, sp);
+                     size_t skip = (bytes_to_jump-bytes_to_copy) /
+                        (sizeof (png_uint_32));
+
+                     do
+                     {
+                        size_t c = bytes_to_copy;
+                        do
+                        {
+                           *dp32++ = *sp32++;
+                           c -= (sizeof (png_uint_32));
+                        }
+                        while (c > 0);
+
+                        if (row_width <= bytes_to_jump)
+                           return;
+
+                        dp32 += skip;
+                        sp32 += skip;
+                        row_width -= bytes_to_jump;
+                     }
+                     while (bytes_to_copy <= row_width);
+
+                     /* Get to here when the row_width truncates the final copy.
+                      * There will be 1-3 bytes left to copy, so don't try the
+                      * 16-bit loop below.
+                      */
+                     dp = (png_bytep)dp32;
+                     sp = (png_const_bytep)sp32;
+                     do
+                        *dp++ = *sp++;
+                     while (--row_width > 0);
+                     return;
+                  }
+
+                  /* Else do it in 16-bit quantities, but only if the size is
+                   * not too large.
+                   */
+                  else
+                  {
+                     png_uint_16p dp16 = png_aligncast(png_uint_16p, dp);
+                     png_const_uint_16p sp16 = png_aligncastconst(
+                        png_const_uint_16p, sp);
+                     size_t skip = (bytes_to_jump-bytes_to_copy) /
+                        (sizeof (png_uint_16));
+
+                     do
+                     {
+                        size_t c = bytes_to_copy;
+                        do
+                        {
+                           *dp16++ = *sp16++;
+                           c -= (sizeof (png_uint_16));
+                        }
+                        while (c > 0);
+
+                        if (row_width <= bytes_to_jump)
+                           return;
+
+                        dp16 += skip;
+                        sp16 += skip;
+                        row_width -= bytes_to_jump;
+                     }
+                     while (bytes_to_copy <= row_width);
+
+                     /* End of row - 1 byte left, bytes_to_copy > row_width: */
+                     dp = (png_bytep)dp16;
+                     sp = (png_const_bytep)sp16;
+                     do
+                        *dp++ = *sp++;
+                     while (--row_width > 0);
+                     return;
+                  }
+               }
+#endif /* PNG_ALIGN_ code */
+
+               /* The true default - use a memcpy: */
+               for (;;)
+               {
+                  memcpy(dp, sp, bytes_to_copy);
+
+                  if (row_width <= bytes_to_jump)
+                     return;
+
+                  sp += bytes_to_jump;
+                  dp += bytes_to_jump;
+                  row_width -= bytes_to_jump;
+                  if (bytes_to_copy > row_width)
+                     bytes_to_copy = row_width;
+               }
+         }
+
+         /* NOT REACHED*/
+      } /* pixel_depth >= 8 */
+
+      /* Here if pixel_depth < 8 to check 'end_ptr' below. */
+   }
+   else
+#endif
+
+   /* If here then the switch above wasn't used so just memcpy the whole row
+    * from the temporary row buffer (notice that this overwrites the end of the
+    * destination row if it is a partial byte.)
+    */
+   memcpy(dp, sp, PNG_ROWBYTES(pixel_depth, row_width));
+
+   /* Restore the overwritten bits from the last byte if necessary. */
+   if (end_ptr != NULL)
+      *end_ptr = (png_byte)((end_byte & end_mask) | (*end_ptr & ~end_mask));
+}
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+void /* PRIVATE */
+png_do_read_interlace(png_row_infop row_info, png_bytep row, int pass,
+   png_uint_32 transformations /* Because these may affect the byte layout */)
+{
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+   /* Offset to next interlace block */
+   static PNG_CONST int png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   png_debug(1, "in png_do_read_interlace");
+   if (row != NULL && row_info != NULL)
+   {
+      png_uint_32 final_width;
+
+      final_width = row_info->width * png_pass_inc[pass];
+
+      switch (row_info->pixel_depth)
+      {
+         case 1:
+         {
+            png_bytep sp = row + (png_size_t)((row_info->width - 1) >> 3);
+            png_bytep dp = row + (png_size_t)((final_width - 1) >> 3);
+            int sshift, dshift;
+            int s_start, s_end, s_inc;
+            int jstop = png_pass_inc[pass];
+            png_byte v;
+            png_uint_32 i;
+            int j;
+
+#ifdef PNG_READ_PACKSWAP_SUPPORTED
+            if (transformations & PNG_PACKSWAP)
+            {
+                sshift = (int)((row_info->width + 7) & 0x07);
+                dshift = (int)((final_width + 7) & 0x07);
+                s_start = 7;
+                s_end = 0;
+                s_inc = -1;
+            }
+
+            else
+#endif
+            {
+                sshift = 7 - (int)((row_info->width + 7) & 0x07);
+                dshift = 7 - (int)((final_width + 7) & 0x07);
+                s_start = 0;
+                s_end = 7;
+                s_inc = 1;
+            }
+
+            for (i = 0; i < row_info->width; i++)
+            {
+               v = (png_byte)((*sp >> sshift) & 0x01);
+               for (j = 0; j < jstop; j++)
+               {
+                  unsigned int tmp = *dp & (0x7f7f >> (7 - dshift));
+                  tmp |= v << dshift;
+                  *dp = (png_byte)(tmp & 0xff);
+
+                  if (dshift == s_end)
+                  {
+                     dshift = s_start;
+                     dp--;
+                  }
+
+                  else
+                     dshift += s_inc;
+               }
+
+               if (sshift == s_end)
+               {
+                  sshift = s_start;
+                  sp--;
+               }
+
+               else
+                  sshift += s_inc;
+            }
+            break;
+         }
+
+         case 2:
+         {
+            png_bytep sp = row + (png_uint_32)((row_info->width - 1) >> 2);
+            png_bytep dp = row + (png_uint_32)((final_width - 1) >> 2);
+            int sshift, dshift;
+            int s_start, s_end, s_inc;
+            int jstop = png_pass_inc[pass];
+            png_uint_32 i;
+
+#ifdef PNG_READ_PACKSWAP_SUPPORTED
+            if (transformations & PNG_PACKSWAP)
+            {
+               sshift = (int)(((row_info->width + 3) & 0x03) << 1);
+               dshift = (int)(((final_width + 3) & 0x03) << 1);
+               s_start = 6;
+               s_end = 0;
+               s_inc = -2;
+            }
+
+            else
+#endif
+            {
+               sshift = (int)((3 - ((row_info->width + 3) & 0x03)) << 1);
+               dshift = (int)((3 - ((final_width + 3) & 0x03)) << 1);
+               s_start = 0;
+               s_end = 6;
+               s_inc = 2;
+            }
+
+            for (i = 0; i < row_info->width; i++)
+            {
+               png_byte v;
+               int j;
+
+               v = (png_byte)((*sp >> sshift) & 0x03);
+               for (j = 0; j < jstop; j++)
+               {
+                  unsigned int tmp = *dp & (0x3f3f >> (6 - dshift));
+                  tmp |= v << dshift;
+                  *dp = (png_byte)(tmp & 0xff);
+
+                  if (dshift == s_end)
+                  {
+                     dshift = s_start;
+                     dp--;
+                  }
+
+                  else
+                     dshift += s_inc;
+               }
+
+               if (sshift == s_end)
+               {
+                  sshift = s_start;
+                  sp--;
+               }
+
+               else
+                  sshift += s_inc;
+            }
+            break;
+         }
+
+         case 4:
+         {
+            png_bytep sp = row + (png_size_t)((row_info->width - 1) >> 1);
+            png_bytep dp = row + (png_size_t)((final_width - 1) >> 1);
+            int sshift, dshift;
+            int s_start, s_end, s_inc;
+            png_uint_32 i;
+            int jstop = png_pass_inc[pass];
+
+#ifdef PNG_READ_PACKSWAP_SUPPORTED
+            if (transformations & PNG_PACKSWAP)
+            {
+               sshift = (int)(((row_info->width + 1) & 0x01) << 2);
+               dshift = (int)(((final_width + 1) & 0x01) << 2);
+               s_start = 4;
+               s_end = 0;
+               s_inc = -4;
+            }
+
+            else
+#endif
+            {
+               sshift = (int)((1 - ((row_info->width + 1) & 0x01)) << 2);
+               dshift = (int)((1 - ((final_width + 1) & 0x01)) << 2);
+               s_start = 0;
+               s_end = 4;
+               s_inc = 4;
+            }
+
+            for (i = 0; i < row_info->width; i++)
+            {
+               png_byte v = (png_byte)((*sp >> sshift) & 0x0f);
+               int j;
+
+               for (j = 0; j < jstop; j++)
+               {
+                  unsigned int tmp = *dp & (0xf0f >> (4 - dshift));
+                  tmp |= v << dshift;
+                  *dp = (png_byte)(tmp & 0xff);
+
+                  if (dshift == s_end)
+                  {
+                     dshift = s_start;
+                     dp--;
+                  }
+
+                  else
+                     dshift += s_inc;
+               }
+
+               if (sshift == s_end)
+               {
+                  sshift = s_start;
+                  sp--;
+               }
+
+               else
+                  sshift += s_inc;
+            }
+            break;
+         }
+
+         default:
+         {
+            png_size_t pixel_bytes = (row_info->pixel_depth >> 3);
+
+            png_bytep sp = row + (png_size_t)(row_info->width - 1)
+                * pixel_bytes;
+
+            png_bytep dp = row + (png_size_t)(final_width - 1) * pixel_bytes;
+
+            int jstop = png_pass_inc[pass];
+            png_uint_32 i;
+
+            for (i = 0; i < row_info->width; i++)
+            {
+               png_byte v[8]; /* SAFE; pixel_depth does not exceed 64 */
+               int j;
+
+               memcpy(v, sp, pixel_bytes);
+
+               for (j = 0; j < jstop; j++)
+               {
+                  memcpy(dp, v, pixel_bytes);
+                  dp -= pixel_bytes;
+               }
+
+               sp -= pixel_bytes;
+            }
+            break;
+         }
+      }
+
+      row_info->width = final_width;
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth, final_width);
+   }
+#ifndef PNG_READ_PACKSWAP_SUPPORTED
+   PNG_UNUSED(transformations)  /* Silence compiler warning */
+#endif
+}
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+
+static void
+png_read_filter_row_sub(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_size_t i;
+   png_size_t istop = row_info->rowbytes;
+   unsigned int bpp = (row_info->pixel_depth + 7) >> 3;
+   png_bytep rp = row + bpp;
+
+   PNG_UNUSED(prev_row)
+
+   for (i = bpp; i < istop; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff);
+      rp++;
+   }
+}
+
+static void
+png_read_filter_row_up(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_size_t i;
+   png_size_t istop = row_info->rowbytes;
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+
+   for (i = 0; i < istop; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff);
+      rp++;
+   }
+}
+
+static void
+png_read_filter_row_avg(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_size_t i;
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   unsigned int bpp = (row_info->pixel_depth + 7) >> 3;
+   png_size_t istop = row_info->rowbytes - bpp;
+
+   for (i = 0; i < bpp; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) +
+         ((int)(*pp++) / 2 )) & 0xff);
+
+      rp++;
+   }
+
+   for (i = 0; i < istop; i++)
+   {
+      *rp = (png_byte)(((int)(*rp) +
+         (int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff);
+
+      rp++;
+   }
+}
+
+static void
+png_read_filter_row_paeth_1byte_pixel(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp_end = row + row_info->rowbytes;
+   int a, c;
+
+   /* First pixel/byte */
+   c = *prev_row++;
+   a = *row + c;
+   *row++ = (png_byte)a;
+
+   /* Remainder */
+   while (row < rp_end)
+   {
+      int b, pa, pb, pc, p;
+
+      a &= 0xff; /* From previous iteration or start */
+      b = *prev_row++;
+
+      p = b - c;
+      pc = a - c;
+
+#     ifdef PNG_USE_ABS
+         pa = abs(p);
+         pb = abs(pc);
+         pc = abs(p + pc);
+#     else
+         pa = p < 0 ? -p : p;
+         pb = pc < 0 ? -pc : pc;
+         pc = (p + pc) < 0 ? -(p + pc) : p + pc;
+#     endif
+
+      /* Find the best predictor, the least of pa, pb, pc favoring the earlier
+       * ones in the case of a tie.
+       */
+      if (pb < pa) pa = pb, a = b;
+      if (pc < pa) a = c;
+
+      /* Calculate the current pixel in a, and move the previous row pixel to c
+       * for the next time round the loop
+       */
+      c = b;
+      a += *row;
+      *row++ = (png_byte)a;
+   }
+}
+
+static void
+png_read_filter_row_paeth_multibyte_pixel(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   int bpp = (row_info->pixel_depth + 7) >> 3;
+   png_bytep rp_end = row + bpp;
+
+   /* Process the first pixel in the row completely (this is the same as 'up'
+    * because there is only one candidate predictor for the first row).
+    */
+   while (row < rp_end)
+   {
+      int a = *row + *prev_row++;
+      *row++ = (png_byte)a;
+   }
+
+   /* Remainder */
+   rp_end += row_info->rowbytes - bpp;
+
+   while (row < rp_end)
+   {
+      int a, b, c, pa, pb, pc, p;
+
+      c = *(prev_row - bpp);
+      a = *(row - bpp);
+      b = *prev_row++;
+
+      p = b - c;
+      pc = a - c;
+
+#     ifdef PNG_USE_ABS
+         pa = abs(p);
+         pb = abs(pc);
+         pc = abs(p + pc);
+#     else
+         pa = p < 0 ? -p : p;
+         pb = pc < 0 ? -pc : pc;
+         pc = (p + pc) < 0 ? -(p + pc) : p + pc;
+#     endif
+
+      if (pb < pa) pa = pb, a = b;
+      if (pc < pa) a = c;
+
+      c = b;
+      a += *row;
+      *row++ = (png_byte)a;
+   }
+}
+
+static void
+png_init_filter_functions(png_structrp pp)
+   /* This function is called once for every PNG image (except for PNG images
+    * that only use PNG_FILTER_VALUE_NONE for all rows) to set the
+    * implementations required to reverse the filtering of PNG rows.  Reversing
+    * the filter is the first transformation performed on the row data.  It is
+    * performed in place, therefore an implementation can be selected based on
+    * the image pixel format.  If the implementation depends on image width then
+    * take care to ensure that it works correctly if the image is interlaced -
+    * interlacing causes the actual row width to vary.
+    */
+{
+   unsigned int bpp = (pp->pixel_depth + 7) >> 3;
+
+   pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub;
+   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up;
+   pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg;
+   if (bpp == 1)
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth_1byte_pixel;
+   else
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth_multibyte_pixel;
+
+#ifdef PNG_FILTER_OPTIMIZATIONS
+   /* To use this define PNG_FILTER_OPTIMIZATIONS as the name of a function to
+    * call to install hardware optimizations for the above functions; simply
+    * replace whatever elements of the pp->read_filter[] array with a hardware
+    * specific (or, for that matter, generic) optimization.
+    *
+    * To see an example of this examine what configure.ac does when
+    * --enable-arm-neon is specified on the command line.
+    */
+   PNG_FILTER_OPTIMIZATIONS(pp, bpp);
+#endif
+}
+
+void /* PRIVATE */
+png_read_filter_row(png_structrp pp, png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row, int filter)
+{
+   /* OPTIMIZATION: DO NOT MODIFY THIS FUNCTION, instead #define
+    * PNG_FILTER_OPTIMIZATIONS to a function that overrides the generic
+    * implementations.  See png_init_filter_functions above.
+    */
+   if (filter > PNG_FILTER_VALUE_NONE && filter < PNG_FILTER_VALUE_LAST)
+   {
+      if (pp->read_filter[0] == NULL)
+         png_init_filter_functions(pp);
+
+      pp->read_filter[filter-1](row_info, row, prev_row);
+   }
+}
+
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+void /* PRIVATE */
+png_read_IDAT_data(png_structrp png_ptr, png_bytep output,
+   png_alloc_size_t avail_out)
+{
+   /* Loop reading IDATs and decompressing the result into output[avail_out] */
+   png_ptr->zstream.next_out = output;
+   png_ptr->zstream.avail_out = 0; /* safety: set below */
+
+   if (output == NULL)
+      avail_out = 0;
+
+   do
+   {
+      int ret;
+      png_byte tmpbuf[PNG_INFLATE_BUF_SIZE];
+
+      if (png_ptr->zstream.avail_in == 0)
+      {
+         uInt avail_in;
+         png_bytep buffer;
+
+         while (png_ptr->idat_size == 0)
+         {
+            png_crc_finish(png_ptr, 0);
+
+            png_ptr->idat_size = png_read_chunk_header(png_ptr);
+            /* This is an error even in the 'check' case because the code just
+             * consumed a non-IDAT header.
+             */
+            if (png_ptr->chunk_name != png_IDAT)
+               png_error(png_ptr, "Not enough image data");
+         }
+
+         avail_in = png_ptr->IDAT_read_size;
+
+         if (avail_in > png_ptr->idat_size)
+            avail_in = (uInt)png_ptr->idat_size;
+
+         /* A PNG with a gradually increasing IDAT size will defeat this attempt
+          * to minimize memory usage by causing lots of re-allocs, but
+          * realistically doing IDAT_read_size re-allocs is not likely to be a
+          * big problem.
+          */
+         buffer = png_read_buffer(png_ptr, avail_in, 0/*error*/);
+
+         png_crc_read(png_ptr, buffer, avail_in);
+         png_ptr->idat_size -= avail_in;
+
+         png_ptr->zstream.next_in = buffer;
+         png_ptr->zstream.avail_in = avail_in;
+      }
+
+      /* And set up the output side. */
+      if (output != NULL) /* standard read */
+      {
+         uInt out = ZLIB_IO_MAX;
+
+         if (out > avail_out)
+            out = (uInt)avail_out;
+
+         avail_out -= out;
+         png_ptr->zstream.avail_out = out;
+      }
+
+      else /* after last row, checking for end */
+      {
+         png_ptr->zstream.next_out = tmpbuf;
+         png_ptr->zstream.avail_out = (sizeof tmpbuf);
+      }
+
+      /* Use NO_FLUSH; this gives zlib the maximum opportunity to optimize the
+       * process.  If the LZ stream is truncated the sequential reader will
+       * terminally damage the stream, above, by reading the chunk header of the
+       * following chunk (it then exits with png_error).
+       *
+       * TODO: deal more elegantly with truncated IDAT lists.
+       */
+      ret = inflate(&png_ptr->zstream, Z_NO_FLUSH);
+
+      /* Take the unconsumed output back. */
+      if (output != NULL)
+         avail_out += png_ptr->zstream.avail_out;
+
+      else /* avail_out counts the extra bytes */
+         avail_out += (sizeof tmpbuf) - png_ptr->zstream.avail_out;
+
+      png_ptr->zstream.avail_out = 0;
+
+      if (ret == Z_STREAM_END)
+      {
+         /* Do this for safety; we won't read any more into this row. */
+         png_ptr->zstream.next_out = NULL;
+
+         png_ptr->mode |= PNG_AFTER_IDAT;
+         png_ptr->flags |= PNG_FLAG_ZSTREAM_ENDED;
+
+         if (png_ptr->zstream.avail_in > 0 || png_ptr->idat_size > 0)
+            png_chunk_benign_error(png_ptr, "Extra compressed data");
+         break;
+      }
+
+      if (ret != Z_OK)
+      {
+         png_zstream_error(png_ptr, ret);
+
+         if (output != NULL)
+            png_chunk_error(png_ptr, png_ptr->zstream.msg);
+
+         else /* checking */
+         {
+            png_chunk_benign_error(png_ptr, png_ptr->zstream.msg);
+            return;
+         }
+      }
+   } while (avail_out > 0);
+
+   if (avail_out > 0)
+   {
+      /* The stream ended before the image; this is the same as too few IDATs so
+       * should be handled the same way.
+       */
+      if (output != NULL)
+         png_error(png_ptr, "Not enough image data");
+
+      else /* the deflate stream contained extra data */
+         png_chunk_benign_error(png_ptr, "Too much image data");
+   }
+}
+
+void /* PRIVATE */
+png_read_finish_IDAT(png_structrp png_ptr)
+{
+   /* We don't need any more data and the stream should have ended, however the
+    * LZ end code may actually not have been processed.  In this case we must
+    * read it otherwise stray unread IDAT data or, more likely, an IDAT chunk
+    * may still remain to be consumed.
+    */
+   if (!(png_ptr->flags & PNG_FLAG_ZSTREAM_ENDED))
+   {
+      /* The NULL causes png_read_IDAT_data to swallow any remaining bytes in
+       * the compressed stream, but the stream may be damaged too, so even after
+       * this call we may need to terminate the zstream ownership.
+       */
+      png_read_IDAT_data(png_ptr, NULL, 0);
+      png_ptr->zstream.next_out = NULL; /* safety */
+
+      /* Now clear everything out for safety; the following may not have been
+       * done.
+       */
+      if (!(png_ptr->flags & PNG_FLAG_ZSTREAM_ENDED))
+      {
+         png_ptr->mode |= PNG_AFTER_IDAT;
+         png_ptr->flags |= PNG_FLAG_ZSTREAM_ENDED;
+      }
+   }
+
+   /* If the zstream has not been released do it now *and* terminate the reading
+    * of the final IDAT chunk.
+    */
+   if (png_ptr->zowner == png_IDAT)
+   {
+      /* Always do this; the pointers otherwise point into the read buffer. */
+      png_ptr->zstream.next_in = NULL;
+      png_ptr->zstream.avail_in = 0;
+
+      /* Now we no longer own the zstream. */
+      png_ptr->zowner = 0;
+
+      /* The slightly weird semantics of the sequential IDAT reading is that we
+       * are always in or at the end of an IDAT chunk, so we always need to do a
+       * crc_finish here.  If idat_size is non-zero we also need to read the
+       * spurious bytes at the end of the chunk now.
+       */
+      (void)png_crc_finish(png_ptr, png_ptr->idat_size);
+   }
+}
+
+void /* PRIVATE */
+png_read_finish_row(png_structrp png_ptr)
+{
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   /* Start of interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_ystart[7] = {0, 0, 4, 0, 2, 0, 1};
+
+   /* Offset to next interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_yinc[7] = {8, 8, 8, 4, 4, 2, 2};
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+
+   png_debug(1, "in png_read_finish_row");
+   png_ptr->row_number++;
+   if (png_ptr->row_number < png_ptr->num_rows)
+      return;
+
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   if (png_ptr->interlaced)
+   {
+      png_ptr->row_number = 0;
+
+      /* TO DO: don't do this if prev_row isn't needed (requires
+       * read-ahead of the next row's filter byte.
+       */
+      memset(png_ptr->prev_row, 0, png_ptr->rowbytes + 1);
+
+      do
+      {
+         png_ptr->pass++;
+
+         if (png_ptr->pass >= 7)
+            break;
+
+         png_ptr->iwidth = (png_ptr->width +
+            png_pass_inc[png_ptr->pass] - 1 -
+            png_pass_start[png_ptr->pass]) /
+            png_pass_inc[png_ptr->pass];
+
+         if (!(png_ptr->transformations & PNG_INTERLACE))
+         {
+            png_ptr->num_rows = (png_ptr->height +
+                png_pass_yinc[png_ptr->pass] - 1 -
+                png_pass_ystart[png_ptr->pass]) /
+                png_pass_yinc[png_ptr->pass];
+         }
+
+         else  /* if (png_ptr->transformations & PNG_INTERLACE) */
+            break; /* libpng deinterlacing sees every row */
+
+      } while (png_ptr->num_rows == 0 || png_ptr->iwidth == 0);
+
+      if (png_ptr->pass < 7)
+         return;
+   }
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+
+   /* Here after at the end of the last row of the last pass. */
+   png_read_finish_IDAT(png_ptr);
+}
+#endif /* PNG_SEQUENTIAL_READ_SUPPORTED */
+
+void /* PRIVATE */
+png_read_start_row(png_structrp png_ptr)
+{
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   /* Start of interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_ystart[7] = {0, 0, 4, 0, 2, 0, 1};
+
+   /* Offset to next interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_yinc[7] = {8, 8, 8, 4, 4, 2, 2};
+#endif
+
+   int max_pixel_depth;
+   png_size_t row_bytes;
+
+   png_debug(1, "in png_read_start_row");
+
+#ifdef PNG_READ_TRANSFORMS_SUPPORTED
+   png_init_read_transformations(png_ptr);
+#endif
+#ifdef PNG_READ_INTERLACING_SUPPORTED
+   if (png_ptr->interlaced)
+   {
+      if (!(png_ptr->transformations & PNG_INTERLACE))
+         png_ptr->num_rows = (png_ptr->height + png_pass_yinc[0] - 1 -
+             png_pass_ystart[0]) / png_pass_yinc[0];
+
+      else
+         png_ptr->num_rows = png_ptr->height;
+
+      png_ptr->iwidth = (png_ptr->width +
+          png_pass_inc[png_ptr->pass] - 1 -
+          png_pass_start[png_ptr->pass]) /
+          png_pass_inc[png_ptr->pass];
+   }
+
+   else
+#endif /* PNG_READ_INTERLACING_SUPPORTED */
+   {
+      png_ptr->num_rows = png_ptr->height;
+      png_ptr->iwidth = png_ptr->width;
+   }
+
+   max_pixel_depth = png_ptr->pixel_depth;
+
+   /* WARNING: * png_read_transform_info (pngrtran.c) performs a simpliar set of
+    * calculations to calculate the final pixel depth, then
+    * png_do_read_transforms actually does the transforms.  This means that the
+    * code which effectively calculates this value is actually repeated in three
+    * separate places.  They must all match.  Innocent changes to the order of
+    * transformations can and will break libpng in a way that causes memory
+    * overwrites.
+    *
+    * TODO: fix this.
+    */
+#ifdef PNG_READ_PACK_SUPPORTED
+   if ((png_ptr->transformations & PNG_PACK) && png_ptr->bit_depth < 8)
+      max_pixel_depth = 8;
+#endif
+
+#ifdef PNG_READ_EXPAND_SUPPORTED
+   if (png_ptr->transformations & PNG_EXPAND)
+   {
+      if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         if (png_ptr->num_trans)
+            max_pixel_depth = 32;
+
+         else
+            max_pixel_depth = 24;
+      }
+
+      else if (png_ptr->color_type == PNG_COLOR_TYPE_GRAY)
+      {
+         if (max_pixel_depth < 8)
+            max_pixel_depth = 8;
+
+         if (png_ptr->num_trans)
+            max_pixel_depth *= 2;
+      }
+
+      else if (png_ptr->color_type == PNG_COLOR_TYPE_RGB)
+      {
+         if (png_ptr->num_trans)
+         {
+            max_pixel_depth *= 4;
+            max_pixel_depth /= 3;
+         }
+      }
+   }
+#endif
+
+#ifdef PNG_READ_EXPAND_16_SUPPORTED
+   if (png_ptr->transformations & PNG_EXPAND_16)
+   {
+#     ifdef PNG_READ_EXPAND_SUPPORTED
+         /* In fact it is an error if it isn't supported, but checking is
+          * the safe way.
+          */
+         if (png_ptr->transformations & PNG_EXPAND)
+         {
+            if (png_ptr->bit_depth < 16)
+               max_pixel_depth *= 2;
+         }
+         else
+#     endif
+         png_ptr->transformations &= ~PNG_EXPAND_16;
+   }
+#endif
+
+#ifdef PNG_READ_FILLER_SUPPORTED
+   if (png_ptr->transformations & (PNG_FILLER))
+   {
+      if (png_ptr->color_type == PNG_COLOR_TYPE_GRAY)
+      {
+         if (max_pixel_depth <= 8)
+            max_pixel_depth = 16;
+
+         else
+            max_pixel_depth = 32;
+      }
+
+      else if (png_ptr->color_type == PNG_COLOR_TYPE_RGB ||
+         png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         if (max_pixel_depth <= 32)
+            max_pixel_depth = 32;
+
+         else
+            max_pixel_depth = 64;
+      }
+   }
+#endif
+
+#ifdef PNG_READ_GRAY_TO_RGB_SUPPORTED
+   if (png_ptr->transformations & PNG_GRAY_TO_RGB)
+   {
+      if (
+#ifdef PNG_READ_EXPAND_SUPPORTED
+          (png_ptr->num_trans && (png_ptr->transformations & PNG_EXPAND)) ||
+#endif
+#ifdef PNG_READ_FILLER_SUPPORTED
+          (png_ptr->transformations & (PNG_FILLER)) ||
+#endif
+          png_ptr->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+      {
+         if (max_pixel_depth <= 16)
+            max_pixel_depth = 32;
+
+         else
+            max_pixel_depth = 64;
+      }
+
+      else
+      {
+         if (max_pixel_depth <= 8)
+         {
+            if (png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+               max_pixel_depth = 32;
+
+            else
+               max_pixel_depth = 24;
+         }
+
+         else if (png_ptr->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+            max_pixel_depth = 64;
+
+         else
+            max_pixel_depth = 48;
+      }
+   }
+#endif
+
+#if defined(PNG_READ_USER_TRANSFORM_SUPPORTED) && \
+defined(PNG_USER_TRANSFORM_PTR_SUPPORTED)
+   if (png_ptr->transformations & PNG_USER_TRANSFORM)
+   {
+      int user_pixel_depth = png_ptr->user_transform_depth *
+         png_ptr->user_transform_channels;
+
+      if (user_pixel_depth > max_pixel_depth)
+         max_pixel_depth = user_pixel_depth;
+   }
+#endif
+
+   /* This value is stored in png_struct and double checked in the row read
+    * code.
+    */
+   png_ptr->maximum_pixel_depth = (png_byte)max_pixel_depth;
+   png_ptr->transformed_pixel_depth = 0; /* calculated on demand */
+
+   /* Align the width on the next larger 8 pixels.  Mainly used
+    * for interlacing
+    */
+   row_bytes = ((png_ptr->width + 7) & ~((png_uint_32)7));
+   /* Calculate the maximum bytes needed, adding a byte and a pixel
+    * for safety's sake
+    */
+   row_bytes = PNG_ROWBYTES(max_pixel_depth, row_bytes) +
+       1 + ((max_pixel_depth + 7) >> 3);
+
+#ifdef PNG_MAX_MALLOC_64K
+   if (row_bytes > (png_uint_32)65536L)
+      png_error(png_ptr, "This image requires a row greater than 64KB");
+#endif
+
+   if (row_bytes + 48 > png_ptr->old_big_row_buf_size)
+   {
+     png_free(png_ptr, png_ptr->big_row_buf);
+     png_free(png_ptr, png_ptr->big_prev_row);
+
+     if (png_ptr->interlaced)
+        png_ptr->big_row_buf = (png_bytep)png_calloc(png_ptr,
+            row_bytes + 48);
+
+     else
+        png_ptr->big_row_buf = (png_bytep)png_malloc(png_ptr, row_bytes + 48);
+
+     png_ptr->big_prev_row = (png_bytep)png_malloc(png_ptr, row_bytes + 48);
+
+#ifdef PNG_ALIGNED_MEMORY_SUPPORTED
+     /* Use 16-byte aligned memory for row_buf with at least 16 bytes
+      * of padding before and after row_buf; treat prev_row similarly.
+      * NOTE: the alignment is to the start of the pixels, one beyond the start
+      * of the buffer, because of the filter byte.  Prior to libpng 1.5.6 this
+      * was incorrect; the filter byte was aligned, which had the exact
+      * opposite effect of that intended.
+      */
+     {
+        png_bytep temp = png_ptr->big_row_buf + 32;
+        int extra = (int)((temp - (png_bytep)0) & 0x0f);
+        png_ptr->row_buf = temp - extra - 1/*filter byte*/;
+
+        temp = png_ptr->big_prev_row + 32;
+        extra = (int)((temp - (png_bytep)0) & 0x0f);
+        png_ptr->prev_row = temp - extra - 1/*filter byte*/;
+     }
+
+#else
+     /* Use 31 bytes of padding before and 17 bytes after row_buf. */
+     png_ptr->row_buf = png_ptr->big_row_buf + 31;
+     png_ptr->prev_row = png_ptr->big_prev_row + 31;
+#endif
+     png_ptr->old_big_row_buf_size = row_bytes + 48;
+   }
+
+#ifdef PNG_MAX_MALLOC_64K
+   if (png_ptr->rowbytes > 65535)
+      png_error(png_ptr, "This image requires a row greater than 64KB");
+
+#endif
+   if (png_ptr->rowbytes > (PNG_SIZE_MAX - 1))
+      png_error(png_ptr, "Row has too many bytes to allocate in memory");
+
+   memset(png_ptr->prev_row, 0, png_ptr->rowbytes + 1);
+
+   png_debug1(3, "width = %u,", png_ptr->width);
+   png_debug1(3, "height = %u,", png_ptr->height);
+   png_debug1(3, "iwidth = %u,", png_ptr->iwidth);
+   png_debug1(3, "num_rows = %u,", png_ptr->num_rows);
+   png_debug1(3, "rowbytes = %lu,", (unsigned long)png_ptr->rowbytes);
+   png_debug1(3, "irowbytes = %lu",
+       (unsigned long)PNG_ROWBYTES(png_ptr->pixel_depth, png_ptr->iwidth) + 1);
+
+   /* The sequential reader needs a buffer for IDAT, but the progressive reader
+    * does not, so free the read buffer now regardless; the sequential reader
+    * reallocates it on demand.
+    */
+   if (png_ptr->read_buffer)
+   {
+      png_bytep buffer = png_ptr->read_buffer;
+
+      png_ptr->read_buffer_size = 0;
+      png_ptr->read_buffer = NULL;
+      png_free(png_ptr, buffer);
+   }
+
+   /* Finally claim the zstream for the inflate of the IDAT data, use the bits
+    * value from the stream (note that this will result in a fatal error if the
+    * IDAT stream has a bogus deflate header window_bits value, but this should
+    * not be happening any longer!)
+    */
+   if (png_inflate_claim(png_ptr, png_IDAT) != Z_OK)
+      png_error(png_ptr, png_ptr->zstream.msg);
+
+   png_ptr->flags |= PNG_FLAG_ROW_INIT;
+}
+#endif /* PNG_READ_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngset.c b/ml/dlib/dlib/external/libpng/pngset.c
new file mode 100644
index 000000000..7e355d1f4
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngset.c
@@ -0,0 +1,1597 @@
+
+/* pngset.c - storage of image information into info struct
+ *
+ * Last changed in libpng 1.6.3 [July 18, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * The functions here are used during reads to store data from the file
+ * into the info struct, and during writes to store application data
+ * into the info struct for writing into the file.  This abstracts the
+ * info struct and allows us to change the structure in the future.
+ */
+
+#include "pngpriv.h"
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+
+#ifdef PNG_bKGD_SUPPORTED
+void PNGAPI
+png_set_bKGD(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_color_16p background)
+{
+   png_debug1(1, "in %s storage function", "bKGD");
+
+   if (png_ptr == NULL || info_ptr == NULL || background == NULL)
+      return;
+
+   info_ptr->background = *background;
+   info_ptr->valid |= PNG_INFO_bKGD;
+}
+#endif
+
+#ifdef PNG_cHRM_SUPPORTED
+void PNGFAPI
+png_set_cHRM_fixed(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_fixed_point white_x, png_fixed_point white_y, png_fixed_point red_x,
+    png_fixed_point red_y, png_fixed_point green_x, png_fixed_point green_y,
+    png_fixed_point blue_x, png_fixed_point blue_y)
+{
+   png_xy xy;
+
+   png_debug1(1, "in %s storage function", "cHRM fixed");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   xy.redx = red_x;
+   xy.redy = red_y;
+   xy.greenx = green_x;
+   xy.greeny = green_y;
+   xy.bluex = blue_x;
+   xy.bluey = blue_y;
+   xy.whitex = white_x;
+   xy.whitey = white_y;
+
+   if (png_colorspace_set_chromaticities(png_ptr, &info_ptr->colorspace, &xy,
+      2/* override with app values*/))
+      info_ptr->colorspace.flags |= PNG_COLORSPACE_FROM_cHRM;
+
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+
+void PNGFAPI
+png_set_cHRM_XYZ_fixed(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_fixed_point int_red_X, png_fixed_point int_red_Y,
+    png_fixed_point int_red_Z, png_fixed_point int_green_X,
+    png_fixed_point int_green_Y, png_fixed_point int_green_Z,
+    png_fixed_point int_blue_X, png_fixed_point int_blue_Y,
+    png_fixed_point int_blue_Z)
+{
+   png_XYZ XYZ;
+
+   png_debug1(1, "in %s storage function", "cHRM XYZ fixed");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   XYZ.red_X = int_red_X;
+   XYZ.red_Y = int_red_Y;
+   XYZ.red_Z = int_red_Z;
+   XYZ.green_X = int_green_X;
+   XYZ.green_Y = int_green_Y;
+   XYZ.green_Z = int_green_Z;
+   XYZ.blue_X = int_blue_X;
+   XYZ.blue_Y = int_blue_Y;
+   XYZ.blue_Z = int_blue_Z;
+
+   if (png_colorspace_set_endpoints(png_ptr, &info_ptr->colorspace, &XYZ, 2))
+      info_ptr->colorspace.flags |= PNG_COLORSPACE_FROM_cHRM;
+
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_cHRM(png_const_structrp png_ptr, png_inforp info_ptr,
+    double white_x, double white_y, double red_x, double red_y,
+    double green_x, double green_y, double blue_x, double blue_y)
+{
+   png_set_cHRM_fixed(png_ptr, info_ptr,
+      png_fixed(png_ptr, white_x, "cHRM White X"),
+      png_fixed(png_ptr, white_y, "cHRM White Y"),
+      png_fixed(png_ptr, red_x, "cHRM Red X"),
+      png_fixed(png_ptr, red_y, "cHRM Red Y"),
+      png_fixed(png_ptr, green_x, "cHRM Green X"),
+      png_fixed(png_ptr, green_y, "cHRM Green Y"),
+      png_fixed(png_ptr, blue_x, "cHRM Blue X"),
+      png_fixed(png_ptr, blue_y, "cHRM Blue Y"));
+}
+
+void PNGAPI
+png_set_cHRM_XYZ(png_const_structrp png_ptr, png_inforp info_ptr, double red_X,
+    double red_Y, double red_Z, double green_X, double green_Y, double green_Z,
+    double blue_X, double blue_Y, double blue_Z)
+{
+   png_set_cHRM_XYZ_fixed(png_ptr, info_ptr,
+      png_fixed(png_ptr, red_X, "cHRM Red X"),
+      png_fixed(png_ptr, red_Y, "cHRM Red Y"),
+      png_fixed(png_ptr, red_Z, "cHRM Red Z"),
+      png_fixed(png_ptr, green_X, "cHRM Red X"),
+      png_fixed(png_ptr, green_Y, "cHRM Red Y"),
+      png_fixed(png_ptr, green_Z, "cHRM Red Z"),
+      png_fixed(png_ptr, blue_X, "cHRM Red X"),
+      png_fixed(png_ptr, blue_Y, "cHRM Red Y"),
+      png_fixed(png_ptr, blue_Z, "cHRM Red Z"));
+}
+#  endif /* PNG_FLOATING_POINT_SUPPORTED */
+
+#endif /* PNG_cHRM_SUPPORTED */
+
+#ifdef PNG_gAMA_SUPPORTED
+void PNGFAPI
+png_set_gAMA_fixed(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_fixed_point file_gamma)
+{
+   png_debug1(1, "in %s storage function", "gAMA");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   png_colorspace_set_gamma(png_ptr, &info_ptr->colorspace, file_gamma);
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_gAMA(png_const_structrp png_ptr, png_inforp info_ptr, double file_gamma)
+{
+   png_set_gAMA_fixed(png_ptr, info_ptr, png_fixed(png_ptr, file_gamma,
+       "png_set_gAMA"));
+}
+#  endif
+#endif
+
+#ifdef PNG_hIST_SUPPORTED
+void PNGAPI
+png_set_hIST(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_uint_16p hist)
+{
+   int i;
+
+   png_debug1(1, "in %s storage function", "hIST");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (info_ptr->num_palette == 0 || info_ptr->num_palette
+       > PNG_MAX_PALETTE_LENGTH)
+   {
+      png_warning(png_ptr,
+          "Invalid palette size, hIST allocation skipped");
+
+      return;
+   }
+
+   png_free_data(png_ptr, info_ptr, PNG_FREE_HIST, 0);
+
+   /* Changed from info->num_palette to PNG_MAX_PALETTE_LENGTH in
+    * version 1.2.1
+    */
+   info_ptr->hist = png_voidcast(png_uint_16p, png_malloc_warn(png_ptr,
+       PNG_MAX_PALETTE_LENGTH * (sizeof (png_uint_16))));
+
+   if (info_ptr->hist == NULL)
+   {
+      png_warning(png_ptr, "Insufficient memory for hIST chunk data");
+      return;
+   }
+
+   info_ptr->free_me |= PNG_FREE_HIST;
+
+   for (i = 0; i < info_ptr->num_palette; i++)
+      info_ptr->hist[i] = hist[i];
+
+   info_ptr->valid |= PNG_INFO_hIST;
+}
+#endif
+
+void PNGAPI
+png_set_IHDR(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_uint_32 width, png_uint_32 height, int bit_depth,
+    int color_type, int interlace_type, int compression_type,
+    int filter_type)
+{
+   png_debug1(1, "in %s storage function", "IHDR");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   info_ptr->width = width;
+   info_ptr->height = height;
+   info_ptr->bit_depth = (png_byte)bit_depth;
+   info_ptr->color_type = (png_byte)color_type;
+   info_ptr->compression_type = (png_byte)compression_type;
+   info_ptr->filter_type = (png_byte)filter_type;
+   info_ptr->interlace_type = (png_byte)interlace_type;
+
+   png_check_IHDR (png_ptr, info_ptr->width, info_ptr->height,
+       info_ptr->bit_depth, info_ptr->color_type, info_ptr->interlace_type,
+       info_ptr->compression_type, info_ptr->filter_type);
+
+   if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      info_ptr->channels = 1;
+
+   else if (info_ptr->color_type & PNG_COLOR_MASK_COLOR)
+      info_ptr->channels = 3;
+
+   else
+      info_ptr->channels = 1;
+
+   if (info_ptr->color_type & PNG_COLOR_MASK_ALPHA)
+      info_ptr->channels++;
+
+   info_ptr->pixel_depth = (png_byte)(info_ptr->channels * info_ptr->bit_depth);
+
+   info_ptr->rowbytes = PNG_ROWBYTES(info_ptr->pixel_depth, width);
+}
+
+#ifdef PNG_oFFs_SUPPORTED
+void PNGAPI
+png_set_oFFs(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_int_32 offset_x, png_int_32 offset_y, int unit_type)
+{
+   png_debug1(1, "in %s storage function", "oFFs");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   info_ptr->x_offset = offset_x;
+   info_ptr->y_offset = offset_y;
+   info_ptr->offset_unit_type = (png_byte)unit_type;
+   info_ptr->valid |= PNG_INFO_oFFs;
+}
+#endif
+
+#ifdef PNG_pCAL_SUPPORTED
+void PNGAPI
+png_set_pCAL(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_charp purpose, png_int_32 X0, png_int_32 X1, int type,
+    int nparams, png_const_charp units, png_charpp params)
+{
+   png_size_t length;
+   int i;
+
+   png_debug1(1, "in %s storage function", "pCAL");
+
+   if (png_ptr == NULL || info_ptr == NULL || purpose == NULL || units == NULL
+      || (nparams > 0 && params == NULL))
+      return;
+
+   length = strlen(purpose) + 1;
+   png_debug1(3, "allocating purpose for info (%lu bytes)",
+       (unsigned long)length);
+
+   /* TODO: validate format of calibration name and unit name */
+
+   /* Check that the type matches the specification. */
+   if (type < 0 || type > 3)
+      png_error(png_ptr, "Invalid pCAL equation type");
+
+   if (nparams < 0 || nparams > 255)
+      png_error(png_ptr, "Invalid pCAL parameter count");
+
+   /* Validate params[nparams] */
+   for (i=0; i<nparams; ++i)
+      if (params[i] == NULL ||
+         !png_check_fp_string(params[i], strlen(params[i])))
+         png_error(png_ptr, "Invalid format for pCAL parameter");
+
+   info_ptr->pcal_purpose = png_voidcast(png_charp,
+      png_malloc_warn(png_ptr, length));
+
+   if (info_ptr->pcal_purpose == NULL)
+   {
+      png_warning(png_ptr, "Insufficient memory for pCAL purpose");
+      return;
+   }
+
+   memcpy(info_ptr->pcal_purpose, purpose, length);
+
+   png_debug(3, "storing X0, X1, type, and nparams in info");
+   info_ptr->pcal_X0 = X0;
+   info_ptr->pcal_X1 = X1;
+   info_ptr->pcal_type = (png_byte)type;
+   info_ptr->pcal_nparams = (png_byte)nparams;
+
+   length = strlen(units) + 1;
+   png_debug1(3, "allocating units for info (%lu bytes)",
+     (unsigned long)length);
+
+   info_ptr->pcal_units = png_voidcast(png_charp,
+      png_malloc_warn(png_ptr, length));
+
+   if (info_ptr->pcal_units == NULL)
+   {
+      png_warning(png_ptr, "Insufficient memory for pCAL units");
+      return;
+   }
+
+   memcpy(info_ptr->pcal_units, units, length);
+
+   info_ptr->pcal_params = png_voidcast(png_charpp, png_malloc_warn(png_ptr,
+       (png_size_t)((nparams + 1) * (sizeof (png_charp)))));
+
+   if (info_ptr->pcal_params == NULL)
+   {
+      png_warning(png_ptr, "Insufficient memory for pCAL params");
+      return;
+   }
+
+   memset(info_ptr->pcal_params, 0, (nparams + 1) * (sizeof (png_charp)));
+
+   for (i = 0; i < nparams; i++)
+   {
+      length = strlen(params[i]) + 1;
+      png_debug2(3, "allocating parameter %d for info (%lu bytes)", i,
+          (unsigned long)length);
+
+      info_ptr->pcal_params[i] = (png_charp)png_malloc_warn(png_ptr, length);
+
+      if (info_ptr->pcal_params[i] == NULL)
+      {
+         png_warning(png_ptr, "Insufficient memory for pCAL parameter");
+         return;
+      }
+
+      memcpy(info_ptr->pcal_params[i], params[i], length);
+   }
+
+   info_ptr->valid |= PNG_INFO_pCAL;
+   info_ptr->free_me |= PNG_FREE_PCAL;
+}
+#endif
+
+#ifdef PNG_sCAL_SUPPORTED
+void PNGAPI
+png_set_sCAL_s(png_const_structrp png_ptr, png_inforp info_ptr,
+    int unit, png_const_charp swidth, png_const_charp sheight)
+{
+   png_size_t lengthw = 0, lengthh = 0;
+
+   png_debug1(1, "in %s storage function", "sCAL");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   /* Double check the unit (should never get here with an invalid
+    * unit unless this is an API call.)
+    */
+   if (unit != 1 && unit != 2)
+      png_error(png_ptr, "Invalid sCAL unit");
+
+   if (swidth == NULL || (lengthw = strlen(swidth)) == 0 ||
+       swidth[0] == 45 /* '-' */ || !png_check_fp_string(swidth, lengthw))
+      png_error(png_ptr, "Invalid sCAL width");
+
+   if (sheight == NULL || (lengthh = strlen(sheight)) == 0 ||
+       sheight[0] == 45 /* '-' */ || !png_check_fp_string(sheight, lengthh))
+      png_error(png_ptr, "Invalid sCAL height");
+
+   info_ptr->scal_unit = (png_byte)unit;
+
+   ++lengthw;
+
+   png_debug1(3, "allocating unit for info (%u bytes)", (unsigned int)lengthw);
+
+   info_ptr->scal_s_width = png_voidcast(png_charp,
+      png_malloc_warn(png_ptr, lengthw));
+
+   if (info_ptr->scal_s_width == NULL)
+   {
+      png_warning(png_ptr, "Memory allocation failed while processing sCAL");
+      return;
+   }
+
+   memcpy(info_ptr->scal_s_width, swidth, lengthw);
+
+   ++lengthh;
+
+   png_debug1(3, "allocating unit for info (%u bytes)", (unsigned int)lengthh);
+
+   info_ptr->scal_s_height = png_voidcast(png_charp,
+      png_malloc_warn(png_ptr, lengthh));
+
+   if (info_ptr->scal_s_height == NULL)
+   {
+      png_free (png_ptr, info_ptr->scal_s_width);
+      info_ptr->scal_s_width = NULL;
+
+      png_warning(png_ptr, "Memory allocation failed while processing sCAL");
+      return;
+   }
+
+   memcpy(info_ptr->scal_s_height, sheight, lengthh);
+
+   info_ptr->valid |= PNG_INFO_sCAL;
+   info_ptr->free_me |= PNG_FREE_SCAL;
+}
+
+#  ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_sCAL(png_const_structrp png_ptr, png_inforp info_ptr, int unit,
+    double width, double height)
+{
+   png_debug1(1, "in %s storage function", "sCAL");
+
+   /* Check the arguments. */
+   if (width <= 0)
+      png_warning(png_ptr, "Invalid sCAL width ignored");
+
+   else if (height <= 0)
+      png_warning(png_ptr, "Invalid sCAL height ignored");
+
+   else
+   {
+      /* Convert 'width' and 'height' to ASCII. */
+      char swidth[PNG_sCAL_MAX_DIGITS+1];
+      char sheight[PNG_sCAL_MAX_DIGITS+1];
+
+      png_ascii_from_fp(png_ptr, swidth, (sizeof swidth), width,
+         PNG_sCAL_PRECISION);
+      png_ascii_from_fp(png_ptr, sheight, (sizeof sheight), height,
+         PNG_sCAL_PRECISION);
+
+      png_set_sCAL_s(png_ptr, info_ptr, unit, swidth, sheight);
+   }
+}
+#  endif
+
+#  ifdef PNG_FIXED_POINT_SUPPORTED
+void PNGAPI
+png_set_sCAL_fixed(png_const_structrp png_ptr, png_inforp info_ptr, int unit,
+    png_fixed_point width, png_fixed_point height)
+{
+   png_debug1(1, "in %s storage function", "sCAL");
+
+   /* Check the arguments. */
+   if (width <= 0)
+      png_warning(png_ptr, "Invalid sCAL width ignored");
+
+   else if (height <= 0)
+      png_warning(png_ptr, "Invalid sCAL height ignored");
+
+   else
+   {
+      /* Convert 'width' and 'height' to ASCII. */
+      char swidth[PNG_sCAL_MAX_DIGITS+1];
+      char sheight[PNG_sCAL_MAX_DIGITS+1];
+
+      png_ascii_from_fixed(png_ptr, swidth, (sizeof swidth), width);
+      png_ascii_from_fixed(png_ptr, sheight, (sizeof sheight), height);
+
+      png_set_sCAL_s(png_ptr, info_ptr, unit, swidth, sheight);
+   }
+}
+#  endif
+#endif
+
+#ifdef PNG_pHYs_SUPPORTED
+void PNGAPI
+png_set_pHYs(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_uint_32 res_x, png_uint_32 res_y, int unit_type)
+{
+   png_debug1(1, "in %s storage function", "pHYs");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   info_ptr->x_pixels_per_unit = res_x;
+   info_ptr->y_pixels_per_unit = res_y;
+   info_ptr->phys_unit_type = (png_byte)unit_type;
+   info_ptr->valid |= PNG_INFO_pHYs;
+}
+#endif
+
+void PNGAPI
+png_set_PLTE(png_structrp png_ptr, png_inforp info_ptr,
+    png_const_colorp palette, int num_palette)
+{
+
+   png_debug1(1, "in %s storage function", "PLTE");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (num_palette < 0 || num_palette > PNG_MAX_PALETTE_LENGTH)
+   {
+      if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+         png_error(png_ptr, "Invalid palette length");
+
+      else
+      {
+         png_warning(png_ptr, "Invalid palette length");
+         return;
+      }
+   }
+
+   if ((num_palette > 0 && palette == NULL) ||
+      (num_palette == 0
+#        ifdef PNG_MNG_FEATURES_SUPPORTED
+            && (png_ptr->mng_features_permitted & PNG_FLAG_MNG_EMPTY_PLTE) == 0
+#        endif
+      ))
+   {
+      png_chunk_report(png_ptr, "Invalid palette", PNG_CHUNK_ERROR);
+      return;
+   }
+
+   /* It may not actually be necessary to set png_ptr->palette here;
+    * we do it for backward compatibility with the way the png_handle_tRNS
+    * function used to do the allocation.
+    *
+    * 1.6.0: the above statement appears to be incorrect; something has to set
+    * the palette inside png_struct on read.
+    */
+   png_free_data(png_ptr, info_ptr, PNG_FREE_PLTE, 0);
+
+   /* Changed in libpng-1.2.1 to allocate PNG_MAX_PALETTE_LENGTH instead
+    * of num_palette entries, in case of an invalid PNG file that has
+    * too-large sample values.
+    */
+   png_ptr->palette = png_voidcast(png_colorp, png_calloc(png_ptr,
+       PNG_MAX_PALETTE_LENGTH * (sizeof (png_color))));
+
+   if (num_palette > 0)
+      memcpy(png_ptr->palette, palette, num_palette * (sizeof (png_color)));
+   info_ptr->palette = png_ptr->palette;
+   info_ptr->num_palette = png_ptr->num_palette = (png_uint_16)num_palette;
+
+   info_ptr->free_me |= PNG_FREE_PLTE;
+
+   info_ptr->valid |= PNG_INFO_PLTE;
+}
+
+#ifdef PNG_sBIT_SUPPORTED
+void PNGAPI
+png_set_sBIT(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_color_8p sig_bit)
+{
+   png_debug1(1, "in %s storage function", "sBIT");
+
+   if (png_ptr == NULL || info_ptr == NULL || sig_bit == NULL)
+      return;
+
+   info_ptr->sig_bit = *sig_bit;
+   info_ptr->valid |= PNG_INFO_sBIT;
+}
+#endif
+
+#ifdef PNG_sRGB_SUPPORTED
+void PNGAPI
+png_set_sRGB(png_const_structrp png_ptr, png_inforp info_ptr, int srgb_intent)
+{
+   png_debug1(1, "in %s storage function", "sRGB");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   (void)png_colorspace_set_sRGB(png_ptr, &info_ptr->colorspace, srgb_intent);
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+
+void PNGAPI
+png_set_sRGB_gAMA_and_cHRM(png_const_structrp png_ptr, png_inforp info_ptr,
+    int srgb_intent)
+{
+   png_debug1(1, "in %s storage function", "sRGB_gAMA_and_cHRM");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (png_colorspace_set_sRGB(png_ptr, &info_ptr->colorspace, srgb_intent))
+   {
+      /* This causes the gAMA and cHRM to be written too */
+      info_ptr->colorspace.flags |=
+         PNG_COLORSPACE_FROM_gAMA|PNG_COLORSPACE_FROM_cHRM;
+   }
+
+   png_colorspace_sync_info(png_ptr, info_ptr);
+}
+#endif /* sRGB */
+
+
+#ifdef PNG_iCCP_SUPPORTED
+void PNGAPI
+png_set_iCCP(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_charp name, int compression_type,
+    png_const_bytep profile, png_uint_32 proflen)
+{
+   png_charp new_iccp_name;
+   png_bytep new_iccp_profile;
+   png_size_t length;
+
+   png_debug1(1, "in %s storage function", "iCCP");
+
+   if (png_ptr == NULL || info_ptr == NULL || name == NULL || profile == NULL)
+      return;
+
+   if (compression_type != PNG_COMPRESSION_TYPE_BASE)
+      png_app_error(png_ptr, "Invalid iCCP compression method");
+
+   /* Set the colorspace first because this validates the profile; do not
+    * override previously set app cHRM or gAMA here (because likely as not the
+    * application knows better than libpng what the correct values are.)  Pass
+    * the info_ptr color_type field to png_colorspace_set_ICC because in the
+    * write case it has not yet been stored in png_ptr.
+    */
+   {
+      int result = png_colorspace_set_ICC(png_ptr, &info_ptr->colorspace, name,
+         proflen, profile, info_ptr->color_type);
+
+      png_colorspace_sync_info(png_ptr, info_ptr);
+
+      /* Don't do any of the copying if the profile was bad, or inconsistent. */
+      if (!result)
+         return;
+
+      /* But do write the gAMA and cHRM chunks from the profile. */
+      info_ptr->colorspace.flags |=
+         PNG_COLORSPACE_FROM_gAMA|PNG_COLORSPACE_FROM_cHRM;
+   }
+
+   length = strlen(name)+1;
+   new_iccp_name = png_voidcast(png_charp, png_malloc_warn(png_ptr, length));
+
+   if (new_iccp_name == NULL)
+   {
+      png_benign_error(png_ptr, "Insufficient memory to process iCCP chunk");
+      return;
+   }
+
+   memcpy(new_iccp_name, name, length);
+   new_iccp_profile = png_voidcast(png_bytep,
+      png_malloc_warn(png_ptr, proflen));
+
+   if (new_iccp_profile == NULL)
+   {
+      png_free(png_ptr, new_iccp_name);
+      png_benign_error(png_ptr,
+          "Insufficient memory to process iCCP profile");
+      return;
+   }
+
+   memcpy(new_iccp_profile, profile, proflen);
+
+   png_free_data(png_ptr, info_ptr, PNG_FREE_ICCP, 0);
+
+   info_ptr->iccp_proflen = proflen;
+   info_ptr->iccp_name = new_iccp_name;
+   info_ptr->iccp_profile = new_iccp_profile;
+   info_ptr->free_me |= PNG_FREE_ICCP;
+   info_ptr->valid |= PNG_INFO_iCCP;
+}
+#endif
+
+#ifdef PNG_TEXT_SUPPORTED
+void PNGAPI
+png_set_text(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_textp text_ptr, int num_text)
+{
+   int ret;
+   ret = png_set_text_2(png_ptr, info_ptr, text_ptr, num_text);
+
+   if (ret)
+      png_error(png_ptr, "Insufficient memory to store text");
+}
+
+int /* PRIVATE */
+png_set_text_2(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_textp text_ptr, int num_text)
+{
+   int i;
+
+   png_debug1(1, "in %lx storage function", png_ptr == NULL ? "unexpected" :
+      (unsigned long)png_ptr->chunk_name);
+
+   if (png_ptr == NULL || info_ptr == NULL || num_text <= 0 || text_ptr == NULL)
+      return(0);
+
+   /* Make sure we have enough space in the "text" array in info_struct
+    * to hold all of the incoming text_ptr objects.  This compare can't overflow
+    * because max_text >= num_text (anyway, subtract of two positive integers
+    * can't overflow in any case.)
+    */
+   if (num_text > info_ptr->max_text - info_ptr->num_text)
+   {
+      int old_num_text = info_ptr->num_text;
+      int max_text;
+      png_textp new_text = NULL;
+
+      /* Calculate an appropriate max_text, checking for overflow. */
+      max_text = old_num_text;
+      if (num_text <= INT_MAX - max_text)
+      {
+         max_text += num_text;
+
+         /* Round up to a multiple of 8 */
+         if (max_text < INT_MAX-8)
+            max_text = (max_text + 8) & ~0x7;
+
+         else
+            max_text = INT_MAX;
+
+         /* Now allocate a new array and copy the old members in, this does all
+          * the overflow checks.
+          */
+         new_text = png_voidcast(png_textp,png_realloc_array(png_ptr,
+            info_ptr->text, old_num_text, max_text-old_num_text,
+            sizeof *new_text));
+      }
+
+      if (new_text == NULL)
+      {
+         png_chunk_report(png_ptr, "too many text chunks",
+            PNG_CHUNK_WRITE_ERROR);
+         return 1;
+      }
+
+      png_free(png_ptr, info_ptr->text);
+
+      info_ptr->text = new_text;
+      info_ptr->free_me |= PNG_FREE_TEXT;
+      info_ptr->max_text = max_text;
+      /* num_text is adjusted below as the entries are copied in */
+
+      png_debug1(3, "allocated %d entries for info_ptr->text", max_text);
+   }
+
+   for (i = 0; i < num_text; i++)
+   {
+      size_t text_length, key_len;
+      size_t lang_len, lang_key_len;
+      png_textp textp = &(info_ptr->text[info_ptr->num_text]);
+
+      if (text_ptr[i].key == NULL)
+          continue;
+
+      if (text_ptr[i].compression < PNG_TEXT_COMPRESSION_NONE ||
+          text_ptr[i].compression >= PNG_TEXT_COMPRESSION_LAST)
+      {
+         png_chunk_report(png_ptr, "text compression mode is out of range",
+            PNG_CHUNK_WRITE_ERROR);
+         continue;
+      }
+
+      key_len = strlen(text_ptr[i].key);
+
+      if (text_ptr[i].compression <= 0)
+      {
+         lang_len = 0;
+         lang_key_len = 0;
+      }
+
+      else
+#  ifdef PNG_iTXt_SUPPORTED
+      {
+         /* Set iTXt data */
+
+         if (text_ptr[i].lang != NULL)
+            lang_len = strlen(text_ptr[i].lang);
+
+         else
+            lang_len = 0;
+
+         if (text_ptr[i].lang_key != NULL)
+            lang_key_len = strlen(text_ptr[i].lang_key);
+
+         else
+            lang_key_len = 0;
+      }
+#  else /* PNG_iTXt_SUPPORTED */
+      {
+         png_chunk_report(png_ptr, "iTXt chunk not supported",
+            PNG_CHUNK_WRITE_ERROR);
+         continue;
+      }
+#  endif
+
+      if (text_ptr[i].text == NULL || text_ptr[i].text[0] == '\0')
+      {
+         text_length = 0;
+#  ifdef PNG_iTXt_SUPPORTED
+         if (text_ptr[i].compression > 0)
+            textp->compression = PNG_ITXT_COMPRESSION_NONE;
+
+         else
+#  endif
+            textp->compression = PNG_TEXT_COMPRESSION_NONE;
+      }
+
+      else
+      {
+         text_length = strlen(text_ptr[i].text);
+         textp->compression = text_ptr[i].compression;
+      }
+
+      textp->key = png_voidcast(png_charp,png_malloc_base(png_ptr,
+          key_len + text_length + lang_len + lang_key_len + 4));
+
+      if (textp->key == NULL)
+      {
+         png_chunk_report(png_ptr, "text chunk: out of memory",
+               PNG_CHUNK_WRITE_ERROR);
+         return 1;
+      }
+
+      png_debug2(2, "Allocated %lu bytes at %p in png_set_text",
+          (unsigned long)(png_uint_32)
+          (key_len + lang_len + lang_key_len + text_length + 4),
+          textp->key);
+
+      memcpy(textp->key, text_ptr[i].key, key_len);
+      *(textp->key + key_len) = '\0';
+
+      if (text_ptr[i].compression > 0)
+      {
+         textp->lang = textp->key + key_len + 1;
+         memcpy(textp->lang, text_ptr[i].lang, lang_len);
+         *(textp->lang + lang_len) = '\0';
+         textp->lang_key = textp->lang + lang_len + 1;
+         memcpy(textp->lang_key, text_ptr[i].lang_key, lang_key_len);
+         *(textp->lang_key + lang_key_len) = '\0';
+         textp->text = textp->lang_key + lang_key_len + 1;
+      }
+
+      else
+      {
+         textp->lang=NULL;
+         textp->lang_key=NULL;
+         textp->text = textp->key + key_len + 1;
+      }
+
+      if (text_length)
+         memcpy(textp->text, text_ptr[i].text, text_length);
+
+      *(textp->text + text_length) = '\0';
+
+#  ifdef PNG_iTXt_SUPPORTED
+      if (textp->compression > 0)
+      {
+         textp->text_length = 0;
+         textp->itxt_length = text_length;
+      }
+
+      else
+#  endif
+      {
+         textp->text_length = text_length;
+         textp->itxt_length = 0;
+      }
+
+      info_ptr->num_text++;
+      png_debug1(3, "transferred text chunk %d", info_ptr->num_text);
+   }
+
+   return(0);
+}
+#endif
+
+#ifdef PNG_tIME_SUPPORTED
+void PNGAPI
+png_set_tIME(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_const_timep mod_time)
+{
+   png_debug1(1, "in %s storage function", "tIME");
+
+   if (png_ptr == NULL || info_ptr == NULL || mod_time == NULL ||
+       (png_ptr->mode & PNG_WROTE_tIME))
+      return;
+
+   if (mod_time->month == 0   || mod_time->month > 12  ||
+       mod_time->day   == 0   || mod_time->day   > 31  ||
+       mod_time->hour  > 23   || mod_time->minute > 59 ||
+       mod_time->second > 60)
+   {
+      png_warning(png_ptr, "Ignoring invalid time value");
+      return;
+   }
+
+   info_ptr->mod_time = *mod_time;
+   info_ptr->valid |= PNG_INFO_tIME;
+}
+#endif
+
+#ifdef PNG_tRNS_SUPPORTED
+void PNGAPI
+png_set_tRNS(png_structrp png_ptr, png_inforp info_ptr,
+    png_const_bytep trans_alpha, int num_trans, png_const_color_16p trans_color)
+{
+   png_debug1(1, "in %s storage function", "tRNS");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (trans_alpha != NULL)
+   {
+       /* It may not actually be necessary to set png_ptr->trans_alpha here;
+        * we do it for backward compatibility with the way the png_handle_tRNS
+        * function used to do the allocation.
+        *
+        * 1.6.0: The above statement is incorrect; png_handle_tRNS effectively
+        * relies on png_set_tRNS storing the information in png_struct
+        * (otherwise it won't be there for the code in pngrtran.c).
+        */
+
+       png_free_data(png_ptr, info_ptr, PNG_FREE_TRNS, 0);
+
+       /* Changed from num_trans to PNG_MAX_PALETTE_LENGTH in version 1.2.1 */
+       png_ptr->trans_alpha = info_ptr->trans_alpha = png_voidcast(png_bytep,
+         png_malloc(png_ptr, PNG_MAX_PALETTE_LENGTH));
+
+       if (num_trans > 0 && num_trans <= PNG_MAX_PALETTE_LENGTH)
+          memcpy(info_ptr->trans_alpha, trans_alpha, (png_size_t)num_trans);
+   }
+
+   if (trans_color != NULL)
+   {
+      int sample_max = (1 << info_ptr->bit_depth);
+
+      if ((info_ptr->color_type == PNG_COLOR_TYPE_GRAY &&
+          trans_color->gray > sample_max) ||
+          (info_ptr->color_type == PNG_COLOR_TYPE_RGB &&
+          (trans_color->red > sample_max ||
+          trans_color->green > sample_max ||
+          trans_color->blue > sample_max)))
+         png_warning(png_ptr,
+            "tRNS chunk has out-of-range samples for bit_depth");
+
+      info_ptr->trans_color = *trans_color;
+
+      if (num_trans == 0)
+         num_trans = 1;
+   }
+
+   info_ptr->num_trans = (png_uint_16)num_trans;
+
+   if (num_trans != 0)
+   {
+      info_ptr->valid |= PNG_INFO_tRNS;
+      info_ptr->free_me |= PNG_FREE_TRNS;
+   }
+}
+#endif
+
+#ifdef PNG_sPLT_SUPPORTED
+void PNGAPI
+png_set_sPLT(png_const_structrp png_ptr,
+    png_inforp info_ptr, png_const_sPLT_tp entries, int nentries)
+/*
+ *  entries        - array of png_sPLT_t structures
+ *                   to be added to the list of palettes
+ *                   in the info structure.
+ *
+ *  nentries       - number of palette structures to be
+ *                   added.
+ */
+{
+   png_sPLT_tp np;
+
+   if (png_ptr == NULL || info_ptr == NULL || nentries <= 0 || entries == NULL)
+      return;
+
+   /* Use the internal realloc function, which checks for all the possible
+    * overflows.  Notice that the parameters are (int) and (size_t)
+    */
+   np = png_voidcast(png_sPLT_tp,png_realloc_array(png_ptr,
+      info_ptr->splt_palettes, info_ptr->splt_palettes_num, nentries,
+      sizeof *np));
+
+   if (np == NULL)
+   {
+      /* Out of memory or too many chunks */
+      png_chunk_report(png_ptr, "too many sPLT chunks", PNG_CHUNK_WRITE_ERROR);
+      return;
+   }
+
+   png_free(png_ptr, info_ptr->splt_palettes);
+   info_ptr->splt_palettes = np;
+   info_ptr->free_me |= PNG_FREE_SPLT;
+
+   np += info_ptr->splt_palettes_num;
+
+   do
+   {
+      png_size_t length;
+
+      /* Skip invalid input entries */
+      if (entries->name == NULL || entries->entries == NULL)
+      {
+         /* png_handle_sPLT doesn't do this, so this is an app error */
+         png_app_error(png_ptr, "png_set_sPLT: invalid sPLT");
+         /* Just skip the invalid entry */
+         continue;
+      }
+
+      np->depth = entries->depth;
+
+      /* In the even of out-of-memory just return - there's no point keeping on
+       * trying to add sPLT chunks.
+       */
+      length = strlen(entries->name) + 1;
+      np->name = png_voidcast(png_charp, png_malloc_base(png_ptr, length));
+
+      if (np->name == NULL)
+         break;
+
+      memcpy(np->name, entries->name, length);
+
+      /* IMPORTANT: we have memory now that won't get freed if something else
+       * goes wrong, this code must free it.  png_malloc_array produces no
+       * warnings, use a png_chunk_report (below) if there is an error.
+       */
+      np->entries = png_voidcast(png_sPLT_entryp, png_malloc_array(png_ptr,
+          entries->nentries, sizeof (png_sPLT_entry)));
+
+      if (np->entries == NULL)
+      {
+         png_free(png_ptr, np->name);
+         break;
+      }
+
+      np->nentries = entries->nentries;
+      /* This multiply can't overflow because png_malloc_array has already
+       * checked it when doing the allocation.
+       */
+      memcpy(np->entries, entries->entries,
+         entries->nentries * sizeof (png_sPLT_entry));
+
+      /* Note that 'continue' skips the advance of the out pointer and out
+       * count, so an invalid entry is not added.
+       */
+      info_ptr->valid |= PNG_INFO_sPLT;
+      ++(info_ptr->splt_palettes_num);
+      ++np;
+   }
+   while (++entries, --nentries);
+
+   if (nentries > 0)
+      png_chunk_report(png_ptr, "sPLT out of memory", PNG_CHUNK_WRITE_ERROR);
+}
+#endif /* PNG_sPLT_SUPPORTED */
+
+#ifdef PNG_STORE_UNKNOWN_CHUNKS_SUPPORTED
+static png_byte
+check_location(png_const_structrp png_ptr, int location)
+{
+   location &= (PNG_HAVE_IHDR|PNG_HAVE_PLTE|PNG_AFTER_IDAT);
+
+   /* New in 1.6.0; copy the location and check it.  This is an API
+    * change, previously the app had to use the
+    * png_set_unknown_chunk_location API below for each chunk.
+    */
+   if (location == 0 && !(png_ptr->mode & PNG_IS_READ_STRUCT))
+   {
+      /* Write struct, so unknown chunks come from the app */
+      png_app_warning(png_ptr,
+         "png_set_unknown_chunks now expects a valid location");
+      /* Use the old behavior */
+      location = (png_byte)(png_ptr->mode &
+         (PNG_HAVE_IHDR|PNG_HAVE_PLTE|PNG_AFTER_IDAT));
+   }
+
+   /* This need not be an internal error - if the app calls
+    * png_set_unknown_chunks on a read pointer it must get the location right.
+    */
+   if (location == 0)
+      png_error(png_ptr, "invalid location in png_set_unknown_chunks");
+
+   /* Now reduce the location to the top-most set bit by removing each least
+    * significant bit in turn.
+    */
+   while (location != (location & -location))
+      location &= ~(location & -location);
+
+   /* The cast is safe because 'location' is a bit mask and only the low four
+    * bits are significant.
+    */
+   return (png_byte)location;
+}
+
+void PNGAPI
+png_set_unknown_chunks(png_const_structrp png_ptr,
+   png_inforp info_ptr, png_const_unknown_chunkp unknowns, int num_unknowns)
+{
+   png_unknown_chunkp np;
+
+   if (png_ptr == NULL || info_ptr == NULL || num_unknowns <= 0 ||
+      unknowns == NULL)
+      return;
+
+   /* Check for the failure cases where support has been disabled at compile
+    * time.  This code is hardly ever compiled - it's here because
+    * STORE_UNKNOWN_CHUNKS is set by both read and write code (compiling in this
+    * code) but may be meaningless if the read or write handling of unknown
+    * chunks is not compiled in.
+    */
+#  if !defined(PNG_READ_UNKNOWN_CHUNKS_SUPPORTED) && \
+      defined(PNG_READ_SUPPORTED)
+      if (png_ptr->mode & PNG_IS_READ_STRUCT)
+      {
+         png_app_error(png_ptr, "no unknown chunk support on read");
+         return;
+      }
+#  endif
+#  if !defined(PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED) && \
+      defined(PNG_WRITE_SUPPORTED)
+      if (!(png_ptr->mode & PNG_IS_READ_STRUCT))
+      {
+         png_app_error(png_ptr, "no unknown chunk support on write");
+         return;
+      }
+#  endif
+
+   /* Prior to 1.6.0 this code used png_malloc_warn; however, this meant that
+    * unknown critical chunks could be lost with just a warning resulting in
+    * undefined behavior.  Now png_chunk_report is used to provide behavior
+    * appropriate to read or write.
+    */
+   np = png_voidcast(png_unknown_chunkp, png_realloc_array(png_ptr,
+         info_ptr->unknown_chunks, info_ptr->unknown_chunks_num, num_unknowns,
+         sizeof *np));
+
+   if (np == NULL)
+   {
+      png_chunk_report(png_ptr, "too many unknown chunks",
+         PNG_CHUNK_WRITE_ERROR);
+      return;
+   }
+
+   png_free(png_ptr, info_ptr->unknown_chunks);
+   info_ptr->unknown_chunks = np; /* safe because it is initialized */
+   info_ptr->free_me |= PNG_FREE_UNKN;
+
+   np += info_ptr->unknown_chunks_num;
+
+   /* Increment unknown_chunks_num each time round the loop to protect the
+    * just-allocated chunk data.
+    */
+   for (; num_unknowns > 0; --num_unknowns, ++unknowns)
+   {
+      memcpy(np->name, unknowns->name, (sizeof np->name));
+      np->name[(sizeof np->name)-1] = '\0';
+      np->location = check_location(png_ptr, unknowns->location);
+
+      if (unknowns->size == 0)
+      {
+         np->data = NULL;
+         np->size = 0;
+      }
+
+      else
+      {
+         np->data = png_voidcast(png_bytep,
+            png_malloc_base(png_ptr, unknowns->size));
+
+         if (np->data == NULL)
+         {
+            png_chunk_report(png_ptr, "unknown chunk: out of memory",
+               PNG_CHUNK_WRITE_ERROR);
+            /* But just skip storing the unknown chunk */
+            continue;
+         }
+
+         memcpy(np->data, unknowns->data, unknowns->size);
+         np->size = unknowns->size;
+      }
+
+      /* These increments are skipped on out-of-memory for the data - the
+       * unknown chunk entry gets overwritten if the png_chunk_report returns.
+       * This is correct in the read case (the chunk is just dropped.)
+       */
+      ++np;
+      ++(info_ptr->unknown_chunks_num);
+   }
+}
+
+void PNGAPI
+png_set_unknown_chunk_location(png_const_structrp png_ptr, png_inforp info_ptr,
+    int chunk, int location)
+{
+   /* This API is pretty pointless in 1.6.0 because the location can be set
+    * before the call to png_set_unknown_chunks.
+    *
+    * TODO: add a png_app_warning in 1.7
+    */
+   if (png_ptr != NULL && info_ptr != NULL && chunk >= 0 &&
+      chunk < info_ptr->unknown_chunks_num)
+   {
+      if ((location & (PNG_HAVE_IHDR|PNG_HAVE_PLTE|PNG_AFTER_IDAT)) == 0)
+      {
+         png_app_error(png_ptr, "invalid unknown chunk location");
+         /* Fake out the pre 1.6.0 behavior: */
+         if ((location & PNG_HAVE_IDAT)) /* undocumented! */
+            location = PNG_AFTER_IDAT;
+
+         else
+            location = PNG_HAVE_IHDR; /* also undocumented */
+      }
+
+      info_ptr->unknown_chunks[chunk].location =
+         check_location(png_ptr, location);
+   }
+}
+#endif
+
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+png_uint_32 PNGAPI
+png_permit_mng_features (png_structrp png_ptr, png_uint_32 mng_features)
+{
+   png_debug(1, "in png_permit_mng_features");
+
+   if (png_ptr == NULL)
+      return 0;
+
+   png_ptr->mng_features_permitted = mng_features & PNG_ALL_MNG_FEATURES;
+
+   return png_ptr->mng_features_permitted;
+}
+#endif
+
+#ifdef PNG_HANDLE_AS_UNKNOWN_SUPPORTED
+static unsigned int
+add_one_chunk(png_bytep list, unsigned int count, png_const_bytep add, int keep)
+{
+   unsigned int i;
+
+   /* Utility function: update the 'keep' state of a chunk if it is already in
+    * the list, otherwise add it to the list.
+    */
+   for (i=0; i<count; ++i, list += 5) if (memcmp(list, add, 4) == 0)
+   {
+      list[4] = (png_byte)keep;
+      return count;
+   }
+
+   if (keep != PNG_HANDLE_CHUNK_AS_DEFAULT)
+   {
+      ++count;
+      memcpy(list, add, 4);
+      list[4] = (png_byte)keep;
+   }
+
+   return count;
+}
+
+void PNGAPI
+png_set_keep_unknown_chunks(png_structrp png_ptr, int keep,
+    png_const_bytep chunk_list, int num_chunks_in)
+{
+   png_bytep new_list;
+   unsigned int num_chunks, old_num_chunks;
+
+   if (png_ptr == NULL)
+      return;
+
+   if (keep < 0 || keep >= PNG_HANDLE_CHUNK_LAST)
+   {
+      png_app_error(png_ptr, "png_set_keep_unknown_chunks: invalid keep");
+      return;
+   }
+
+   if (num_chunks_in <= 0)
+   {
+      png_ptr->unknown_default = keep;
+
+      /* '0' means just set the flags, so stop here */
+      if (num_chunks_in == 0)
+        return;
+   }
+
+   if (num_chunks_in < 0)
+   {
+      /* Ignore all unknown chunks and all chunks recognized by
+       * libpng except for IHDR, PLTE, tRNS, IDAT, and IEND
+       */
+      static PNG_CONST png_byte chunks_to_ignore[] = {
+         98,  75,  71,  68, '\0',  /* bKGD */
+         99,  72,  82,  77, '\0',  /* cHRM */
+        103,  65,  77,  65, '\0',  /* gAMA */
+        104,  73,  83,  84, '\0',  /* hIST */
+        105,  67,  67,  80, '\0',  /* iCCP */
+        105,  84,  88, 116, '\0',  /* iTXt */
+        111,  70,  70, 115, '\0',  /* oFFs */
+        112,  67,  65,  76, '\0',  /* pCAL */
+        112,  72,  89, 115, '\0',  /* pHYs */
+        115,  66,  73,  84, '\0',  /* sBIT */
+        115,  67,  65,  76, '\0',  /* sCAL */
+        115,  80,  76,  84, '\0',  /* sPLT */
+        115,  84,  69,  82, '\0',  /* sTER */
+        115,  82,  71,  66, '\0',  /* sRGB */
+        116,  69,  88, 116, '\0',  /* tEXt */
+        116,  73,  77,  69, '\0',  /* tIME */
+        122,  84,  88, 116, '\0'   /* zTXt */
+      };
+
+      chunk_list = chunks_to_ignore;
+      num_chunks = (sizeof chunks_to_ignore)/5;
+   }
+
+   else /* num_chunks_in > 0 */
+   {
+      if (chunk_list == NULL)
+      {
+         /* Prior to 1.6.0 this was silently ignored, now it is an app_error
+          * which can be switched off.
+          */
+         png_app_error(png_ptr, "png_set_keep_unknown_chunks: no chunk list");
+         return;
+      }
+
+      num_chunks = num_chunks_in;
+   }
+
+   old_num_chunks = png_ptr->num_chunk_list;
+   if (png_ptr->chunk_list == NULL)
+      old_num_chunks = 0;
+
+   /* Since num_chunks is always restricted to UINT_MAX/5 this can't overflow.
+    */
+   if (num_chunks + old_num_chunks > UINT_MAX/5)
+   {
+      png_app_error(png_ptr, "png_set_keep_unknown_chunks: too many chunks");
+      return;
+   }
+
+   /* If these chunks are being reset to the default then no more memory is
+    * required because add_one_chunk above doesn't extend the list if the 'keep'
+    * parameter is the default.
+    */
+   if (keep)
+   {
+      new_list = png_voidcast(png_bytep, png_malloc(png_ptr,
+          5 * (num_chunks + old_num_chunks)));
+
+      if (old_num_chunks > 0)
+         memcpy(new_list, png_ptr->chunk_list, 5*old_num_chunks);
+   }
+
+   else if (old_num_chunks > 0)
+      new_list = png_ptr->chunk_list;
+
+   else
+      new_list = NULL;
+
+   /* Add the new chunks together with each one's handling code.  If the chunk
+    * already exists the code is updated, otherwise the chunk is added to the
+    * end.  (In libpng 1.6.0 order no longer matters because this code enforces
+    * the earlier convention that the last setting is the one that is used.)
+    */
+   if (new_list != NULL)
+   {
+      png_const_bytep inlist;
+      png_bytep outlist;
+      unsigned int i;
+
+      for (i=0; i<num_chunks; ++i)
+         old_num_chunks = add_one_chunk(new_list, old_num_chunks,
+            chunk_list+5*i, keep);
+
+      /* Now remove any spurious 'default' entries. */
+      num_chunks = 0;
+      for (i=0, inlist=outlist=new_list; i<old_num_chunks; ++i, inlist += 5)
+         if (inlist[4])
+         {
+            if (outlist != inlist)
+               memcpy(outlist, inlist, 5);
+            outlist += 5;
+            ++num_chunks;
+         }
+
+      /* This means the application has removed all the specialized handling. */
+      if (num_chunks == 0)
+      {
+         if (png_ptr->chunk_list != new_list)
+            png_free(png_ptr, new_list);
+
+         new_list = NULL;
+      }
+   }
+
+   else
+      num_chunks = 0;
+
+   png_ptr->num_chunk_list = num_chunks;
+
+   if (png_ptr->chunk_list != new_list)
+   {
+      if (png_ptr->chunk_list != NULL)
+         png_free(png_ptr, png_ptr->chunk_list);
+
+      png_ptr->chunk_list = new_list;
+   }
+}
+#endif
+
+#ifdef PNG_READ_USER_CHUNKS_SUPPORTED
+void PNGAPI
+png_set_read_user_chunk_fn(png_structrp png_ptr, png_voidp user_chunk_ptr,
+    png_user_chunk_ptr read_user_chunk_fn)
+{
+   png_debug(1, "in png_set_read_user_chunk_fn");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->read_user_chunk_fn = read_user_chunk_fn;
+   png_ptr->user_chunk_ptr = user_chunk_ptr;
+}
+#endif
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+void PNGAPI
+png_set_rows(png_const_structrp png_ptr, png_inforp info_ptr,
+    png_bytepp row_pointers)
+{
+   png_debug1(1, "in %s storage function", "rows");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (info_ptr->row_pointers && (info_ptr->row_pointers != row_pointers))
+      png_free_data(png_ptr, info_ptr, PNG_FREE_ROWS, 0);
+
+   info_ptr->row_pointers = row_pointers;
+
+   if (row_pointers)
+      info_ptr->valid |= PNG_INFO_IDAT;
+}
+#endif
+
+void PNGAPI
+png_set_compression_buffer_size(png_structrp png_ptr, png_size_t size)
+{
+    if (png_ptr == NULL)
+       return;
+
+    if (size == 0 || size > PNG_UINT_31_MAX)
+       png_error(png_ptr, "invalid compression buffer size");
+
+#  ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+      if (png_ptr->mode & PNG_IS_READ_STRUCT)
+      {
+         png_ptr->IDAT_read_size = (png_uint_32)size; /* checked above */
+         return;
+      }
+#  endif
+
+#  ifdef PNG_WRITE_SUPPORTED
+      if (!(png_ptr->mode & PNG_IS_READ_STRUCT))
+      {
+         if (png_ptr->zowner != 0)
+         {
+            png_warning(png_ptr,
+              "Compression buffer size cannot be changed because it is in use");
+            return;
+         }
+
+         if (size > ZLIB_IO_MAX)
+         {
+            png_warning(png_ptr,
+               "Compression buffer size limited to system maximum");
+            size = ZLIB_IO_MAX; /* must fit */
+         }
+
+         else if (size < 6)
+         {
+            /* Deflate will potentially go into an infinite loop on a SYNC_FLUSH
+             * if this is permitted.
+             */
+            png_warning(png_ptr,
+               "Compression buffer size cannot be reduced below 6");
+            return;
+         }
+
+         if (png_ptr->zbuffer_size != size)
+         {
+            png_free_buffer_list(png_ptr, &png_ptr->zbuffer_list);
+            png_ptr->zbuffer_size = (uInt)size;
+         }
+      }
+#  endif
+}
+
+void PNGAPI
+png_set_invalid(png_const_structrp png_ptr, png_inforp info_ptr, int mask)
+{
+   if (png_ptr && info_ptr)
+      info_ptr->valid &= ~mask;
+}
+
+
+#ifdef PNG_SET_USER_LIMITS_SUPPORTED
+/* This function was added to libpng 1.2.6 */
+void PNGAPI
+png_set_user_limits (png_structrp png_ptr, png_uint_32 user_width_max,
+    png_uint_32 user_height_max)
+{
+   /* Images with dimensions larger than these limits will be
+    * rejected by png_set_IHDR().  To accept any PNG datastream
+    * regardless of dimensions, set both limits to 0x7ffffffL.
+    */
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->user_width_max = user_width_max;
+   png_ptr->user_height_max = user_height_max;
+}
+
+/* This function was added to libpng 1.4.0 */
+void PNGAPI
+png_set_chunk_cache_max (png_structrp png_ptr, png_uint_32 user_chunk_cache_max)
+{
+    if (png_ptr)
+       png_ptr->user_chunk_cache_max = user_chunk_cache_max;
+}
+
+/* This function was added to libpng 1.4.1 */
+void PNGAPI
+png_set_chunk_malloc_max (png_structrp png_ptr,
+    png_alloc_size_t user_chunk_malloc_max)
+{
+   if (png_ptr)
+      png_ptr->user_chunk_malloc_max = user_chunk_malloc_max;
+}
+#endif /* ?PNG_SET_USER_LIMITS_SUPPORTED */
+
+
+#ifdef PNG_BENIGN_ERRORS_SUPPORTED
+void PNGAPI
+png_set_benign_errors(png_structrp png_ptr, int allowed)
+{
+   png_debug(1, "in png_set_benign_errors");
+
+   /* If allowed is 1, png_benign_error() is treated as a warning.
+    *
+    * If allowed is 0, png_benign_error() is treated as an error (which
+    * is the default behavior if png_set_benign_errors() is not called).
+    */
+
+   if (allowed)
+      png_ptr->flags |= PNG_FLAG_BENIGN_ERRORS_WARN |
+         PNG_FLAG_APP_WARNINGS_WARN | PNG_FLAG_APP_ERRORS_WARN;
+
+   else
+      png_ptr->flags &= ~(PNG_FLAG_BENIGN_ERRORS_WARN |
+         PNG_FLAG_APP_WARNINGS_WARN | PNG_FLAG_APP_ERRORS_WARN);
+}
+#endif /* PNG_BENIGN_ERRORS_SUPPORTED */
+
+#ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   /* Whether to report invalid palette index; added at libng-1.5.10.
+    * It is possible for an indexed (color-type==3) PNG file to contain
+    * pixels with invalid (out-of-range) indexes if the PLTE chunk has
+    * fewer entries than the image's bit-depth would allow. We recover
+    * from this gracefully by filling any incomplete palette with zeroes
+    * (opaque black).  By default, when this occurs libpng will issue
+    * a benign error.  This API can be used to override that behavior.
+    */
+void PNGAPI
+png_set_check_for_invalid_index(png_structrp png_ptr, int allowed)
+{
+   png_debug(1, "in png_set_check_for_invalid_index");
+
+   if (allowed > 0)
+      png_ptr->num_palette_max = 0;
+
+   else
+      png_ptr->num_palette_max = -1;
+}
+#endif
+#endif /* PNG_READ_SUPPORTED || PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngstruct.h b/ml/dlib/dlib/external/libpng/pngstruct.h
new file mode 100644
index 000000000..d58c02884
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngstruct.h
@@ -0,0 +1,489 @@
+
+/* pngstruct.h - header file for PNG reference library
+ *
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * Last changed in libpng 1.6.1 [March 28, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* The structure that holds the information to read and write PNG files.
+ * The only people who need to care about what is inside of this are the
+ * people who will be modifying the library for their own special needs.
+ * It should NOT be accessed directly by an application.
+ */
+
+#ifndef PNGSTRUCT_H
+#define PNGSTRUCT_H
+/* zlib.h defines the structure z_stream, an instance of which is included
+ * in this structure and is required for decompressing the LZ compressed
+ * data in PNG files.
+ */
+#ifndef ZLIB_CONST
+   /* We must ensure that zlib uses 'const' in declarations. */
+#  define ZLIB_CONST
+#endif
+#include "zlib.h"
+#ifdef const
+   /* zlib.h sometimes #defines const to nothing, undo this. */
+#  undef const
+#endif
+
+/* zlib.h has mediocre z_const use before 1.2.6, this stuff is for compatibility
+ * with older builds.
+ */
+#if ZLIB_VERNUM < 0x1260
+#  define PNGZ_MSG_CAST(s) png_constcast(char*,s)
+#  define PNGZ_INPUT_CAST(b) png_constcast(png_bytep,b)
+#else
+#  define PNGZ_MSG_CAST(s) (s)
+#  define PNGZ_INPUT_CAST(b) (b)
+#endif
+
+/* zlib.h declares a magic type 'uInt' that limits the amount of data that zlib
+ * can handle at once.  This type need be no larger than 16 bits (so maximum of
+ * 65535), this define allows us to discover how big it is, but limited by the
+ * maximuum for png_size_t.  The value can be overriden in a library build
+ * (pngusr.h, or set it in CPPFLAGS) and it works to set it to a considerably
+ * lower value (e.g. 255 works).  A lower value may help memory usage (slightly)
+ * and may even improve performance on some systems (and degrade it on others.)
+ */
+#ifndef ZLIB_IO_MAX
+#  define ZLIB_IO_MAX ((uInt)-1)
+#endif
+
+#ifdef PNG_WRITE_SUPPORTED
+/* The type of a compression buffer list used by the write code. */
+typedef struct png_compression_buffer
+{
+   struct png_compression_buffer *next;
+   png_byte                       output[1]; /* actually zbuf_size */
+} png_compression_buffer, *png_compression_bufferp;
+
+#define PNG_COMPRESSION_BUFFER_SIZE(pp)\
+   (offsetof(png_compression_buffer, output) + (pp)->zbuffer_size)
+#endif
+
+/* Colorspace support; structures used in png_struct, png_info and in internal
+ * functions to hold and communicate information about the color space.
+ *
+ * PNG_COLORSPACE_SUPPORTED is only required if the application will perform
+ * colorspace corrections, otherwise all the colorspace information can be
+ * skipped and the size of libpng can be reduced (significantly) by compiling
+ * out the colorspace support.
+ */
+#ifdef PNG_COLORSPACE_SUPPORTED
+/* The chromaticities of the red, green and blue colorants and the chromaticity
+ * of the corresponding white point (i.e. of rgb(1.0,1.0,1.0)).
+ */
+typedef struct png_xy
+{
+   png_fixed_point redx, redy;
+   png_fixed_point greenx, greeny;
+   png_fixed_point bluex, bluey;
+   png_fixed_point whitex, whitey;
+} png_xy;
+
+/* The same data as above but encoded as CIE XYZ values.  When this data comes
+ * from chromaticities the sum of the Y values is assumed to be 1.0
+ */
+typedef struct png_XYZ
+{
+   png_fixed_point red_X, red_Y, red_Z;
+   png_fixed_point green_X, green_Y, green_Z;
+   png_fixed_point blue_X, blue_Y, blue_Z;
+} png_XYZ;
+#endif /* COLORSPACE */
+
+#if defined(PNG_COLORSPACE_SUPPORTED) || defined(PNG_GAMMA_SUPPORTED)
+/* A colorspace is all the above plus, potentially, profile information,
+ * however at present libpng does not use the profile internally so it is only
+ * stored in the png_info struct (if iCCP is supported.)  The rendering intent
+ * is retained here and is checked.
+ *
+ * The file gamma encoding information is also stored here and gamma correction
+ * is done by libpng, whereas color correction must currently be done by the
+ * application.
+ */
+typedef struct png_colorspace
+{
+#ifdef PNG_GAMMA_SUPPORTED
+   png_fixed_point gamma;        /* File gamma */
+#endif
+
+#ifdef PNG_COLORSPACE_SUPPORTED
+   png_xy      end_points_xy;    /* End points as chromaticities */
+   png_XYZ     end_points_XYZ;   /* End points as CIE XYZ colorant values */
+   png_uint_16 rendering_intent; /* Rendering intent of a profile */
+#endif
+
+   /* Flags are always defined to simplify the code. */
+   png_uint_16 flags;            /* As defined below */
+} png_colorspace, * PNG_RESTRICT png_colorspacerp;
+
+typedef const png_colorspace * PNG_RESTRICT png_const_colorspacerp;
+
+/* General flags for the 'flags' field */
+#define PNG_COLORSPACE_HAVE_GAMMA           0x0001
+#define PNG_COLORSPACE_HAVE_ENDPOINTS       0x0002
+#define PNG_COLORSPACE_HAVE_INTENT          0x0004
+#define PNG_COLORSPACE_FROM_gAMA            0x0008
+#define PNG_COLORSPACE_FROM_cHRM            0x0010
+#define PNG_COLORSPACE_FROM_sRGB            0x0020
+#define PNG_COLORSPACE_ENDPOINTS_MATCH_sRGB 0x0040
+#define PNG_COLORSPACE_MATCHES_sRGB         0x0080 /* exact match on profile */
+#define PNG_COLORSPACE_INVALID              0x8000
+#define PNG_COLORSPACE_CANCEL(flags)        (0xffff ^ (flags))
+#endif /* COLORSPACE || GAMMA */
+
+struct png_struct_def
+{
+#ifdef PNG_SETJMP_SUPPORTED
+   jmp_buf jmp_buf_local;     /* New name in 1.6.0 for jmp_buf in png_struct */
+   png_longjmp_ptr longjmp_fn;/* setjmp non-local goto function. */
+   jmp_buf *jmp_buf_ptr;      /* passed to longjmp_fn */
+   size_t jmp_buf_size;       /* size of the above, if allocated */
+#endif
+   png_error_ptr error_fn;    /* function for printing errors and aborting */
+#ifdef PNG_WARNINGS_SUPPORTED
+   png_error_ptr warning_fn;  /* function for printing warnings */
+#endif
+   png_voidp error_ptr;       /* user supplied struct for error functions */
+   png_rw_ptr write_data_fn;  /* function for writing output data */
+   png_rw_ptr read_data_fn;   /* function for reading input data */
+   png_voidp io_ptr;          /* ptr to application struct for I/O functions */
+
+#ifdef PNG_READ_USER_TRANSFORM_SUPPORTED
+   png_user_transform_ptr read_user_transform_fn; /* user read transform */
+#endif
+
+#ifdef PNG_WRITE_USER_TRANSFORM_SUPPORTED
+   png_user_transform_ptr write_user_transform_fn; /* user write transform */
+#endif
+
+/* These were added in libpng-1.0.2 */
+#ifdef PNG_USER_TRANSFORM_PTR_SUPPORTED
+#if defined(PNG_READ_USER_TRANSFORM_SUPPORTED) || \
+    defined(PNG_WRITE_USER_TRANSFORM_SUPPORTED)
+   png_voidp user_transform_ptr; /* user supplied struct for user transform */
+   png_byte user_transform_depth;    /* bit depth of user transformed pixels */
+   png_byte user_transform_channels; /* channels in user transformed pixels */
+#endif
+#endif
+
+   png_uint_32 mode;          /* tells us where we are in the PNG file */
+   png_uint_32 flags;         /* flags indicating various things to libpng */
+   png_uint_32 transformations; /* which transformations to perform */
+
+   png_uint_32 zowner;        /* ID (chunk type) of zstream owner, 0 if none */
+   z_stream    zstream;       /* decompression structure */
+
+#ifdef PNG_WRITE_SUPPORTED
+   png_compression_bufferp zbuffer_list; /* Created on demand during write */
+   uInt                    zbuffer_size; /* size of the actual buffer */
+
+   int zlib_level;            /* holds zlib compression level */
+   int zlib_method;           /* holds zlib compression method */
+   int zlib_window_bits;      /* holds zlib compression window bits */
+   int zlib_mem_level;        /* holds zlib compression memory level */
+   int zlib_strategy;         /* holds zlib compression strategy */
+#endif
+/* Added at libpng 1.5.4 */
+#ifdef PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED
+   int zlib_text_level;            /* holds zlib compression level */
+   int zlib_text_method;           /* holds zlib compression method */
+   int zlib_text_window_bits;      /* holds zlib compression window bits */
+   int zlib_text_mem_level;        /* holds zlib compression memory level */
+   int zlib_text_strategy;         /* holds zlib compression strategy */
+#endif
+/* End of material added at libpng 1.5.4 */
+/* Added at libpng 1.6.0 */
+#ifdef PNG_WRITE_SUPPORTED
+   int zlib_set_level;        /* Actual values set into the zstream on write */
+   int zlib_set_method;
+   int zlib_set_window_bits;
+   int zlib_set_mem_level;
+   int zlib_set_strategy;
+#endif
+
+   png_uint_32 width;         /* width of image in pixels */
+   png_uint_32 height;        /* height of image in pixels */
+   png_uint_32 num_rows;      /* number of rows in current pass */
+   png_uint_32 usr_width;     /* width of row at start of write */
+   png_size_t rowbytes;       /* size of row in bytes */
+   png_uint_32 iwidth;        /* width of current interlaced row in pixels */
+   png_uint_32 row_number;    /* current row in interlace pass */
+   png_uint_32 chunk_name;    /* PNG_CHUNK() id of current chunk */
+   png_bytep prev_row;        /* buffer to save previous (unfiltered) row.
+                               * This is a pointer into big_prev_row
+                               */
+   png_bytep row_buf;         /* buffer to save current (unfiltered) row.
+                               * This is a pointer into big_row_buf
+                               */
+#ifdef PNG_WRITE_SUPPORTED
+   png_bytep sub_row;         /* buffer to save "sub" row when filtering */
+   png_bytep up_row;          /* buffer to save "up" row when filtering */
+   png_bytep avg_row;         /* buffer to save "avg" row when filtering */
+   png_bytep paeth_row;       /* buffer to save "Paeth" row when filtering */
+#endif
+   png_size_t info_rowbytes;  /* Added in 1.5.4: cache of updated row bytes */
+
+   png_uint_32 idat_size;     /* current IDAT size for read */
+   png_uint_32 crc;           /* current chunk CRC value */
+   png_colorp palette;        /* palette from the input file */
+   png_uint_16 num_palette;   /* number of color entries in palette */
+
+/* Added at libpng-1.5.10 */
+#ifdef PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   int num_palette_max;       /* maximum palette index found in IDAT */
+#endif
+
+   png_uint_16 num_trans;     /* number of transparency values */
+   png_byte compression;      /* file compression type (always 0) */
+   png_byte filter;           /* file filter type (always 0) */
+   png_byte interlaced;       /* PNG_INTERLACE_NONE, PNG_INTERLACE_ADAM7 */
+   png_byte pass;             /* current interlace pass (0 - 6) */
+   png_byte do_filter;        /* row filter flags (see PNG_FILTER_ below ) */
+   png_byte color_type;       /* color type of file */
+   png_byte bit_depth;        /* bit depth of file */
+   png_byte usr_bit_depth;    /* bit depth of users row: write only */
+   png_byte pixel_depth;      /* number of bits per pixel */
+   png_byte channels;         /* number of channels in file */
+#ifdef PNG_WRITE_SUPPORTED
+   png_byte usr_channels;     /* channels at start of write: write only */
+#endif
+   png_byte sig_bytes;        /* magic bytes read/written from start of file */
+   png_byte maximum_pixel_depth;
+                              /* pixel depth used for the row buffers */
+   png_byte transformed_pixel_depth;
+                              /* pixel depth after read/write transforms */
+#if defined(PNG_READ_FILLER_SUPPORTED) || defined(PNG_WRITE_FILLER_SUPPORTED)
+   png_uint_16 filler;           /* filler bytes for pixel expansion */
+#endif
+
+#if defined(PNG_bKGD_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED) ||\
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED)
+   png_byte background_gamma_type;
+   png_fixed_point background_gamma;
+   png_color_16 background;   /* background color in screen gamma space */
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   png_color_16 background_1; /* background normalized to gamma 1.0 */
+#endif
+#endif /* PNG_bKGD_SUPPORTED */
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+   png_flush_ptr output_flush_fn; /* Function for flushing output */
+   png_uint_32 flush_dist;    /* how many rows apart to flush, 0 - no flush */
+   png_uint_32 flush_rows;    /* number of rows written since last flush */
+#endif
+
+#ifdef PNG_READ_GAMMA_SUPPORTED
+   int gamma_shift;      /* number of "insignificant" bits in 16-bit gamma */
+   png_fixed_point screen_gamma; /* screen gamma value (display_exponent) */
+
+   png_bytep gamma_table;     /* gamma table for 8-bit depth files */
+   png_uint_16pp gamma_16_table; /* gamma table for 16-bit depth files */
+#if defined(PNG_READ_BACKGROUND_SUPPORTED) || \
+   defined(PNG_READ_ALPHA_MODE_SUPPORTED) || \
+   defined(PNG_READ_RGB_TO_GRAY_SUPPORTED)
+   png_bytep gamma_from_1;    /* converts from 1.0 to screen */
+   png_bytep gamma_to_1;      /* converts from file to 1.0 */
+   png_uint_16pp gamma_16_from_1; /* converts from 1.0 to screen */
+   png_uint_16pp gamma_16_to_1; /* converts from file to 1.0 */
+#endif /* READ_BACKGROUND || READ_ALPHA_MODE || RGB_TO_GRAY */
+#endif
+
+#if defined(PNG_READ_GAMMA_SUPPORTED) || defined(PNG_sBIT_SUPPORTED)
+   png_color_8 sig_bit;       /* significant bits in each available channel */
+#endif
+
+#if defined(PNG_READ_SHIFT_SUPPORTED) || defined(PNG_WRITE_SHIFT_SUPPORTED)
+   png_color_8 shift;         /* shift for significant bit tranformation */
+#endif
+
+#if defined(PNG_tRNS_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED) \
+ || defined(PNG_READ_EXPAND_SUPPORTED) || defined(PNG_READ_BACKGROUND_SUPPORTED)
+   png_bytep trans_alpha;           /* alpha values for paletted files */
+   png_color_16 trans_color;  /* transparent color for non-paletted files */
+#endif
+
+   png_read_status_ptr read_row_fn;   /* called after each row is decoded */
+   png_write_status_ptr write_row_fn; /* called after each row is encoded */
+#ifdef PNG_PROGRESSIVE_READ_SUPPORTED
+   png_progressive_info_ptr info_fn; /* called after header data fully read */
+   png_progressive_row_ptr row_fn;   /* called after a prog. row is decoded */
+   png_progressive_end_ptr end_fn;   /* called after image is complete */
+   png_bytep save_buffer_ptr;        /* current location in save_buffer */
+   png_bytep save_buffer;            /* buffer for previously read data */
+   png_bytep current_buffer_ptr;     /* current location in current_buffer */
+   png_bytep current_buffer;         /* buffer for recently used data */
+   png_uint_32 push_length;          /* size of current input chunk */
+   png_uint_32 skip_length;          /* bytes to skip in input data */
+   png_size_t save_buffer_size;      /* amount of data now in save_buffer */
+   png_size_t save_buffer_max;       /* total size of save_buffer */
+   png_size_t buffer_size;           /* total amount of available input data */
+   png_size_t current_buffer_size;   /* amount of data now in current_buffer */
+   int process_mode;                 /* what push library is currently doing */
+   int cur_palette;                  /* current push library palette index */
+
+#endif /* PNG_PROGRESSIVE_READ_SUPPORTED */
+
+#if defined(__TURBOC__) && !defined(_Windows) && !defined(__FLAT__)
+/* For the Borland special 64K segment handler */
+   png_bytepp offset_table_ptr;
+   png_bytep offset_table;
+   png_uint_16 offset_table_number;
+   png_uint_16 offset_table_count;
+   png_uint_16 offset_table_count_free;
+#endif
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+   png_bytep palette_lookup; /* lookup table for quantizing */
+   png_bytep quantize_index; /* index translation for palette files */
+#endif
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+   png_byte heuristic_method;        /* heuristic for row filter selection */
+   png_byte num_prev_filters;        /* number of weights for previous rows */
+   png_bytep prev_filters;           /* filter type(s) of previous row(s) */
+   png_uint_16p filter_weights;      /* weight(s) for previous line(s) */
+   png_uint_16p inv_filter_weights;  /* 1/weight(s) for previous line(s) */
+   png_uint_16p filter_costs;        /* relative filter calculation cost */
+   png_uint_16p inv_filter_costs;    /* 1/relative filter calculation cost */
+#endif
+
+   /* Options */
+#ifdef PNG_SET_OPTION_SUPPORTED
+   png_byte options;           /* On/off state (up to 4 options) */
+#endif
+
+#if PNG_LIBPNG_VER < 10700
+/* To do: remove this from libpng-1.7 */
+#ifdef PNG_TIME_RFC1123_SUPPORTED
+   char time_buffer[29]; /* String to hold RFC 1123 time text */
+#endif
+#endif
+
+/* New members added in libpng-1.0.6 */
+
+   png_uint_32 free_me;    /* flags items libpng is responsible for freeing */
+
+#ifdef PNG_USER_CHUNKS_SUPPORTED
+   png_voidp user_chunk_ptr;
+#ifdef PNG_READ_USER_CHUNKS_SUPPORTED
+   png_user_chunk_ptr read_user_chunk_fn; /* user read chunk handler */
+#endif
+#endif
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+   int          unknown_default; /* As PNG_HANDLE_* */
+   unsigned int num_chunk_list;  /* Number of entries in the list */
+   png_bytep    chunk_list;      /* List of png_byte[5]; the textual chunk name
+                                  * followed by a PNG_HANDLE_* byte */
+#endif
+
+/* New members added in libpng-1.0.3 */
+#ifdef PNG_READ_RGB_TO_GRAY_SUPPORTED
+   png_byte rgb_to_gray_status;
+   /* Added in libpng 1.5.5 to record setting of coefficients: */
+   png_byte rgb_to_gray_coefficients_set;
+   /* These were changed from png_byte in libpng-1.0.6 */
+   png_uint_16 rgb_to_gray_red_coeff;
+   png_uint_16 rgb_to_gray_green_coeff;
+   /* deleted in 1.5.5: rgb_to_gray_blue_coeff; */
+#endif
+
+/* New member added in libpng-1.0.4 (renamed in 1.0.9) */
+#if defined(PNG_MNG_FEATURES_SUPPORTED)
+/* Changed from png_byte to png_uint_32 at version 1.2.0 */
+   png_uint_32 mng_features_permitted;
+#endif
+
+/* New member added in libpng-1.0.9, ifdef'ed out in 1.0.12, enabled in 1.2.0 */
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   png_byte filter_type;
+#endif
+
+/* New members added in libpng-1.2.0 */
+
+/* New members added in libpng-1.0.2 but first enabled by default in 1.2.0 */
+#ifdef PNG_USER_MEM_SUPPORTED
+   png_voidp mem_ptr;             /* user supplied struct for mem functions */
+   png_malloc_ptr malloc_fn;      /* function for allocating memory */
+   png_free_ptr free_fn;          /* function for freeing memory */
+#endif
+
+/* New member added in libpng-1.0.13 and 1.2.0 */
+   png_bytep big_row_buf;         /* buffer to save current (unfiltered) row */
+
+#ifdef PNG_READ_QUANTIZE_SUPPORTED
+/* The following three members were added at version 1.0.14 and 1.2.4 */
+   png_bytep quantize_sort;          /* working sort array */
+   png_bytep index_to_palette;       /* where the original index currently is
+                                        in the palette */
+   png_bytep palette_to_index;       /* which original index points to this
+                                         palette color */
+#endif
+
+/* New members added in libpng-1.0.16 and 1.2.6 */
+   png_byte compression_type;
+
+#ifdef PNG_USER_LIMITS_SUPPORTED
+   png_uint_32 user_width_max;
+   png_uint_32 user_height_max;
+
+   /* Added in libpng-1.4.0: Total number of sPLT, text, and unknown
+    * chunks that can be stored (0 means unlimited).
+    */
+   png_uint_32 user_chunk_cache_max;
+
+   /* Total memory that a zTXt, sPLT, iTXt, iCCP, or unknown chunk
+    * can occupy when decompressed.  0 means unlimited.
+    */
+   png_alloc_size_t user_chunk_malloc_max;
+#endif
+
+/* New member added in libpng-1.0.25 and 1.2.17 */
+#ifdef PNG_READ_UNKNOWN_CHUNKS_SUPPORTED
+   /* Temporary storage for unknown chunk that the library doesn't recognize,
+    * used while reading the chunk.
+    */
+   png_unknown_chunk unknown_chunk;
+#endif
+
+/* New member added in libpng-1.2.26 */
+  png_size_t old_big_row_buf_size;
+
+#ifdef PNG_READ_SUPPORTED
+/* New member added in libpng-1.2.30 */
+  png_bytep        read_buffer;      /* buffer for reading chunk data */
+  png_alloc_size_t read_buffer_size; /* current size of the buffer */
+#endif
+#ifdef PNG_SEQUENTIAL_READ_SUPPORTED
+  uInt             IDAT_read_size;   /* limit on read buffer size for IDAT */
+#endif
+
+#ifdef PNG_IO_STATE_SUPPORTED
+/* New member added in libpng-1.4.0 */
+   png_uint_32 io_state;
+#endif
+
+/* New member added in libpng-1.5.6 */
+   png_bytep big_prev_row;
+
+/* New member added in libpng-1.5.7 */
+   void (*read_filter[PNG_FILTER_VALUE_LAST-1])(png_row_infop row_info,
+      png_bytep row, png_const_bytep prev_row);
+
+#ifdef PNG_READ_SUPPORTED
+#if defined(PNG_COLORSPACE_SUPPORTED) || defined(PNG_GAMMA_SUPPORTED)
+   png_colorspace   colorspace;
+#endif
+#endif
+};
+#endif /* PNGSTRUCT_H */
diff --git a/ml/dlib/dlib/external/libpng/pngtrans.c b/ml/dlib/dlib/external/libpng/pngtrans.c
new file mode 100644
index 000000000..8f8bc5d9e
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngtrans.c
@@ -0,0 +1,841 @@
+
+/* pngtrans.c - transforms the data in a row (used by both readers and writers)
+ *
+ * Last changed in libpng 1.6.2 [April 25, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+
+#if defined(PNG_READ_SUPPORTED) || defined(PNG_WRITE_SUPPORTED)
+
+#if defined(PNG_READ_BGR_SUPPORTED) || defined(PNG_WRITE_BGR_SUPPORTED)
+/* Turn on BGR-to-RGB mapping */
+void PNGAPI
+png_set_bgr(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_bgr");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_BGR;
+}
+#endif
+
+#if defined(PNG_READ_SWAP_SUPPORTED) || defined(PNG_WRITE_SWAP_SUPPORTED)
+/* Turn on 16 bit byte swapping */
+void PNGAPI
+png_set_swap(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_swap");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (png_ptr->bit_depth == 16)
+      png_ptr->transformations |= PNG_SWAP_BYTES;
+}
+#endif
+
+#if defined(PNG_READ_PACK_SUPPORTED) || defined(PNG_WRITE_PACK_SUPPORTED)
+/* Turn on pixel packing */
+void PNGAPI
+png_set_packing(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_packing");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (png_ptr->bit_depth < 8)
+   {
+      png_ptr->transformations |= PNG_PACK;
+      png_ptr->usr_bit_depth = 8;
+   }
+}
+#endif
+
+#if defined(PNG_READ_PACKSWAP_SUPPORTED)||defined(PNG_WRITE_PACKSWAP_SUPPORTED)
+/* Turn on packed pixel swapping */
+void PNGAPI
+png_set_packswap(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_packswap");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (png_ptr->bit_depth < 8)
+      png_ptr->transformations |= PNG_PACKSWAP;
+}
+#endif
+
+#if defined(PNG_READ_SHIFT_SUPPORTED) || defined(PNG_WRITE_SHIFT_SUPPORTED)
+void PNGAPI
+png_set_shift(png_structrp png_ptr, png_const_color_8p true_bits)
+{
+   png_debug(1, "in png_set_shift");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_SHIFT;
+   png_ptr->shift = *true_bits;
+}
+#endif
+
+#if defined(PNG_READ_INTERLACING_SUPPORTED) || \
+    defined(PNG_WRITE_INTERLACING_SUPPORTED)
+int PNGAPI
+png_set_interlace_handling(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_interlace handling");
+
+   if (png_ptr && png_ptr->interlaced)
+   {
+      png_ptr->transformations |= PNG_INTERLACE;
+      return (7);
+   }
+
+   return (1);
+}
+#endif
+
+#if defined(PNG_READ_FILLER_SUPPORTED) || defined(PNG_WRITE_FILLER_SUPPORTED)
+/* Add a filler byte on read, or remove a filler or alpha byte on write.
+ * The filler type has changed in v0.95 to allow future 2-byte fillers
+ * for 48-bit input data, as well as to avoid problems with some compilers
+ * that don't like bytes as parameters.
+ */
+void PNGAPI
+png_set_filler(png_structrp png_ptr, png_uint_32 filler, int filler_loc)
+{
+   png_debug(1, "in png_set_filler");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* In libpng 1.6 it is possible to determine whether this is a read or write
+    * operation and therefore to do more checking here for a valid call.
+    */
+   if (png_ptr->mode & PNG_IS_READ_STRUCT)
+   {
+#     ifdef PNG_READ_FILLER_SUPPORTED
+         /* On read png_set_filler is always valid, regardless of the base PNG
+          * format, because other transformations can give a format where the
+          * filler code can execute (basically an 8 or 16-bit component RGB or G
+          * format.)
+          *
+          * NOTE: usr_channels is not used by the read code!  (This has led to
+          * confusion in the past.)  The filler is only used in the read code.
+          */
+         png_ptr->filler = (png_uint_16)filler;
+#     else
+         png_app_error(png_ptr, "png_set_filler not supported on read");
+         PNG_UNUSED(filler) /* not used in the write case */
+         return;
+#     endif
+   }
+
+   else /* write */
+   {
+#     ifdef PNG_WRITE_FILLER_SUPPORTED
+         /* On write the usr_channels parameter must be set correctly at the
+          * start to record the number of channels in the app-supplied data.
+          */
+         switch (png_ptr->color_type)
+         {
+            case PNG_COLOR_TYPE_RGB:
+               png_ptr->usr_channels = 4;
+               break;
+
+            case PNG_COLOR_TYPE_GRAY:
+               if (png_ptr->bit_depth >= 8)
+               {
+                  png_ptr->usr_channels = 2;
+                  break;
+               }
+
+               else
+               {
+                  /* There simply isn't any code in libpng to strip out bits
+                   * from bytes when the components are less than a byte in
+                   * size!
+                   */
+                  png_app_error(png_ptr,
+                     "png_set_filler is invalid for low bit depth gray output");
+                  return;
+               }
+
+            default:
+               png_app_error(png_ptr,
+                  "png_set_filler: inappropriate color type");
+               return;
+         }
+#     else
+         png_app_error(png_ptr, "png_set_filler not supported on write");
+         return;
+#     endif
+   }
+
+   /* Here on success - libpng supports the operation, set the transformation
+    * and the flag to say where the filler channel is.
+    */
+   png_ptr->transformations |= PNG_FILLER;
+
+   if (filler_loc == PNG_FILLER_AFTER)
+      png_ptr->flags |= PNG_FLAG_FILLER_AFTER;
+
+   else
+      png_ptr->flags &= ~PNG_FLAG_FILLER_AFTER;
+}
+
+/* Added to libpng-1.2.7 */
+void PNGAPI
+png_set_add_alpha(png_structrp png_ptr, png_uint_32 filler, int filler_loc)
+{
+   png_debug(1, "in png_set_add_alpha");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_set_filler(png_ptr, filler, filler_loc);
+   /* The above may fail to do anything. */
+   if (png_ptr->transformations & PNG_FILLER)
+      png_ptr->transformations |= PNG_ADD_ALPHA;
+}
+
+#endif
+
+#if defined(PNG_READ_SWAP_ALPHA_SUPPORTED) || \
+    defined(PNG_WRITE_SWAP_ALPHA_SUPPORTED)
+void PNGAPI
+png_set_swap_alpha(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_swap_alpha");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_SWAP_ALPHA;
+}
+#endif
+
+#if defined(PNG_READ_INVERT_ALPHA_SUPPORTED) || \
+    defined(PNG_WRITE_INVERT_ALPHA_SUPPORTED)
+void PNGAPI
+png_set_invert_alpha(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_invert_alpha");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_INVERT_ALPHA;
+}
+#endif
+
+#if defined(PNG_READ_INVERT_SUPPORTED) || defined(PNG_WRITE_INVERT_SUPPORTED)
+void PNGAPI
+png_set_invert_mono(png_structrp png_ptr)
+{
+   png_debug(1, "in png_set_invert_mono");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_INVERT_MONO;
+}
+
+/* Invert monochrome grayscale data */
+void /* PRIVATE */
+png_do_invert(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_invert");
+
+  /* This test removed from libpng version 1.0.13 and 1.2.0:
+   *   if (row_info->bit_depth == 1 &&
+   */
+   if (row_info->color_type == PNG_COLOR_TYPE_GRAY)
+   {
+      png_bytep rp = row;
+      png_size_t i;
+      png_size_t istop = row_info->rowbytes;
+
+      for (i = 0; i < istop; i++)
+      {
+         *rp = (png_byte)(~(*rp));
+         rp++;
+      }
+   }
+
+   else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA &&
+      row_info->bit_depth == 8)
+   {
+      png_bytep rp = row;
+      png_size_t i;
+      png_size_t istop = row_info->rowbytes;
+
+      for (i = 0; i < istop; i += 2)
+      {
+         *rp = (png_byte)(~(*rp));
+         rp += 2;
+      }
+   }
+
+#ifdef PNG_16BIT_SUPPORTED
+   else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA &&
+      row_info->bit_depth == 16)
+   {
+      png_bytep rp = row;
+      png_size_t i;
+      png_size_t istop = row_info->rowbytes;
+
+      for (i = 0; i < istop; i += 4)
+      {
+         *rp = (png_byte)(~(*rp));
+         *(rp + 1) = (png_byte)(~(*(rp + 1)));
+         rp += 4;
+      }
+   }
+#endif
+}
+#endif
+
+#ifdef PNG_16BIT_SUPPORTED
+#if defined(PNG_READ_SWAP_SUPPORTED) || defined(PNG_WRITE_SWAP_SUPPORTED)
+/* Swaps byte order on 16 bit depth images */
+void /* PRIVATE */
+png_do_swap(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_swap");
+
+   if (row_info->bit_depth == 16)
+   {
+      png_bytep rp = row;
+      png_uint_32 i;
+      png_uint_32 istop= row_info->width * row_info->channels;
+
+      for (i = 0; i < istop; i++, rp += 2)
+      {
+         png_byte t = *rp;
+         *rp = *(rp + 1);
+         *(rp + 1) = t;
+      }
+   }
+}
+#endif
+#endif
+
+#if defined(PNG_READ_PACKSWAP_SUPPORTED)||defined(PNG_WRITE_PACKSWAP_SUPPORTED)
+static PNG_CONST png_byte onebppswaptable[256] = {
+   0x00, 0x80, 0x40, 0xC0, 0x20, 0xA0, 0x60, 0xE0,
+   0x10, 0x90, 0x50, 0xD0, 0x30, 0xB0, 0x70, 0xF0,
+   0x08, 0x88, 0x48, 0xC8, 0x28, 0xA8, 0x68, 0xE8,
+   0x18, 0x98, 0x58, 0xD8, 0x38, 0xB8, 0x78, 0xF8,
+   0x04, 0x84, 0x44, 0xC4, 0x24, 0xA4, 0x64, 0xE4,
+   0x14, 0x94, 0x54, 0xD4, 0x34, 0xB4, 0x74, 0xF4,
+   0x0C, 0x8C, 0x4C, 0xCC, 0x2C, 0xAC, 0x6C, 0xEC,
+   0x1C, 0x9C, 0x5C, 0xDC, 0x3C, 0xBC, 0x7C, 0xFC,
+   0x02, 0x82, 0x42, 0xC2, 0x22, 0xA2, 0x62, 0xE2,
+   0x12, 0x92, 0x52, 0xD2, 0x32, 0xB2, 0x72, 0xF2,
+   0x0A, 0x8A, 0x4A, 0xCA, 0x2A, 0xAA, 0x6A, 0xEA,
+   0x1A, 0x9A, 0x5A, 0xDA, 0x3A, 0xBA, 0x7A, 0xFA,
+   0x06, 0x86, 0x46, 0xC6, 0x26, 0xA6, 0x66, 0xE6,
+   0x16, 0x96, 0x56, 0xD6, 0x36, 0xB6, 0x76, 0xF6,
+   0x0E, 0x8E, 0x4E, 0xCE, 0x2E, 0xAE, 0x6E, 0xEE,
+   0x1E, 0x9E, 0x5E, 0xDE, 0x3E, 0xBE, 0x7E, 0xFE,
+   0x01, 0x81, 0x41, 0xC1, 0x21, 0xA1, 0x61, 0xE1,
+   0x11, 0x91, 0x51, 0xD1, 0x31, 0xB1, 0x71, 0xF1,
+   0x09, 0x89, 0x49, 0xC9, 0x29, 0xA9, 0x69, 0xE9,
+   0x19, 0x99, 0x59, 0xD9, 0x39, 0xB9, 0x79, 0xF9,
+   0x05, 0x85, 0x45, 0xC5, 0x25, 0xA5, 0x65, 0xE5,
+   0x15, 0x95, 0x55, 0xD5, 0x35, 0xB5, 0x75, 0xF5,
+   0x0D, 0x8D, 0x4D, 0xCD, 0x2D, 0xAD, 0x6D, 0xED,
+   0x1D, 0x9D, 0x5D, 0xDD, 0x3D, 0xBD, 0x7D, 0xFD,
+   0x03, 0x83, 0x43, 0xC3, 0x23, 0xA3, 0x63, 0xE3,
+   0x13, 0x93, 0x53, 0xD3, 0x33, 0xB3, 0x73, 0xF3,
+   0x0B, 0x8B, 0x4B, 0xCB, 0x2B, 0xAB, 0x6B, 0xEB,
+   0x1B, 0x9B, 0x5B, 0xDB, 0x3B, 0xBB, 0x7B, 0xFB,
+   0x07, 0x87, 0x47, 0xC7, 0x27, 0xA7, 0x67, 0xE7,
+   0x17, 0x97, 0x57, 0xD7, 0x37, 0xB7, 0x77, 0xF7,
+   0x0F, 0x8F, 0x4F, 0xCF, 0x2F, 0xAF, 0x6F, 0xEF,
+   0x1F, 0x9F, 0x5F, 0xDF, 0x3F, 0xBF, 0x7F, 0xFF
+};
+
+static PNG_CONST png_byte twobppswaptable[256] = {
+   0x00, 0x40, 0x80, 0xC0, 0x10, 0x50, 0x90, 0xD0,
+   0x20, 0x60, 0xA0, 0xE0, 0x30, 0x70, 0xB0, 0xF0,
+   0x04, 0x44, 0x84, 0xC4, 0x14, 0x54, 0x94, 0xD4,
+   0x24, 0x64, 0xA4, 0xE4, 0x34, 0x74, 0xB4, 0xF4,
+   0x08, 0x48, 0x88, 0xC8, 0x18, 0x58, 0x98, 0xD8,
+   0x28, 0x68, 0xA8, 0xE8, 0x38, 0x78, 0xB8, 0xF8,
+   0x0C, 0x4C, 0x8C, 0xCC, 0x1C, 0x5C, 0x9C, 0xDC,
+   0x2C, 0x6C, 0xAC, 0xEC, 0x3C, 0x7C, 0xBC, 0xFC,
+   0x01, 0x41, 0x81, 0xC1, 0x11, 0x51, 0x91, 0xD1,
+   0x21, 0x61, 0xA1, 0xE1, 0x31, 0x71, 0xB1, 0xF1,
+   0x05, 0x45, 0x85, 0xC5, 0x15, 0x55, 0x95, 0xD5,
+   0x25, 0x65, 0xA5, 0xE5, 0x35, 0x75, 0xB5, 0xF5,
+   0x09, 0x49, 0x89, 0xC9, 0x19, 0x59, 0x99, 0xD9,
+   0x29, 0x69, 0xA9, 0xE9, 0x39, 0x79, 0xB9, 0xF9,
+   0x0D, 0x4D, 0x8D, 0xCD, 0x1D, 0x5D, 0x9D, 0xDD,
+   0x2D, 0x6D, 0xAD, 0xED, 0x3D, 0x7D, 0xBD, 0xFD,
+   0x02, 0x42, 0x82, 0xC2, 0x12, 0x52, 0x92, 0xD2,
+   0x22, 0x62, 0xA2, 0xE2, 0x32, 0x72, 0xB2, 0xF2,
+   0x06, 0x46, 0x86, 0xC6, 0x16, 0x56, 0x96, 0xD6,
+   0x26, 0x66, 0xA6, 0xE6, 0x36, 0x76, 0xB6, 0xF6,
+   0x0A, 0x4A, 0x8A, 0xCA, 0x1A, 0x5A, 0x9A, 0xDA,
+   0x2A, 0x6A, 0xAA, 0xEA, 0x3A, 0x7A, 0xBA, 0xFA,
+   0x0E, 0x4E, 0x8E, 0xCE, 0x1E, 0x5E, 0x9E, 0xDE,
+   0x2E, 0x6E, 0xAE, 0xEE, 0x3E, 0x7E, 0xBE, 0xFE,
+   0x03, 0x43, 0x83, 0xC3, 0x13, 0x53, 0x93, 0xD3,
+   0x23, 0x63, 0xA3, 0xE3, 0x33, 0x73, 0xB3, 0xF3,
+   0x07, 0x47, 0x87, 0xC7, 0x17, 0x57, 0x97, 0xD7,
+   0x27, 0x67, 0xA7, 0xE7, 0x37, 0x77, 0xB7, 0xF7,
+   0x0B, 0x4B, 0x8B, 0xCB, 0x1B, 0x5B, 0x9B, 0xDB,
+   0x2B, 0x6B, 0xAB, 0xEB, 0x3B, 0x7B, 0xBB, 0xFB,
+   0x0F, 0x4F, 0x8F, 0xCF, 0x1F, 0x5F, 0x9F, 0xDF,
+   0x2F, 0x6F, 0xAF, 0xEF, 0x3F, 0x7F, 0xBF, 0xFF
+};
+
+static PNG_CONST png_byte fourbppswaptable[256] = {
+   0x00, 0x10, 0x20, 0x30, 0x40, 0x50, 0x60, 0x70,
+   0x80, 0x90, 0xA0, 0xB0, 0xC0, 0xD0, 0xE0, 0xF0,
+   0x01, 0x11, 0x21, 0x31, 0x41, 0x51, 0x61, 0x71,
+   0x81, 0x91, 0xA1, 0xB1, 0xC1, 0xD1, 0xE1, 0xF1,
+   0x02, 0x12, 0x22, 0x32, 0x42, 0x52, 0x62, 0x72,
+   0x82, 0x92, 0xA2, 0xB2, 0xC2, 0xD2, 0xE2, 0xF2,
+   0x03, 0x13, 0x23, 0x33, 0x43, 0x53, 0x63, 0x73,
+   0x83, 0x93, 0xA3, 0xB3, 0xC3, 0xD3, 0xE3, 0xF3,
+   0x04, 0x14, 0x24, 0x34, 0x44, 0x54, 0x64, 0x74,
+   0x84, 0x94, 0xA4, 0xB4, 0xC4, 0xD4, 0xE4, 0xF4,
+   0x05, 0x15, 0x25, 0x35, 0x45, 0x55, 0x65, 0x75,
+   0x85, 0x95, 0xA5, 0xB5, 0xC5, 0xD5, 0xE5, 0xF5,
+   0x06, 0x16, 0x26, 0x36, 0x46, 0x56, 0x66, 0x76,
+   0x86, 0x96, 0xA6, 0xB6, 0xC6, 0xD6, 0xE6, 0xF6,
+   0x07, 0x17, 0x27, 0x37, 0x47, 0x57, 0x67, 0x77,
+   0x87, 0x97, 0xA7, 0xB7, 0xC7, 0xD7, 0xE7, 0xF7,
+   0x08, 0x18, 0x28, 0x38, 0x48, 0x58, 0x68, 0x78,
+   0x88, 0x98, 0xA8, 0xB8, 0xC8, 0xD8, 0xE8, 0xF8,
+   0x09, 0x19, 0x29, 0x39, 0x49, 0x59, 0x69, 0x79,
+   0x89, 0x99, 0xA9, 0xB9, 0xC9, 0xD9, 0xE9, 0xF9,
+   0x0A, 0x1A, 0x2A, 0x3A, 0x4A, 0x5A, 0x6A, 0x7A,
+   0x8A, 0x9A, 0xAA, 0xBA, 0xCA, 0xDA, 0xEA, 0xFA,
+   0x0B, 0x1B, 0x2B, 0x3B, 0x4B, 0x5B, 0x6B, 0x7B,
+   0x8B, 0x9B, 0xAB, 0xBB, 0xCB, 0xDB, 0xEB, 0xFB,
+   0x0C, 0x1C, 0x2C, 0x3C, 0x4C, 0x5C, 0x6C, 0x7C,
+   0x8C, 0x9C, 0xAC, 0xBC, 0xCC, 0xDC, 0xEC, 0xFC,
+   0x0D, 0x1D, 0x2D, 0x3D, 0x4D, 0x5D, 0x6D, 0x7D,
+   0x8D, 0x9D, 0xAD, 0xBD, 0xCD, 0xDD, 0xED, 0xFD,
+   0x0E, 0x1E, 0x2E, 0x3E, 0x4E, 0x5E, 0x6E, 0x7E,
+   0x8E, 0x9E, 0xAE, 0xBE, 0xCE, 0xDE, 0xEE, 0xFE,
+   0x0F, 0x1F, 0x2F, 0x3F, 0x4F, 0x5F, 0x6F, 0x7F,
+   0x8F, 0x9F, 0xAF, 0xBF, 0xCF, 0xDF, 0xEF, 0xFF
+};
+
+/* Swaps pixel packing order within bytes */
+void /* PRIVATE */
+png_do_packswap(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_packswap");
+
+   if (row_info->bit_depth < 8)
+   {
+      png_bytep rp;
+      png_const_bytep end, table;
+
+      end = row + row_info->rowbytes;
+
+      if (row_info->bit_depth == 1)
+         table = onebppswaptable;
+
+      else if (row_info->bit_depth == 2)
+         table = twobppswaptable;
+
+      else if (row_info->bit_depth == 4)
+         table = fourbppswaptable;
+
+      else
+         return;
+
+      for (rp = row; rp < end; rp++)
+         *rp = table[*rp];
+   }
+}
+#endif /* PNG_READ_PACKSWAP_SUPPORTED or PNG_WRITE_PACKSWAP_SUPPORTED */
+
+#if defined(PNG_WRITE_FILLER_SUPPORTED) || \
+    defined(PNG_READ_STRIP_ALPHA_SUPPORTED)
+/* Remove a channel - this used to be 'png_do_strip_filler' but it used a
+ * somewhat weird combination of flags to determine what to do.  All the calls
+ * to png_do_strip_filler are changed in 1.5.2 to call this instead with the
+ * correct arguments.
+ *
+ * The routine isn't general - the channel must be the channel at the start or
+ * end (not in the middle) of each pixel.
+ */
+void /* PRIVATE */
+png_do_strip_channel(png_row_infop row_info, png_bytep row, int at_start)
+{
+   png_bytep sp = row; /* source pointer */
+   png_bytep dp = row; /* destination pointer */
+   png_bytep ep = row + row_info->rowbytes; /* One beyond end of row */
+
+   /* At the start sp will point to the first byte to copy and dp to where
+    * it is copied to.  ep always points just beyond the end of the row, so
+    * the loop simply copies (channels-1) channels until sp reaches ep.
+    *
+    * at_start:        0 -- convert AG, XG, ARGB, XRGB, AAGG, XXGG, etc.
+    *            nonzero -- convert GA, GX, RGBA, RGBX, GGAA, RRGGBBXX, etc.
+    */
+
+   /* GA, GX, XG cases */
+   if (row_info->channels == 2)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         if (at_start) /* Skip initial filler */
+            ++sp;
+         else          /* Skip initial channel and, for sp, the filler */
+            sp += 2, ++dp;
+
+         /* For a 1 pixel wide image there is nothing to do */
+         while (sp < ep)
+            *dp++ = *sp, sp += 2;
+
+         row_info->pixel_depth = 8;
+      }
+
+      else if (row_info->bit_depth == 16)
+      {
+         if (at_start) /* Skip initial filler */
+            sp += 2;
+         else          /* Skip initial channel and, for sp, the filler */
+            sp += 4, dp += 2;
+
+         while (sp < ep)
+            *dp++ = *sp++, *dp++ = *sp, sp += 3;
+
+         row_info->pixel_depth = 16;
+      }
+
+      else
+         return; /* bad bit depth */
+
+      row_info->channels = 1;
+
+      /* Finally fix the color type if it records an alpha channel */
+      if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+         row_info->color_type = PNG_COLOR_TYPE_GRAY;
+   }
+
+   /* RGBA, RGBX, XRGB cases */
+   else if (row_info->channels == 4)
+   {
+      if (row_info->bit_depth == 8)
+      {
+         if (at_start) /* Skip initial filler */
+            ++sp;
+         else          /* Skip initial channels and, for sp, the filler */
+            sp += 4, dp += 3;
+
+         /* Note that the loop adds 3 to dp and 4 to sp each time. */
+         while (sp < ep)
+            *dp++ = *sp++, *dp++ = *sp++, *dp++ = *sp, sp += 2;
+
+         row_info->pixel_depth = 24;
+      }
+
+      else if (row_info->bit_depth == 16)
+      {
+         if (at_start) /* Skip initial filler */
+            sp += 2;
+         else          /* Skip initial channels and, for sp, the filler */
+            sp += 8, dp += 6;
+
+         while (sp < ep)
+         {
+            /* Copy 6 bytes, skip 2 */
+            *dp++ = *sp++, *dp++ = *sp++;
+            *dp++ = *sp++, *dp++ = *sp++;
+            *dp++ = *sp++, *dp++ = *sp, sp += 3;
+         }
+
+         row_info->pixel_depth = 48;
+      }
+
+      else
+         return; /* bad bit depth */
+
+      row_info->channels = 3;
+
+      /* Finally fix the color type if it records an alpha channel */
+      if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+         row_info->color_type = PNG_COLOR_TYPE_RGB;
+   }
+
+   else
+      return; /* The filler channel has gone already */
+
+   /* Fix the rowbytes value. */
+   row_info->rowbytes = dp-row;
+}
+#endif
+
+#if defined(PNG_READ_BGR_SUPPORTED) || defined(PNG_WRITE_BGR_SUPPORTED)
+/* Swaps red and blue bytes within a pixel */
+void /* PRIVATE */
+png_do_bgr(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_bgr");
+
+   if ((row_info->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      png_uint_32 row_width = row_info->width;
+      if (row_info->bit_depth == 8)
+      {
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+         {
+            png_bytep rp;
+            png_uint_32 i;
+
+            for (i = 0, rp = row; i < row_width; i++, rp += 3)
+            {
+               png_byte save = *rp;
+               *rp = *(rp + 2);
+               *(rp + 2) = save;
+            }
+         }
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+         {
+            png_bytep rp;
+            png_uint_32 i;
+
+            for (i = 0, rp = row; i < row_width; i++, rp += 4)
+            {
+               png_byte save = *rp;
+               *rp = *(rp + 2);
+               *(rp + 2) = save;
+            }
+         }
+      }
+
+#ifdef PNG_16BIT_SUPPORTED
+      else if (row_info->bit_depth == 16)
+      {
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+         {
+            png_bytep rp;
+            png_uint_32 i;
+
+            for (i = 0, rp = row; i < row_width; i++, rp += 6)
+            {
+               png_byte save = *rp;
+               *rp = *(rp + 4);
+               *(rp + 4) = save;
+               save = *(rp + 1);
+               *(rp + 1) = *(rp + 5);
+               *(rp + 5) = save;
+            }
+         }
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+         {
+            png_bytep rp;
+            png_uint_32 i;
+
+            for (i = 0, rp = row; i < row_width; i++, rp += 8)
+            {
+               png_byte save = *rp;
+               *rp = *(rp + 4);
+               *(rp + 4) = save;
+               save = *(rp + 1);
+               *(rp + 1) = *(rp + 5);
+               *(rp + 5) = save;
+            }
+         }
+      }
+#endif
+   }
+}
+#endif /* PNG_READ_BGR_SUPPORTED or PNG_WRITE_BGR_SUPPORTED */
+
+#if defined(PNG_READ_CHECK_FOR_INVALID_INDEX_SUPPORTED) || \
+    defined(PNG_WRITE_CHECK_FOR_INVALID_INDEX_SUPPORTED)
+/* Added at libpng-1.5.10 */
+void /* PRIVATE */
+png_do_check_palette_indexes(png_structrp png_ptr, png_row_infop row_info)
+{
+   if (png_ptr->num_palette < (1 << row_info->bit_depth) &&
+      png_ptr->num_palette > 0) /* num_palette can be 0 in MNG files */
+   {
+      /* Calculations moved outside switch in an attempt to stop different
+       * compiler warnings.  'padding' is in *bits* within the last byte, it is
+       * an 'int' because pixel_depth becomes an 'int' in the expression below,
+       * and this calculation is used because it avoids warnings that other
+       * forms produced on either GCC or MSVC.
+       */
+      int padding = (-row_info->pixel_depth * row_info->width) & 7;
+      png_bytep rp = png_ptr->row_buf + row_info->rowbytes;
+
+      switch (row_info->bit_depth)
+      {
+         case 1:
+         {
+            /* in this case, all bytes must be 0 so we don't need
+             * to unpack the pixels except for the rightmost one.
+             */
+            for (; rp > png_ptr->row_buf; rp--)
+            {
+              if (*rp >> padding != 0)
+                 png_ptr->num_palette_max = 1;
+              padding = 0;
+            }
+
+            break;
+         }
+
+         case 2:
+         {
+            for (; rp > png_ptr->row_buf; rp--)
+            {
+              int i = ((*rp >> padding) & 0x03);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              i = (((*rp >> padding) >> 2) & 0x03);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              i = (((*rp >> padding) >> 4) & 0x03);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              i = (((*rp >> padding) >> 6) & 0x03);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              padding = 0;
+            }
+
+            break;
+         }
+
+         case 4:
+         {
+            for (; rp > png_ptr->row_buf; rp--)
+            {
+              int i = ((*rp >> padding) & 0x0f);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              i = (((*rp >> padding) >> 4) & 0x0f);
+
+              if (i > png_ptr->num_palette_max)
+                 png_ptr->num_palette_max = i;
+
+              padding = 0;
+            }
+
+            break;
+         }
+
+         case 8:
+         {
+            for (; rp > png_ptr->row_buf; rp--)
+            {
+               if (*rp > png_ptr->num_palette_max)
+                  png_ptr->num_palette_max = (int) *rp;
+            }
+
+            break;
+         }
+
+         default:
+            break;
+      }
+   }
+}
+#endif /* PNG_CHECK_FOR_INVALID_INDEX_SUPPORTED */
+
+#if defined(PNG_READ_USER_TRANSFORM_SUPPORTED) || \
+    defined(PNG_WRITE_USER_TRANSFORM_SUPPORTED)
+#ifdef PNG_USER_TRANSFORM_PTR_SUPPORTED
+void PNGAPI
+png_set_user_transform_info(png_structrp png_ptr, png_voidp
+   user_transform_ptr, int user_transform_depth, int user_transform_channels)
+{
+   png_debug(1, "in png_set_user_transform_info");
+
+   if (png_ptr == NULL)
+      return;
+
+#ifdef PNG_READ_USER_TRANSFORM_SUPPORTED
+   if ((png_ptr->mode & PNG_IS_READ_STRUCT) != 0 &&
+      (png_ptr->flags & PNG_FLAG_ROW_INIT) != 0)
+   {
+      png_app_error(png_ptr,
+            "info change after png_start_read_image or png_read_update_info");
+      return;
+   }
+#endif
+
+   png_ptr->user_transform_ptr = user_transform_ptr;
+   png_ptr->user_transform_depth = (png_byte)user_transform_depth;
+   png_ptr->user_transform_channels = (png_byte)user_transform_channels;
+}
+#endif
+
+/* This function returns a pointer to the user_transform_ptr associated with
+ * the user transform functions.  The application should free any memory
+ * associated with this pointer before png_write_destroy and png_read_destroy
+ * are called.
+ */
+#ifdef PNG_USER_TRANSFORM_PTR_SUPPORTED
+png_voidp PNGAPI
+png_get_user_transform_ptr(png_const_structrp png_ptr)
+{
+   if (png_ptr == NULL)
+      return (NULL);
+
+   return png_ptr->user_transform_ptr;
+}
+#endif
+
+#ifdef PNG_USER_TRANSFORM_INFO_SUPPORTED
+png_uint_32 PNGAPI
+png_get_current_row_number(png_const_structrp png_ptr)
+{
+   /* See the comments in png.h - this is the sub-image row when reading and
+    * interlaced image.
+    */
+   if (png_ptr != NULL)
+      return png_ptr->row_number;
+
+   return PNG_UINT_32_MAX; /* help the app not to fail silently */
+}
+
+png_byte PNGAPI
+png_get_current_pass_number(png_const_structrp png_ptr)
+{
+   if (png_ptr != NULL)
+      return png_ptr->pass;
+   return 8; /* invalid */
+}
+#endif /* PNG_USER_TRANSFORM_INFO_SUPPORTED */
+#endif /* PNG_READ_USER_TRANSFORM_SUPPORTED ||
+          PNG_WRITE_USER_TRANSFORM_SUPPORTED */
+#endif /* PNG_READ_SUPPORTED || PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngwio.c b/ml/dlib/dlib/external/libpng/pngwio.c
new file mode 100644
index 000000000..e3289dfe4
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngwio.c
@@ -0,0 +1,164 @@
+
+/* pngwio.c - functions for data output
+ *
+ * Last changed in libpng 1.6.0 [February 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ *
+ * This file provides a location for all output.  Users who need
+ * special handling are expected to write functions that have the same
+ * arguments as these and perform similar functions, but that possibly
+ * use different output methods.  Note that you shouldn't change these
+ * functions, but rather write replacement functions and then change
+ * them at run time with png_set_write_fn(...).
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_WRITE_SUPPORTED
+
+/* Write the data to whatever output you are using.  The default routine
+ * writes to a file pointer.  Note that this routine sometimes gets called
+ * with very small lengths, so you should implement some kind of simple
+ * buffering if you are using unbuffered writes.  This should never be asked
+ * to write more than 64K on a 16 bit machine.
+ */
+
+void /* PRIVATE */
+png_write_data(png_structrp png_ptr, png_const_bytep data, png_size_t length)
+{
+   /* NOTE: write_data_fn must not change the buffer! */
+   if (png_ptr->write_data_fn != NULL )
+      (*(png_ptr->write_data_fn))(png_ptr, png_constcast(png_bytep,data),
+         length);
+
+   else
+      png_error(png_ptr, "Call to NULL write function");
+}
+
+#ifdef PNG_STDIO_SUPPORTED
+/* This is the function that does the actual writing of data.  If you are
+ * not writing to a standard C stream, you should create a replacement
+ * write_data function and use it at run time with png_set_write_fn(), rather
+ * than changing the library.
+ */
+void PNGCBAPI
+png_default_write_data(png_structp png_ptr, png_bytep data, png_size_t length)
+{
+   png_size_t check;
+
+   if (png_ptr == NULL)
+      return;
+
+   check = fwrite(data, 1, length, (png_FILE_p)(png_ptr->io_ptr));
+
+   if (check != length)
+      png_error(png_ptr, "Write Error");
+}
+#endif
+
+/* This function is called to output any data pending writing (normally
+ * to disk).  After png_flush is called, there should be no data pending
+ * writing in any buffers.
+ */
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+void /* PRIVATE */
+png_flush(png_structrp png_ptr)
+{
+   if (png_ptr->output_flush_fn != NULL)
+      (*(png_ptr->output_flush_fn))(png_ptr);
+}
+
+#  ifdef PNG_STDIO_SUPPORTED
+void PNGCBAPI
+png_default_flush(png_structp png_ptr)
+{
+   png_FILE_p io_ptr;
+
+   if (png_ptr == NULL)
+      return;
+
+   io_ptr = png_voidcast(png_FILE_p, (png_ptr->io_ptr));
+   fflush(io_ptr);
+}
+#  endif
+#endif
+
+/* This function allows the application to supply new output functions for
+ * libpng if standard C streams aren't being used.
+ *
+ * This function takes as its arguments:
+ * png_ptr       - pointer to a png output data structure
+ * io_ptr        - pointer to user supplied structure containing info about
+ *                 the output functions.  May be NULL.
+ * write_data_fn - pointer to a new output function that takes as its
+ *                 arguments a pointer to a png_struct, a pointer to
+ *                 data to be written, and a 32-bit unsigned int that is
+ *                 the number of bytes to be written.  The new write
+ *                 function should call png_error(png_ptr, "Error msg")
+ *                 to exit and output any fatal error messages.  May be
+ *                 NULL, in which case libpng's default function will
+ *                 be used.
+ * flush_data_fn - pointer to a new flush function that takes as its
+ *                 arguments a pointer to a png_struct.  After a call to
+ *                 the flush function, there should be no data in any buffers
+ *                 or pending transmission.  If the output method doesn't do
+ *                 any buffering of output, a function prototype must still be
+ *                 supplied although it doesn't have to do anything.  If
+ *                 PNG_WRITE_FLUSH_SUPPORTED is not defined at libpng compile
+ *                 time, output_flush_fn will be ignored, although it must be
+ *                 supplied for compatibility.  May be NULL, in which case
+ *                 libpng's default function will be used, if
+ *                 PNG_WRITE_FLUSH_SUPPORTED is defined.  This is not
+ *                 a good idea if io_ptr does not point to a standard
+ *                 *FILE structure.
+ */
+void PNGAPI
+png_set_write_fn(png_structrp png_ptr, png_voidp io_ptr,
+    png_rw_ptr write_data_fn, png_flush_ptr output_flush_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->io_ptr = io_ptr;
+
+#ifdef PNG_STDIO_SUPPORTED
+   if (write_data_fn != NULL)
+      png_ptr->write_data_fn = write_data_fn;
+
+   else
+      png_ptr->write_data_fn = png_default_write_data;
+#else
+   png_ptr->write_data_fn = write_data_fn;
+#endif
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+#  ifdef PNG_STDIO_SUPPORTED
+
+   if (output_flush_fn != NULL)
+      png_ptr->output_flush_fn = output_flush_fn;
+
+   else
+      png_ptr->output_flush_fn = png_default_flush;
+
+#  else
+   png_ptr->output_flush_fn = output_flush_fn;
+#  endif
+#endif /* PNG_WRITE_FLUSH_SUPPORTED */
+
+   /* It is an error to read while writing a png file */
+   if (png_ptr->read_data_fn != NULL)
+   {
+      png_ptr->read_data_fn = NULL;
+
+      png_warning(png_ptr,
+          "Can't set both read_data_fn and write_data_fn in the"
+          " same structure");
+   }
+}
+#endif /* PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngwrite.c b/ml/dlib/dlib/external/libpng/pngwrite.c
new file mode 100644
index 000000000..b71a3d345
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngwrite.c
@@ -0,0 +1,2330 @@
+
+/* pngwrite.c - general routines to write a PNG file
+ *
+ * Last changed in libpng 1.6.2 [April 25, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+#if defined(PNG_SIMPLIFIED_WRITE_SUPPORTED) && defined(PNG_STDIO_SUPPORTED)
+#  include <errno.h>
+#endif
+
+#ifdef PNG_WRITE_SUPPORTED
+
+#ifdef PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED
+/* Write out all the unknown chunks for the current given location */
+static void
+write_unknown_chunks(png_structrp png_ptr, png_const_inforp info_ptr,
+   unsigned int where)
+{
+   if (info_ptr->unknown_chunks_num)
+   {
+      png_const_unknown_chunkp up;
+
+      png_debug(5, "writing extra chunks");
+
+      for (up = info_ptr->unknown_chunks;
+           up < info_ptr->unknown_chunks + info_ptr->unknown_chunks_num;
+           ++up)
+         if (up->location & where)
+      {
+         /* If per-chunk unknown chunk handling is enabled use it, otherwise
+          * just write the chunks the application has set.
+          */
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+         int keep = png_handle_as_unknown(png_ptr, up->name);
+
+         /* NOTE: this code is radically different from the read side in the
+          * matter of handling an ancillary unknown chunk.  In the read side
+          * the default behavior is to discard it, in the code below the default
+          * behavior is to write it.  Critical chunks are, however, only
+          * written if explicitly listed or if the default is set to write all
+          * unknown chunks.
+          *
+          * The default handling is also slightly weird - it is not possible to
+          * stop the writing of all unsafe-to-copy chunks!
+          *
+          * TODO: REVIEW: this would seem to be a bug.
+          */
+         if (keep != PNG_HANDLE_CHUNK_NEVER &&
+             ((up->name[3] & 0x20) /* safe-to-copy overrides everything */ ||
+              keep == PNG_HANDLE_CHUNK_ALWAYS ||
+              (keep == PNG_HANDLE_CHUNK_AS_DEFAULT &&
+               png_ptr->unknown_default == PNG_HANDLE_CHUNK_ALWAYS)))
+#endif
+         {
+            /* TODO: review, what is wrong with a zero length unknown chunk? */
+            if (up->size == 0)
+               png_warning(png_ptr, "Writing zero-length unknown chunk");
+
+            png_write_chunk(png_ptr, up->name, up->data, up->size);
+         }
+      }
+   }
+}
+#endif /* PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED */
+
+/* Writes all the PNG information.  This is the suggested way to use the
+ * library.  If you have a new chunk to add, make a function to write it,
+ * and put it in the correct location here.  If you want the chunk written
+ * after the image data, put it in png_write_end().  I strongly encourage
+ * you to supply a PNG_INFO_ flag, and check info_ptr->valid before writing
+ * the chunk, as that will keep the code from breaking if you want to just
+ * write a plain PNG file.  If you have long comments, I suggest writing
+ * them in png_write_end(), and compressing them.
+ */
+void PNGAPI
+png_write_info_before_PLTE(png_structrp png_ptr, png_const_inforp info_ptr)
+{
+   png_debug(1, "in png_write_info_before_PLTE");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   if (!(png_ptr->mode & PNG_WROTE_INFO_BEFORE_PLTE))
+   {
+   /* Write PNG signature */
+   png_write_sig(png_ptr);
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   if ((png_ptr->mode&PNG_HAVE_PNG_SIGNATURE) && \
+       (png_ptr->mng_features_permitted))
+   {
+      png_warning(png_ptr, "MNG features are not allowed in a PNG datastream");
+      png_ptr->mng_features_permitted = 0;
+   }
+#endif
+
+   /* Write IHDR information. */
+   png_write_IHDR(png_ptr, info_ptr->width, info_ptr->height,
+       info_ptr->bit_depth, info_ptr->color_type, info_ptr->compression_type,
+       info_ptr->filter_type,
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+       info_ptr->interlace_type
+#else
+       0
+#endif
+      );
+
+   /* The rest of these check to see if the valid field has the appropriate
+    * flag set, and if it does, writes the chunk.
+    *
+    * 1.6.0: COLORSPACE support controls the writing of these chunks too, and
+    * the chunks will be written if the WRITE routine is there and information
+    * is available in the COLORSPACE.  (See png_colorspace_sync_info in png.c
+    * for where the valid flags get set.)
+    *
+    * Under certain circumstances the colorspace can be invalidated without
+    * syncing the info_struct 'valid' flags; this happens if libpng detects and
+    * error and calls png_error while the color space is being set, yet the
+    * application continues writing the PNG.  So check the 'invalid' flag here
+    * too.
+    */
+#ifdef PNG_GAMMA_SUPPORTED
+#  ifdef PNG_WRITE_gAMA_SUPPORTED
+      if (!(info_ptr->colorspace.flags & PNG_COLORSPACE_INVALID) &&
+         (info_ptr->colorspace.flags & PNG_COLORSPACE_FROM_gAMA) &&
+         (info_ptr->valid & PNG_INFO_gAMA))
+         png_write_gAMA_fixed(png_ptr, info_ptr->colorspace.gamma);
+#  endif
+#endif
+
+#ifdef PNG_COLORSPACE_SUPPORTED
+   /* Write only one of sRGB or an ICC profile.  If a profile was supplied
+    * and it matches one of the known sRGB ones issue a warning.
+    */
+#  ifdef PNG_WRITE_iCCP_SUPPORTED
+      if (!(info_ptr->colorspace.flags & PNG_COLORSPACE_INVALID) &&
+         (info_ptr->valid & PNG_INFO_iCCP))
+      {
+#        ifdef PNG_WRITE_sRGB_SUPPORTED
+            if (info_ptr->valid & PNG_INFO_sRGB)
+               png_app_warning(png_ptr,
+                  "profile matches sRGB but writing iCCP instead");
+#        endif
+
+         png_write_iCCP(png_ptr, info_ptr->iccp_name,
+            info_ptr->iccp_profile);
+      }
+#     ifdef PNG_WRITE_sRGB_SUPPORTED
+         else
+#     endif
+#  endif
+
+#  ifdef PNG_WRITE_sRGB_SUPPORTED
+      if (!(info_ptr->colorspace.flags & PNG_COLORSPACE_INVALID) &&
+         (info_ptr->valid & PNG_INFO_sRGB))
+         png_write_sRGB(png_ptr, info_ptr->colorspace.rendering_intent);
+#  endif /* WRITE_sRGB */
+#endif /* COLORSPACE */
+
+#ifdef PNG_WRITE_sBIT_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_sBIT)
+      png_write_sBIT(png_ptr, &(info_ptr->sig_bit), info_ptr->color_type);
+#endif
+
+#ifdef PNG_COLORSPACE_SUPPORTED
+#  ifdef PNG_WRITE_cHRM_SUPPORTED
+      if (!(info_ptr->colorspace.flags & PNG_COLORSPACE_INVALID) &&
+         (info_ptr->colorspace.flags & PNG_COLORSPACE_FROM_cHRM) &&
+         (info_ptr->valid & PNG_INFO_cHRM))
+         png_write_cHRM_fixed(png_ptr, &info_ptr->colorspace.end_points_xy);
+#  endif
+#endif
+
+#ifdef PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED
+      write_unknown_chunks(png_ptr, info_ptr, PNG_HAVE_IHDR);
+#endif
+
+      png_ptr->mode |= PNG_WROTE_INFO_BEFORE_PLTE;
+   }
+}
+
+void PNGAPI
+png_write_info(png_structrp png_ptr, png_const_inforp info_ptr)
+{
+#if defined(PNG_WRITE_TEXT_SUPPORTED) || defined(PNG_WRITE_sPLT_SUPPORTED)
+   int i;
+#endif
+
+   png_debug(1, "in png_write_info");
+
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   png_write_info_before_PLTE(png_ptr, info_ptr);
+
+   if (info_ptr->valid & PNG_INFO_PLTE)
+      png_write_PLTE(png_ptr, info_ptr->palette,
+          (png_uint_32)info_ptr->num_palette);
+
+   else if (info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      png_error(png_ptr, "Valid palette required for paletted images");
+
+#ifdef PNG_WRITE_tRNS_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_tRNS)
+   {
+#ifdef PNG_WRITE_INVERT_ALPHA_SUPPORTED
+      /* Invert the alpha channel (in tRNS) */
+      if ((png_ptr->transformations & PNG_INVERT_ALPHA) &&
+          info_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         int j;
+         for (j = 0; j<(int)info_ptr->num_trans; j++)
+            info_ptr->trans_alpha[j] =
+               (png_byte)(255 - info_ptr->trans_alpha[j]);
+      }
+#endif
+      png_write_tRNS(png_ptr, info_ptr->trans_alpha, &(info_ptr->trans_color),
+          info_ptr->num_trans, info_ptr->color_type);
+   }
+#endif
+#ifdef PNG_WRITE_bKGD_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_bKGD)
+      png_write_bKGD(png_ptr, &(info_ptr->background), info_ptr->color_type);
+#endif
+
+#ifdef PNG_WRITE_hIST_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_hIST)
+      png_write_hIST(png_ptr, info_ptr->hist, info_ptr->num_palette);
+#endif
+
+#ifdef PNG_WRITE_oFFs_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_oFFs)
+      png_write_oFFs(png_ptr, info_ptr->x_offset, info_ptr->y_offset,
+          info_ptr->offset_unit_type);
+#endif
+
+#ifdef PNG_WRITE_pCAL_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_pCAL)
+      png_write_pCAL(png_ptr, info_ptr->pcal_purpose, info_ptr->pcal_X0,
+          info_ptr->pcal_X1, info_ptr->pcal_type, info_ptr->pcal_nparams,
+          info_ptr->pcal_units, info_ptr->pcal_params);
+#endif
+
+#ifdef PNG_WRITE_sCAL_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_sCAL)
+      png_write_sCAL_s(png_ptr, (int)info_ptr->scal_unit,
+          info_ptr->scal_s_width, info_ptr->scal_s_height);
+#endif /* sCAL */
+
+#ifdef PNG_WRITE_pHYs_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_pHYs)
+      png_write_pHYs(png_ptr, info_ptr->x_pixels_per_unit,
+          info_ptr->y_pixels_per_unit, info_ptr->phys_unit_type);
+#endif /* pHYs */
+
+#ifdef PNG_WRITE_tIME_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_tIME)
+   {
+      png_write_tIME(png_ptr, &(info_ptr->mod_time));
+      png_ptr->mode |= PNG_WROTE_tIME;
+   }
+#endif /* tIME */
+
+#ifdef PNG_WRITE_sPLT_SUPPORTED
+   if (info_ptr->valid & PNG_INFO_sPLT)
+      for (i = 0; i < (int)info_ptr->splt_palettes_num; i++)
+         png_write_sPLT(png_ptr, info_ptr->splt_palettes + i);
+#endif /* sPLT */
+
+#ifdef PNG_WRITE_TEXT_SUPPORTED
+   /* Check to see if we need to write text chunks */
+   for (i = 0; i < info_ptr->num_text; i++)
+   {
+      png_debug2(2, "Writing header text chunk %d, type %d", i,
+          info_ptr->text[i].compression);
+      /* An internationalized chunk? */
+      if (info_ptr->text[i].compression > 0)
+      {
+#ifdef PNG_WRITE_iTXt_SUPPORTED
+         /* Write international chunk */
+         png_write_iTXt(png_ptr,
+             info_ptr->text[i].compression,
+             info_ptr->text[i].key,
+             info_ptr->text[i].lang,
+             info_ptr->text[i].lang_key,
+             info_ptr->text[i].text);
+#else
+          png_warning(png_ptr, "Unable to write international text");
+#endif
+          /* Mark this chunk as written */
+          info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_NONE_WR;
+      }
+
+      /* If we want a compressed text chunk */
+      else if (info_ptr->text[i].compression == PNG_TEXT_COMPRESSION_zTXt)
+      {
+#ifdef PNG_WRITE_zTXt_SUPPORTED
+         /* Write compressed chunk */
+         png_write_zTXt(png_ptr, info_ptr->text[i].key,
+             info_ptr->text[i].text, 0,
+             info_ptr->text[i].compression);
+#else
+         png_warning(png_ptr, "Unable to write compressed text");
+#endif
+         /* Mark this chunk as written */
+         info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_zTXt_WR;
+      }
+
+      else if (info_ptr->text[i].compression == PNG_TEXT_COMPRESSION_NONE)
+      {
+#ifdef PNG_WRITE_tEXt_SUPPORTED
+         /* Write uncompressed chunk */
+         png_write_tEXt(png_ptr, info_ptr->text[i].key,
+             info_ptr->text[i].text,
+             0);
+         /* Mark this chunk as written */
+         info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_NONE_WR;
+#else
+         /* Can't get here */
+         png_warning(png_ptr, "Unable to write uncompressed text");
+#endif
+      }
+   }
+#endif /* tEXt */
+
+#ifdef PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED
+   write_unknown_chunks(png_ptr, info_ptr, PNG_HAVE_PLTE);
+#endif
+}
+
+/* Writes the end of the PNG file.  If you don't want to write comments or
+ * time information, you can pass NULL for info.  If you already wrote these
+ * in png_write_info(), do not write them again here.  If you have long
+ * comments, I suggest writing them here, and compressing them.
+ */
+void PNGAPI
+png_write_end(png_structrp png_ptr, png_inforp info_ptr)
+{
+   png_debug(1, "in png_write_end");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (!(png_ptr->mode & PNG_HAVE_IDAT))
+      png_error(png_ptr, "No IDATs written into file");
+
+#ifdef PNG_WRITE_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   if (png_ptr->num_palette_max > png_ptr->num_palette)
+      png_benign_error(png_ptr, "Wrote palette index exceeding num_palette");
+#endif
+
+   /* See if user wants us to write information chunks */
+   if (info_ptr != NULL)
+   {
+#ifdef PNG_WRITE_TEXT_SUPPORTED
+      int i; /* local index variable */
+#endif
+#ifdef PNG_WRITE_tIME_SUPPORTED
+      /* Check to see if user has supplied a time chunk */
+      if ((info_ptr->valid & PNG_INFO_tIME) &&
+          !(png_ptr->mode & PNG_WROTE_tIME))
+         png_write_tIME(png_ptr, &(info_ptr->mod_time));
+
+#endif
+#ifdef PNG_WRITE_TEXT_SUPPORTED
+      /* Loop through comment chunks */
+      for (i = 0; i < info_ptr->num_text; i++)
+      {
+         png_debug2(2, "Writing trailer text chunk %d, type %d", i,
+            info_ptr->text[i].compression);
+         /* An internationalized chunk? */
+         if (info_ptr->text[i].compression > 0)
+         {
+#ifdef PNG_WRITE_iTXt_SUPPORTED
+            /* Write international chunk */
+            png_write_iTXt(png_ptr,
+                info_ptr->text[i].compression,
+                info_ptr->text[i].key,
+                info_ptr->text[i].lang,
+                info_ptr->text[i].lang_key,
+                info_ptr->text[i].text);
+#else
+            png_warning(png_ptr, "Unable to write international text");
+#endif
+            /* Mark this chunk as written */
+            info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_NONE_WR;
+         }
+
+         else if (info_ptr->text[i].compression >= PNG_TEXT_COMPRESSION_zTXt)
+         {
+#ifdef PNG_WRITE_zTXt_SUPPORTED
+            /* Write compressed chunk */
+            png_write_zTXt(png_ptr, info_ptr->text[i].key,
+                info_ptr->text[i].text, 0,
+                info_ptr->text[i].compression);
+#else
+            png_warning(png_ptr, "Unable to write compressed text");
+#endif
+            /* Mark this chunk as written */
+            info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_zTXt_WR;
+         }
+
+         else if (info_ptr->text[i].compression == PNG_TEXT_COMPRESSION_NONE)
+         {
+#ifdef PNG_WRITE_tEXt_SUPPORTED
+            /* Write uncompressed chunk */
+            png_write_tEXt(png_ptr, info_ptr->text[i].key,
+                info_ptr->text[i].text, 0);
+#else
+            png_warning(png_ptr, "Unable to write uncompressed text");
+#endif
+
+            /* Mark this chunk as written */
+            info_ptr->text[i].compression = PNG_TEXT_COMPRESSION_NONE_WR;
+         }
+      }
+#endif
+#ifdef PNG_WRITE_UNKNOWN_CHUNKS_SUPPORTED
+      write_unknown_chunks(png_ptr, info_ptr, PNG_AFTER_IDAT);
+#endif
+   }
+
+   png_ptr->mode |= PNG_AFTER_IDAT;
+
+   /* Write end of PNG file */
+   png_write_IEND(png_ptr);
+   /* This flush, added in libpng-1.0.8, removed from libpng-1.0.9beta03,
+    * and restored again in libpng-1.2.30, may cause some applications that
+    * do not set png_ptr->output_flush_fn to crash.  If your application
+    * experiences a problem, please try building libpng with
+    * PNG_WRITE_FLUSH_AFTER_IEND_SUPPORTED defined, and report the event to
+    * png-mng-implement at lists.sf.net .
+    */
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+#  ifdef PNG_WRITE_FLUSH_AFTER_IEND_SUPPORTED
+   png_flush(png_ptr);
+#  endif
+#endif
+}
+
+#ifdef PNG_CONVERT_tIME_SUPPORTED
+void PNGAPI
+png_convert_from_struct_tm(png_timep ptime, PNG_CONST struct tm * ttime)
+{
+   png_debug(1, "in png_convert_from_struct_tm");
+
+   ptime->year = (png_uint_16)(1900 + ttime->tm_year);
+   ptime->month = (png_byte)(ttime->tm_mon + 1);
+   ptime->day = (png_byte)ttime->tm_mday;
+   ptime->hour = (png_byte)ttime->tm_hour;
+   ptime->minute = (png_byte)ttime->tm_min;
+   ptime->second = (png_byte)ttime->tm_sec;
+}
+
+void PNGAPI
+png_convert_from_time_t(png_timep ptime, time_t ttime)
+{
+   struct tm *tbuf;
+
+   png_debug(1, "in png_convert_from_time_t");
+
+   tbuf = gmtime(&ttime);
+   png_convert_from_struct_tm(ptime, tbuf);
+}
+#endif
+
+/* Initialize png_ptr structure, and allocate any memory needed */
+PNG_FUNCTION(png_structp,PNGAPI
+png_create_write_struct,(png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn),PNG_ALLOCATED)
+{
+#ifndef PNG_USER_MEM_SUPPORTED
+   png_structrp png_ptr = png_create_png_struct(user_png_ver, error_ptr,
+      error_fn, warn_fn, NULL, NULL, NULL);
+#else
+   return png_create_write_struct_2(user_png_ver, error_ptr, error_fn,
+       warn_fn, NULL, NULL, NULL);
+}
+
+/* Alternate initialize png_ptr structure, and allocate any memory needed */
+PNG_FUNCTION(png_structp,PNGAPI
+png_create_write_struct_2,(png_const_charp user_png_ver, png_voidp error_ptr,
+    png_error_ptr error_fn, png_error_ptr warn_fn, png_voidp mem_ptr,
+    png_malloc_ptr malloc_fn, png_free_ptr free_fn),PNG_ALLOCATED)
+{
+   png_structrp png_ptr = png_create_png_struct(user_png_ver, error_ptr,
+      error_fn, warn_fn, mem_ptr, malloc_fn, free_fn);
+#endif /* PNG_USER_MEM_SUPPORTED */
+   if (png_ptr != NULL)
+   {
+      /* Set the zlib control values to defaults; they can be overridden by the
+       * application after the struct has been created.
+       */
+      png_ptr->zbuffer_size = PNG_ZBUF_SIZE;
+
+      /* The 'zlib_strategy' setting is irrelevant because png_default_claim in
+       * pngwutil.c defaults it according to whether or not filters will be
+       * used, and ignores this setting.
+       */
+      png_ptr->zlib_strategy = PNG_Z_DEFAULT_STRATEGY;
+      png_ptr->zlib_level = PNG_Z_DEFAULT_COMPRESSION;
+      png_ptr->zlib_mem_level = 8;
+      png_ptr->zlib_window_bits = 15;
+      png_ptr->zlib_method = 8;
+
+#ifdef PNG_WRITE_COMPRESSED_TEXT_SUPPORTED
+      png_ptr->zlib_text_strategy = PNG_TEXT_Z_DEFAULT_STRATEGY;
+      png_ptr->zlib_text_level = PNG_TEXT_Z_DEFAULT_COMPRESSION;
+      png_ptr->zlib_text_mem_level = 8;
+      png_ptr->zlib_text_window_bits = 15;
+      png_ptr->zlib_text_method = 8;
+#endif /* PNG_WRITE_COMPRESSED_TEXT_SUPPORTED */
+
+      /* This is a highly dubious configuration option; by default it is off,
+       * but it may be appropriate for private builds that are testing
+       * extensions not conformant to the current specification, or of
+       * applications that must not fail to write at all costs!
+       */
+#ifdef PNG_BENIGN_WRITE_ERRORS_SUPPORTED
+      png_ptr->flags |= PNG_FLAG_BENIGN_ERRORS_WARN;
+      /* In stable builds only warn if an application error can be completely
+       * handled.
+       */
+#endif
+
+      /* App warnings are warnings in release (or release candidate) builds but
+       * are errors during development.
+       */
+#if PNG_LIBPNG_BUILD_BASE_TYPE >= PNG_LIBPNG_BUILD_RC
+      png_ptr->flags |= PNG_FLAG_APP_WARNINGS_WARN;
+#endif
+
+      /* TODO: delay this, it can be done in png_init_io() (if the app doesn't
+       * do it itself) avoiding setting the default function if it is not
+       * required.
+       */
+      png_set_write_fn(png_ptr, NULL, NULL, NULL);
+   }
+
+   return png_ptr;
+}
+
+
+/* Write a few rows of image data.  If the image is interlaced,
+ * either you will have to write the 7 sub images, or, if you
+ * have called png_set_interlace_handling(), you will have to
+ * "write" the image seven times.
+ */
+void PNGAPI
+png_write_rows(png_structrp png_ptr, png_bytepp row,
+    png_uint_32 num_rows)
+{
+   png_uint_32 i; /* row counter */
+   png_bytepp rp; /* row pointer */
+
+   png_debug(1, "in png_write_rows");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* Loop through the rows */
+   for (i = 0, rp = row; i < num_rows; i++, rp++)
+   {
+      png_write_row(png_ptr, *rp);
+   }
+}
+
+/* Write the image.  You only need to call this function once, even
+ * if you are writing an interlaced image.
+ */
+void PNGAPI
+png_write_image(png_structrp png_ptr, png_bytepp image)
+{
+   png_uint_32 i; /* row index */
+   int pass, num_pass; /* pass variables */
+   png_bytepp rp; /* points to current row */
+
+   if (png_ptr == NULL)
+      return;
+
+   png_debug(1, "in png_write_image");
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* Initialize interlace handling.  If image is not interlaced,
+    * this will set pass to 1
+    */
+   num_pass = png_set_interlace_handling(png_ptr);
+#else
+   num_pass = 1;
+#endif
+   /* Loop through passes */
+   for (pass = 0; pass < num_pass; pass++)
+   {
+      /* Loop through image */
+      for (i = 0, rp = image; i < png_ptr->height; i++, rp++)
+      {
+         png_write_row(png_ptr, *rp);
+      }
+   }
+}
+
+/* Called by user to write a row of image data */
+void PNGAPI
+png_write_row(png_structrp png_ptr, png_const_bytep row)
+{
+   /* 1.5.6: moved from png_struct to be a local structure: */
+   png_row_info row_info;
+
+   if (png_ptr == NULL)
+      return;
+
+   png_debug2(1, "in png_write_row (row %u, pass %d)",
+      png_ptr->row_number, png_ptr->pass);
+
+   /* Initialize transformations and other stuff if first time */
+   if (png_ptr->row_number == 0 && png_ptr->pass == 0)
+   {
+      /* Make sure we wrote the header info */
+      if (!(png_ptr->mode & PNG_WROTE_INFO_BEFORE_PLTE))
+         png_error(png_ptr,
+             "png_write_info was never called before png_write_row");
+
+      /* Check for transforms that have been set but were defined out */
+#if !defined(PNG_WRITE_INVERT_SUPPORTED) && defined(PNG_READ_INVERT_SUPPORTED)
+      if (png_ptr->transformations & PNG_INVERT_MONO)
+         png_warning(png_ptr, "PNG_WRITE_INVERT_SUPPORTED is not defined");
+#endif
+
+#if !defined(PNG_WRITE_FILLER_SUPPORTED) && defined(PNG_READ_FILLER_SUPPORTED)
+      if (png_ptr->transformations & PNG_FILLER)
+         png_warning(png_ptr, "PNG_WRITE_FILLER_SUPPORTED is not defined");
+#endif
+#if !defined(PNG_WRITE_PACKSWAP_SUPPORTED) && \
+    defined(PNG_READ_PACKSWAP_SUPPORTED)
+      if (png_ptr->transformations & PNG_PACKSWAP)
+         png_warning(png_ptr,
+             "PNG_WRITE_PACKSWAP_SUPPORTED is not defined");
+#endif
+
+#if !defined(PNG_WRITE_PACK_SUPPORTED) && defined(PNG_READ_PACK_SUPPORTED)
+      if (png_ptr->transformations & PNG_PACK)
+         png_warning(png_ptr, "PNG_WRITE_PACK_SUPPORTED is not defined");
+#endif
+
+#if !defined(PNG_WRITE_SHIFT_SUPPORTED) && defined(PNG_READ_SHIFT_SUPPORTED)
+      if (png_ptr->transformations & PNG_SHIFT)
+         png_warning(png_ptr, "PNG_WRITE_SHIFT_SUPPORTED is not defined");
+#endif
+
+#if !defined(PNG_WRITE_BGR_SUPPORTED) && defined(PNG_READ_BGR_SUPPORTED)
+      if (png_ptr->transformations & PNG_BGR)
+         png_warning(png_ptr, "PNG_WRITE_BGR_SUPPORTED is not defined");
+#endif
+
+#if !defined(PNG_WRITE_SWAP_SUPPORTED) && defined(PNG_READ_SWAP_SUPPORTED)
+      if (png_ptr->transformations & PNG_SWAP_BYTES)
+         png_warning(png_ptr, "PNG_WRITE_SWAP_SUPPORTED is not defined");
+#endif
+
+      png_write_start_row(png_ptr);
+   }
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* If interlaced and not interested in row, return */
+   if (png_ptr->interlaced && (png_ptr->transformations & PNG_INTERLACE))
+   {
+      switch (png_ptr->pass)
+      {
+         case 0:
+            if (png_ptr->row_number & 0x07)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 1:
+            if ((png_ptr->row_number & 0x07) || png_ptr->width < 5)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 2:
+            if ((png_ptr->row_number & 0x07) != 4)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 3:
+            if ((png_ptr->row_number & 0x03) || png_ptr->width < 3)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 4:
+            if ((png_ptr->row_number & 0x03) != 2)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 5:
+            if ((png_ptr->row_number & 0x01) || png_ptr->width < 2)
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         case 6:
+            if (!(png_ptr->row_number & 0x01))
+            {
+               png_write_finish_row(png_ptr);
+               return;
+            }
+            break;
+
+         default: /* error: ignore it */
+            break;
+      }
+   }
+#endif
+
+   /* Set up row info for transformations */
+   row_info.color_type = png_ptr->color_type;
+   row_info.width = png_ptr->usr_width;
+   row_info.channels = png_ptr->usr_channels;
+   row_info.bit_depth = png_ptr->usr_bit_depth;
+   row_info.pixel_depth = (png_byte)(row_info.bit_depth * row_info.channels);
+   row_info.rowbytes = PNG_ROWBYTES(row_info.pixel_depth, row_info.width);
+
+   png_debug1(3, "row_info->color_type = %d", row_info.color_type);
+   png_debug1(3, "row_info->width = %u", row_info.width);
+   png_debug1(3, "row_info->channels = %d", row_info.channels);
+   png_debug1(3, "row_info->bit_depth = %d", row_info.bit_depth);
+   png_debug1(3, "row_info->pixel_depth = %d", row_info.pixel_depth);
+   png_debug1(3, "row_info->rowbytes = %lu", (unsigned long)row_info.rowbytes);
+
+   /* Copy user's row into buffer, leaving room for filter byte. */
+   memcpy(png_ptr->row_buf + 1, row, row_info.rowbytes);
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* Handle interlacing */
+   if (png_ptr->interlaced && png_ptr->pass < 6 &&
+       (png_ptr->transformations & PNG_INTERLACE))
+   {
+      png_do_write_interlace(&row_info, png_ptr->row_buf + 1, png_ptr->pass);
+      /* This should always get caught above, but still ... */
+      if (!(row_info.width))
+      {
+         png_write_finish_row(png_ptr);
+         return;
+      }
+   }
+#endif
+
+#ifdef PNG_WRITE_TRANSFORMS_SUPPORTED
+   /* Handle other transformations */
+   if (png_ptr->transformations)
+      png_do_write_transformations(png_ptr, &row_info);
+#endif
+
+   /* At this point the row_info pixel depth must match the 'transformed' depth,
+    * which is also the output depth.
+    */
+   if (row_info.pixel_depth != png_ptr->pixel_depth ||
+      row_info.pixel_depth != png_ptr->transformed_pixel_depth)
+      png_error(png_ptr, "internal write transform logic error");
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   /* Write filter_method 64 (intrapixel differencing) only if
+    * 1. Libpng was compiled with PNG_MNG_FEATURES_SUPPORTED and
+    * 2. Libpng did not write a PNG signature (this filter_method is only
+    *    used in PNG datastreams that are embedded in MNG datastreams) and
+    * 3. The application called png_permit_mng_features with a mask that
+    *    included PNG_FLAG_MNG_FILTER_64 and
+    * 4. The filter_method is 64 and
+    * 5. The color_type is RGB or RGBA
+    */
+   if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) &&
+       (png_ptr->filter_type == PNG_INTRAPIXEL_DIFFERENCING))
+   {
+      /* Intrapixel differencing */
+      png_do_write_intrapixel(&row_info, png_ptr->row_buf + 1);
+   }
+#endif
+
+/* Added at libpng-1.5.10 */
+#ifdef PNG_WRITE_CHECK_FOR_INVALID_INDEX_SUPPORTED
+   /* Check for out-of-range palette index */
+   if (row_info.color_type == PNG_COLOR_TYPE_PALETTE &&
+       png_ptr->num_palette_max >= 0)
+      png_do_check_palette_indexes(png_ptr, &row_info);
+#endif
+
+   /* Find a filter if necessary, filter the row and write it out. */
+   png_write_find_filter(png_ptr, &row_info);
+
+   if (png_ptr->write_row_fn != NULL)
+      (*(png_ptr->write_row_fn))(png_ptr, png_ptr->row_number, png_ptr->pass);
+}
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+/* Set the automatic flush interval or 0 to turn flushing off */
+void PNGAPI
+png_set_flush(png_structrp png_ptr, int nrows)
+{
+   png_debug(1, "in png_set_flush");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->flush_dist = (nrows < 0 ? 0 : nrows);
+}
+
+/* Flush the current output buffers now */
+void PNGAPI
+png_write_flush(png_structrp png_ptr)
+{
+   png_debug(1, "in png_write_flush");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* We have already written out all of the data */
+   if (png_ptr->row_number >= png_ptr->num_rows)
+      return;
+
+   png_compress_IDAT(png_ptr, NULL, 0, Z_SYNC_FLUSH);
+   png_ptr->flush_rows = 0;
+   png_flush(png_ptr);
+}
+#endif /* PNG_WRITE_FLUSH_SUPPORTED */
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+static void png_reset_filter_heuristics(png_structrp png_ptr);/* forward decl */
+#endif
+
+/* Free any memory used in png_ptr struct without freeing the struct itself. */
+static void
+png_write_destroy(png_structrp png_ptr)
+{
+   png_debug(1, "in png_write_destroy");
+
+   /* Free any memory zlib uses */
+   if (png_ptr->flags & PNG_FLAG_ZSTREAM_INITIALIZED)
+      deflateEnd(&png_ptr->zstream);
+
+   /* Free our memory.  png_free checks NULL for us. */
+   png_free_buffer_list(png_ptr, &png_ptr->zbuffer_list);
+   png_free(png_ptr, png_ptr->row_buf);
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+   png_free(png_ptr, png_ptr->prev_row);
+   png_free(png_ptr, png_ptr->sub_row);
+   png_free(png_ptr, png_ptr->up_row);
+   png_free(png_ptr, png_ptr->avg_row);
+   png_free(png_ptr, png_ptr->paeth_row);
+#endif
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+   /* Use this to save a little code space, it doesn't free the filter_costs */
+   png_reset_filter_heuristics(png_ptr);
+   png_free(png_ptr, png_ptr->filter_costs);
+   png_free(png_ptr, png_ptr->inv_filter_costs);
+#endif
+
+#ifdef PNG_SET_UNKNOWN_CHUNKS_SUPPORTED
+   png_free(png_ptr, png_ptr->chunk_list);
+#endif
+
+   /* The error handling and memory handling information is left intact at this
+    * point: the jmp_buf may still have to be freed.  See png_destroy_png_struct
+    * for how this happens.
+    */
+}
+
+/* Free all memory used by the write.
+ * In libpng 1.6.0 this API changed quietly to no longer accept a NULL value for
+ * *png_ptr_ptr.  Prior to 1.6.0 it would accept such a value and it would free
+ * the passed in info_structs but it would quietly fail to free any of the data
+ * inside them.  In 1.6.0 it quietly does nothing (it has to be quiet because it
+ * has no png_ptr.)
+ */
+void PNGAPI
+png_destroy_write_struct(png_structpp png_ptr_ptr, png_infopp info_ptr_ptr)
+{
+   png_debug(1, "in png_destroy_write_struct");
+
+   if (png_ptr_ptr != NULL)
+   {
+      png_structrp png_ptr = *png_ptr_ptr;
+
+      if (png_ptr != NULL) /* added in libpng 1.6.0 */
+      {
+         png_destroy_info_struct(png_ptr, info_ptr_ptr);
+
+         *png_ptr_ptr = NULL;
+         png_write_destroy(png_ptr);
+         png_destroy_png_struct(png_ptr);
+      }
+   }
+}
+
+/* Allow the application to select one or more row filters to use. */
+void PNGAPI
+png_set_filter(png_structrp png_ptr, int method, int filters)
+{
+   png_debug(1, "in png_set_filter");
+
+   if (png_ptr == NULL)
+      return;
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   if ((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) &&
+       (method == PNG_INTRAPIXEL_DIFFERENCING))
+      method = PNG_FILTER_TYPE_BASE;
+
+#endif
+   if (method == PNG_FILTER_TYPE_BASE)
+   {
+      switch (filters & (PNG_ALL_FILTERS | 0x07))
+      {
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+         case 5:
+         case 6:
+         case 7: png_app_error(png_ptr, "Unknown row filter for method 0");
+            /* FALL THROUGH */
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+         case PNG_FILTER_VALUE_NONE:
+            png_ptr->do_filter = PNG_FILTER_NONE; break;
+
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+         case PNG_FILTER_VALUE_SUB:
+            png_ptr->do_filter = PNG_FILTER_SUB; break;
+
+         case PNG_FILTER_VALUE_UP:
+            png_ptr->do_filter = PNG_FILTER_UP; break;
+
+         case PNG_FILTER_VALUE_AVG:
+            png_ptr->do_filter = PNG_FILTER_AVG; break;
+
+         case PNG_FILTER_VALUE_PAETH:
+            png_ptr->do_filter = PNG_FILTER_PAETH; break;
+
+         default:
+            png_ptr->do_filter = (png_byte)filters; break;
+#else
+         default:
+            png_app_error(png_ptr, "Unknown row filter for method 0");
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+      }
+
+      /* If we have allocated the row_buf, this means we have already started
+       * with the image and we should have allocated all of the filter buffers
+       * that have been selected.  If prev_row isn't already allocated, then
+       * it is too late to start using the filters that need it, since we
+       * will be missing the data in the previous row.  If an application
+       * wants to start and stop using particular filters during compression,
+       * it should start out with all of the filters, and then add and
+       * remove them after the start of compression.
+       */
+      if (png_ptr->row_buf != NULL)
+      {
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+         if ((png_ptr->do_filter & PNG_FILTER_SUB) && png_ptr->sub_row == NULL)
+         {
+            png_ptr->sub_row = (png_bytep)png_malloc(png_ptr,
+                (png_ptr->rowbytes + 1));
+            png_ptr->sub_row[0] = PNG_FILTER_VALUE_SUB;
+         }
+
+         if ((png_ptr->do_filter & PNG_FILTER_UP) && png_ptr->up_row == NULL)
+         {
+            if (png_ptr->prev_row == NULL)
+            {
+               png_warning(png_ptr, "Can't add Up filter after starting");
+               png_ptr->do_filter = (png_byte)(png_ptr->do_filter &
+                   ~PNG_FILTER_UP);
+            }
+
+            else
+            {
+               png_ptr->up_row = (png_bytep)png_malloc(png_ptr,
+                   (png_ptr->rowbytes + 1));
+               png_ptr->up_row[0] = PNG_FILTER_VALUE_UP;
+            }
+         }
+
+         if ((png_ptr->do_filter & PNG_FILTER_AVG) && png_ptr->avg_row == NULL)
+         {
+            if (png_ptr->prev_row == NULL)
+            {
+               png_warning(png_ptr, "Can't add Average filter after starting");
+               png_ptr->do_filter = (png_byte)(png_ptr->do_filter &
+                   ~PNG_FILTER_AVG);
+            }
+
+            else
+            {
+               png_ptr->avg_row = (png_bytep)png_malloc(png_ptr,
+                   (png_ptr->rowbytes + 1));
+               png_ptr->avg_row[0] = PNG_FILTER_VALUE_AVG;
+            }
+         }
+
+         if ((png_ptr->do_filter & PNG_FILTER_PAETH) &&
+             png_ptr->paeth_row == NULL)
+         {
+            if (png_ptr->prev_row == NULL)
+            {
+               png_warning(png_ptr, "Can't add Paeth filter after starting");
+               png_ptr->do_filter &= (png_byte)(~PNG_FILTER_PAETH);
+            }
+
+            else
+            {
+               png_ptr->paeth_row = (png_bytep)png_malloc(png_ptr,
+                   (png_ptr->rowbytes + 1));
+               png_ptr->paeth_row[0] = PNG_FILTER_VALUE_PAETH;
+            }
+         }
+
+         if (png_ptr->do_filter == PNG_NO_FILTERS)
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+            png_ptr->do_filter = PNG_FILTER_NONE;
+      }
+   }
+   else
+      png_error(png_ptr, "Unknown custom filter method");
+}
+
+/* This allows us to influence the way in which libpng chooses the "best"
+ * filter for the current scanline.  While the "minimum-sum-of-absolute-
+ * differences metric is relatively fast and effective, there is some
+ * question as to whether it can be improved upon by trying to keep the
+ * filtered data going to zlib more consistent, hopefully resulting in
+ * better compression.
+ */
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED      /* GRR 970116 */
+/* Convenience reset API. */
+static void
+png_reset_filter_heuristics(png_structrp png_ptr)
+{
+   /* Clear out any old values in the 'weights' - this must be done because if
+    * the app calls set_filter_heuristics multiple times with different
+    * 'num_weights' values we would otherwise potentially have wrong sized
+    * arrays.
+    */
+   png_ptr->num_prev_filters = 0;
+   png_ptr->heuristic_method = PNG_FILTER_HEURISTIC_UNWEIGHTED;
+   if (png_ptr->prev_filters != NULL)
+   {
+      png_bytep old = png_ptr->prev_filters;
+      png_ptr->prev_filters = NULL;
+      png_free(png_ptr, old);
+   }
+   if (png_ptr->filter_weights != NULL)
+   {
+      png_uint_16p old = png_ptr->filter_weights;
+      png_ptr->filter_weights = NULL;
+      png_free(png_ptr, old);
+   }
+
+   if (png_ptr->inv_filter_weights != NULL)
+   {
+      png_uint_16p old = png_ptr->inv_filter_weights;
+      png_ptr->inv_filter_weights = NULL;
+      png_free(png_ptr, old);
+   }
+
+   /* Leave the filter_costs - this array is fixed size. */
+}
+
+static int
+png_init_filter_heuristics(png_structrp png_ptr, int heuristic_method,
+   int num_weights)
+{
+   if (png_ptr == NULL)
+      return 0;
+
+   /* Clear out the arrays */
+   png_reset_filter_heuristics(png_ptr);
+
+   /* Check arguments; the 'reset' function makes the correct settings for the
+    * unweighted case, but we must handle the weight case by initializing the
+    * arrays for the caller.
+    */
+   if (heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+   {
+      int i;
+
+      if (num_weights > 0)
+      {
+         png_ptr->prev_filters = (png_bytep)png_malloc(png_ptr,
+             (png_uint_32)((sizeof (png_byte)) * num_weights));
+
+         /* To make sure that the weighting starts out fairly */
+         for (i = 0; i < num_weights; i++)
+         {
+            png_ptr->prev_filters[i] = 255;
+         }
+
+         png_ptr->filter_weights = (png_uint_16p)png_malloc(png_ptr,
+             (png_uint_32)((sizeof (png_uint_16)) * num_weights));
+
+         png_ptr->inv_filter_weights = (png_uint_16p)png_malloc(png_ptr,
+             (png_uint_32)((sizeof (png_uint_16)) * num_weights));
+
+         for (i = 0; i < num_weights; i++)
+         {
+            png_ptr->inv_filter_weights[i] =
+            png_ptr->filter_weights[i] = PNG_WEIGHT_FACTOR;
+         }
+
+         /* Safe to set this now */
+         png_ptr->num_prev_filters = (png_byte)num_weights;
+      }
+
+      /* If, in the future, there are other filter methods, this would
+       * need to be based on png_ptr->filter.
+       */
+      if (png_ptr->filter_costs == NULL)
+      {
+         png_ptr->filter_costs = (png_uint_16p)png_malloc(png_ptr,
+             (png_uint_32)((sizeof (png_uint_16)) * PNG_FILTER_VALUE_LAST));
+
+         png_ptr->inv_filter_costs = (png_uint_16p)png_malloc(png_ptr,
+             (png_uint_32)((sizeof (png_uint_16)) * PNG_FILTER_VALUE_LAST));
+      }
+
+      for (i = 0; i < PNG_FILTER_VALUE_LAST; i++)
+      {
+         png_ptr->inv_filter_costs[i] =
+         png_ptr->filter_costs[i] = PNG_COST_FACTOR;
+      }
+
+      /* All the arrays are inited, safe to set this: */
+      png_ptr->heuristic_method = PNG_FILTER_HEURISTIC_WEIGHTED;
+
+      /* Return the 'ok' code. */
+      return 1;
+   }
+   else if (heuristic_method == PNG_FILTER_HEURISTIC_DEFAULT ||
+      heuristic_method == PNG_FILTER_HEURISTIC_UNWEIGHTED)
+   {
+      return 1;
+   }
+   else
+   {
+      png_warning(png_ptr, "Unknown filter heuristic method");
+      return 0;
+   }
+}
+
+/* Provide floating and fixed point APIs */
+#ifdef PNG_FLOATING_POINT_SUPPORTED
+void PNGAPI
+png_set_filter_heuristics(png_structrp png_ptr, int heuristic_method,
+    int num_weights, png_const_doublep filter_weights,
+    png_const_doublep filter_costs)
+{
+   png_debug(1, "in png_set_filter_heuristics");
+
+   /* The internal API allocates all the arrays and ensures that the elements of
+    * those arrays are set to the default value.
+    */
+   if (!png_init_filter_heuristics(png_ptr, heuristic_method, num_weights))
+      return;
+
+   /* If using the weighted method copy in the weights. */
+   if (heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+   {
+      int i;
+      for (i = 0; i < num_weights; i++)
+      {
+         if (filter_weights[i] <= 0.0)
+         {
+            png_ptr->inv_filter_weights[i] =
+            png_ptr->filter_weights[i] = PNG_WEIGHT_FACTOR;
+         }
+
+         else
+         {
+            png_ptr->inv_filter_weights[i] =
+                (png_uint_16)(PNG_WEIGHT_FACTOR*filter_weights[i]+.5);
+
+            png_ptr->filter_weights[i] =
+                (png_uint_16)(PNG_WEIGHT_FACTOR/filter_weights[i]+.5);
+         }
+      }
+
+      /* Here is where we set the relative costs of the different filters.  We
+       * should take the desired compression level into account when setting
+       * the costs, so that Paeth, for instance, has a high relative cost at low
+       * compression levels, while it has a lower relative cost at higher
+       * compression settings.  The filter types are in order of increasing
+       * relative cost, so it would be possible to do this with an algorithm.
+       */
+      for (i = 0; i < PNG_FILTER_VALUE_LAST; i++) if (filter_costs[i] >= 1.0)
+      {
+         png_ptr->inv_filter_costs[i] =
+             (png_uint_16)(PNG_COST_FACTOR / filter_costs[i] + .5);
+
+         png_ptr->filter_costs[i] =
+             (png_uint_16)(PNG_COST_FACTOR * filter_costs[i] + .5);
+      }
+   }
+}
+#endif /* FLOATING_POINT */
+
+#ifdef PNG_FIXED_POINT_SUPPORTED
+void PNGAPI
+png_set_filter_heuristics_fixed(png_structrp png_ptr, int heuristic_method,
+    int num_weights, png_const_fixed_point_p filter_weights,
+    png_const_fixed_point_p filter_costs)
+{
+   png_debug(1, "in png_set_filter_heuristics_fixed");
+
+   /* The internal API allocates all the arrays and ensures that the elements of
+    * those arrays are set to the default value.
+    */
+   if (!png_init_filter_heuristics(png_ptr, heuristic_method, num_weights))
+      return;
+
+   /* If using the weighted method copy in the weights. */
+   if (heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+   {
+      int i;
+      for (i = 0; i < num_weights; i++)
+      {
+         if (filter_weights[i] <= 0)
+         {
+            png_ptr->inv_filter_weights[i] =
+            png_ptr->filter_weights[i] = PNG_WEIGHT_FACTOR;
+         }
+
+         else
+         {
+            png_ptr->inv_filter_weights[i] = (png_uint_16)
+               ((PNG_WEIGHT_FACTOR*filter_weights[i]+PNG_FP_HALF)/PNG_FP_1);
+
+            png_ptr->filter_weights[i] = (png_uint_16)((PNG_WEIGHT_FACTOR*
+               PNG_FP_1+(filter_weights[i]/2))/filter_weights[i]);
+         }
+      }
+
+      /* Here is where we set the relative costs of the different filters.  We
+       * should take the desired compression level into account when setting
+       * the costs, so that Paeth, for instance, has a high relative cost at low
+       * compression levels, while it has a lower relative cost at higher
+       * compression settings.  The filter types are in order of increasing
+       * relative cost, so it would be possible to do this with an algorithm.
+       */
+      for (i = 0; i < PNG_FILTER_VALUE_LAST; i++)
+         if (filter_costs[i] >= PNG_FP_1)
+      {
+         png_uint_32 tmp;
+
+         /* Use a 32 bit unsigned temporary here because otherwise the
+          * intermediate value will be a 32 bit *signed* integer (ANSI rules)
+          * and this will get the wrong answer on division.
+          */
+         tmp = PNG_COST_FACTOR*PNG_FP_1 + (filter_costs[i]/2);
+         tmp /= filter_costs[i];
+
+         png_ptr->inv_filter_costs[i] = (png_uint_16)tmp;
+
+         tmp = PNG_COST_FACTOR * filter_costs[i] + PNG_FP_HALF;
+         tmp /= PNG_FP_1;
+
+         png_ptr->filter_costs[i] = (png_uint_16)tmp;
+      }
+   }
+}
+#endif /* FIXED_POINT */
+#endif /* PNG_WRITE_WEIGHTED_FILTER_SUPPORTED */
+
+void PNGAPI
+png_set_compression_level(png_structrp png_ptr, int level)
+{
+   png_debug(1, "in png_set_compression_level");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->zlib_level = level;
+}
+
+void PNGAPI
+png_set_compression_mem_level(png_structrp png_ptr, int mem_level)
+{
+   png_debug(1, "in png_set_compression_mem_level");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->zlib_mem_level = mem_level;
+}
+
+void PNGAPI
+png_set_compression_strategy(png_structrp png_ptr, int strategy)
+{
+   png_debug(1, "in png_set_compression_strategy");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* The flag setting here prevents the libpng dynamic selection of strategy.
+    */
+   png_ptr->flags |= PNG_FLAG_ZLIB_CUSTOM_STRATEGY;
+   png_ptr->zlib_strategy = strategy;
+}
+
+/* If PNG_WRITE_OPTIMIZE_CMF_SUPPORTED is defined, libpng will use a
+ * smaller value of window_bits if it can do so safely.
+ */
+void PNGAPI
+png_set_compression_window_bits(png_structrp png_ptr, int window_bits)
+{
+   if (png_ptr == NULL)
+      return;
+
+   /* Prior to 1.6.0 this would warn but then set the window_bits value, this
+    * meant that negative window bits values could be selected which would cause
+    * libpng to write a non-standard PNG file with raw deflate or gzip
+    * compressed IDAT or ancillary chunks.  Such files can be read and there is
+    * no warning on read, so this seems like a very bad idea.
+    */
+   if (window_bits > 15)
+   {
+      png_warning(png_ptr, "Only compression windows <= 32k supported by PNG");
+      window_bits = 15;
+   }
+
+   else if (window_bits < 8)
+   {
+      png_warning(png_ptr, "Only compression windows >= 256 supported by PNG");
+      window_bits = 8;
+   }
+
+   png_ptr->zlib_window_bits = window_bits;
+}
+
+void PNGAPI
+png_set_compression_method(png_structrp png_ptr, int method)
+{
+   png_debug(1, "in png_set_compression_method");
+
+   if (png_ptr == NULL)
+      return;
+
+   /* This would produce an invalid PNG file if it worked, but it doesn't and
+    * deflate will fault it, so it is harmless to just warn here.
+    */
+   if (method != 8)
+      png_warning(png_ptr, "Only compression method 8 is supported by PNG");
+
+   png_ptr->zlib_method = method;
+}
+
+/* The following were added to libpng-1.5.4 */
+#ifdef PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED
+void PNGAPI
+png_set_text_compression_level(png_structrp png_ptr, int level)
+{
+   png_debug(1, "in png_set_text_compression_level");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->zlib_text_level = level;
+}
+
+void PNGAPI
+png_set_text_compression_mem_level(png_structrp png_ptr, int mem_level)
+{
+   png_debug(1, "in png_set_text_compression_mem_level");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->zlib_text_mem_level = mem_level;
+}
+
+void PNGAPI
+png_set_text_compression_strategy(png_structrp png_ptr, int strategy)
+{
+   png_debug(1, "in png_set_text_compression_strategy");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->zlib_text_strategy = strategy;
+}
+
+/* If PNG_WRITE_OPTIMIZE_CMF_SUPPORTED is defined, libpng will use a
+ * smaller value of window_bits if it can do so safely.
+ */
+void PNGAPI
+png_set_text_compression_window_bits(png_structrp png_ptr, int window_bits)
+{
+   if (png_ptr == NULL)
+      return;
+
+   if (window_bits > 15)
+   {
+      png_warning(png_ptr, "Only compression windows <= 32k supported by PNG");
+      window_bits = 15;
+   }
+
+   else if (window_bits < 8)
+   {
+      png_warning(png_ptr, "Only compression windows >= 256 supported by PNG");
+      window_bits = 8;
+   }
+
+   png_ptr->zlib_text_window_bits = window_bits;
+}
+
+void PNGAPI
+png_set_text_compression_method(png_structrp png_ptr, int method)
+{
+   png_debug(1, "in png_set_text_compression_method");
+
+   if (png_ptr == NULL)
+      return;
+
+   if (method != 8)
+      png_warning(png_ptr, "Only compression method 8 is supported by PNG");
+
+   png_ptr->zlib_text_method = method;
+}
+#endif /* PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED */
+/* end of API added to libpng-1.5.4 */
+
+void PNGAPI
+png_set_write_status_fn(png_structrp png_ptr, png_write_status_ptr write_row_fn)
+{
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->write_row_fn = write_row_fn;
+}
+
+#ifdef PNG_WRITE_USER_TRANSFORM_SUPPORTED
+void PNGAPI
+png_set_write_user_transform_fn(png_structrp png_ptr, png_user_transform_ptr
+    write_user_transform_fn)
+{
+   png_debug(1, "in png_set_write_user_transform_fn");
+
+   if (png_ptr == NULL)
+      return;
+
+   png_ptr->transformations |= PNG_USER_TRANSFORM;
+   png_ptr->write_user_transform_fn = write_user_transform_fn;
+}
+#endif
+
+
+#ifdef PNG_INFO_IMAGE_SUPPORTED
+void PNGAPI
+png_write_png(png_structrp png_ptr, png_inforp info_ptr,
+    int transforms, voidp params)
+{
+   if (png_ptr == NULL || info_ptr == NULL)
+      return;
+
+   /* Write the file header information. */
+   png_write_info(png_ptr, info_ptr);
+
+   /* ------ these transformations don't touch the info structure ------- */
+
+#ifdef PNG_WRITE_INVERT_SUPPORTED
+   /* Invert monochrome pixels */
+   if (transforms & PNG_TRANSFORM_INVERT_MONO)
+      png_set_invert_mono(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_SHIFT_SUPPORTED
+   /* Shift the pixels up to a legal bit depth and fill in
+    * as appropriate to correctly scale the image.
+    */
+   if ((transforms & PNG_TRANSFORM_SHIFT)
+       && (info_ptr->valid & PNG_INFO_sBIT))
+      png_set_shift(png_ptr, &info_ptr->sig_bit);
+#endif
+
+#ifdef PNG_WRITE_PACK_SUPPORTED
+   /* Pack pixels into bytes */
+   if (transforms & PNG_TRANSFORM_PACKING)
+       png_set_packing(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_SWAP_ALPHA_SUPPORTED
+   /* Swap location of alpha bytes from ARGB to RGBA */
+   if (transforms & PNG_TRANSFORM_SWAP_ALPHA)
+      png_set_swap_alpha(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_FILLER_SUPPORTED
+   /* Pack XRGB/RGBX/ARGB/RGBA into RGB (4 channels -> 3 channels) */
+   if (transforms & PNG_TRANSFORM_STRIP_FILLER_AFTER)
+      png_set_filler(png_ptr, 0, PNG_FILLER_AFTER);
+
+   else if (transforms & PNG_TRANSFORM_STRIP_FILLER_BEFORE)
+      png_set_filler(png_ptr, 0, PNG_FILLER_BEFORE);
+#endif
+
+#ifdef PNG_WRITE_BGR_SUPPORTED
+   /* Flip BGR pixels to RGB */
+   if (transforms & PNG_TRANSFORM_BGR)
+      png_set_bgr(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_SWAP_SUPPORTED
+   /* Swap bytes of 16-bit files to most significant byte first */
+   if (transforms & PNG_TRANSFORM_SWAP_ENDIAN)
+      png_set_swap(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_PACKSWAP_SUPPORTED
+   /* Swap bits of 1, 2, 4 bit packed pixel formats */
+   if (transforms & PNG_TRANSFORM_PACKSWAP)
+      png_set_packswap(png_ptr);
+#endif
+
+#ifdef PNG_WRITE_INVERT_ALPHA_SUPPORTED
+   /* Invert the alpha channel from opacity to transparency */
+   if (transforms & PNG_TRANSFORM_INVERT_ALPHA)
+      png_set_invert_alpha(png_ptr);
+#endif
+
+   /* ----------------------- end of transformations ------------------- */
+
+   /* Write the bits */
+   if (info_ptr->valid & PNG_INFO_IDAT)
+       png_write_image(png_ptr, info_ptr->row_pointers);
+
+   /* It is REQUIRED to call this to finish writing the rest of the file */
+   png_write_end(png_ptr, info_ptr);
+
+   PNG_UNUSED(transforms)   /* Quiet compiler warnings */
+   PNG_UNUSED(params)
+}
+#endif
+
+
+#ifdef PNG_SIMPLIFIED_WRITE_SUPPORTED
+#ifdef PNG_STDIO_SUPPORTED /* currently required for png_image_write_* */
+/* Initialize the write structure - general purpose utility. */
+static int
+png_image_write_init(png_imagep image)
+{
+   png_structp png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, image,
+          png_safe_error, png_safe_warning);
+
+   if (png_ptr != NULL)
+   {
+      png_infop info_ptr = png_create_info_struct(png_ptr);
+
+      if (info_ptr != NULL)
+      {
+         png_controlp control = png_voidcast(png_controlp,
+            png_malloc_warn(png_ptr, (sizeof *control)));
+
+         if (control != NULL)
+         {
+            memset(control, 0, (sizeof *control));
+
+            control->png_ptr = png_ptr;
+            control->info_ptr = info_ptr;
+            control->for_write = 1;
+
+            image->opaque = control;
+            return 1;
+         }
+
+         /* Error clean up */
+         png_destroy_info_struct(png_ptr, &info_ptr);
+      }
+
+      png_destroy_write_struct(&png_ptr, NULL);
+   }
+
+   return png_image_error(image, "png_image_write_: out of memory");
+}
+
+/* Arguments to png_image_write_main: */
+typedef struct
+{
+   /* Arguments: */
+   png_imagep      image;
+   png_const_voidp buffer;
+   png_int_32      row_stride;
+   png_const_voidp colormap;
+   int             convert_to_8bit;
+   /* Local variables: */
+   png_const_voidp first_row;
+   ptrdiff_t       row_bytes;
+   png_voidp       local_row;
+} png_image_write_control;
+
+/* Write png_uint_16 input to a 16-bit PNG; the png_ptr has already been set to
+ * do any necessary byte swapping.  The component order is defined by the
+ * png_image format value.
+ */
+static int
+png_write_image_16bit(png_voidp argument)
+{
+   png_image_write_control *display = png_voidcast(png_image_write_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+
+   png_const_uint_16p input_row = png_voidcast(png_const_uint_16p,
+      display->first_row);
+   png_uint_16p output_row = png_voidcast(png_uint_16p, display->local_row);
+   png_uint_16p row_end;
+   const int channels = (image->format & PNG_FORMAT_FLAG_COLOR) ? 3 : 1;
+   int aindex = 0;
+   png_uint_32 y = image->height;
+
+   if (image->format & PNG_FORMAT_FLAG_ALPHA)
+   {
+      if (image->format & PNG_FORMAT_FLAG_AFIRST)
+      {
+         aindex = -1;
+         ++input_row; /* To point to the first component */
+         ++output_row;
+      }
+
+      else
+         aindex = channels;
+   }
+
+   else
+      png_error(png_ptr, "png_write_image: internal call error");
+
+   /* Work out the output row end and count over this, note that the increment
+    * above to 'row' means that row_end can actually be beyond the end of the
+    * row; this is correct.
+    */
+   row_end = output_row + image->width * (channels+1);
+
+   while (y-- > 0)
+   {
+      png_const_uint_16p in_ptr = input_row;
+      png_uint_16p out_ptr = output_row;
+
+      while (out_ptr < row_end)
+      {
+         const png_uint_16 alpha = in_ptr[aindex];
+         png_uint_32 reciprocal = 0;
+         int c;
+
+         out_ptr[aindex] = alpha;
+
+         /* Calculate a reciprocal.  The correct calculation is simply
+          * component/alpha*65535 << 15. (I.e. 15 bits of precision); this
+          * allows correct rounding by adding .5 before the shift.  'reciprocal'
+          * is only initialized when required.
+          */
+         if (alpha > 0 && alpha < 65535)
+            reciprocal = ((0xffff<<15)+(alpha>>1))/alpha;
+
+         c = channels;
+         do /* always at least one channel */
+         {
+            png_uint_16 component = *in_ptr++;
+
+            /* The following gives 65535 for an alpha of 0, which is fine,
+             * otherwise if 0/0 is represented as some other value there is more
+             * likely to be a discontinuity which will probably damage
+             * compression when moving from a fully transparent area to a
+             * nearly transparent one.  (The assumption here is that opaque
+             * areas tend not to be 0 intensity.)
+             */
+            if (component >= alpha)
+               component = 65535;
+
+            /* component<alpha, so component/alpha is less than one and
+             * component*reciprocal is less than 2^31.
+             */
+            else if (component > 0 && alpha < 65535)
+            {
+               png_uint_32 calc = component * reciprocal;
+               calc += 16384; /* round to nearest */
+               component = (png_uint_16)(calc >> 15);
+            }
+
+            *out_ptr++ = component;
+         }
+         while (--c > 0);
+
+         /* Skip to next component (skip the intervening alpha channel) */
+         ++in_ptr;
+         ++out_ptr;
+      }
+
+      png_write_row(png_ptr, png_voidcast(png_const_bytep, display->local_row));
+      input_row += display->row_bytes/(sizeof (png_uint_16));
+   }
+
+   return 1;
+}
+
+/* Given 16-bit input (1 to 4 channels) write 8-bit output.  If an alpha channel
+ * is present it must be removed from the components, the components are then
+ * written in sRGB encoding.  No components are added or removed.
+ *
+ * Calculate an alpha reciprocal to reverse pre-multiplication.  As above the
+ * calculation can be done to 15 bits of accuracy; however, the output needs to
+ * be scaled in the range 0..255*65535, so include that scaling here.
+ */
+#define UNP_RECIPROCAL(alpha) ((((0xffff*0xff)<<7)+(alpha>>1))/alpha)
+
+static png_byte
+png_unpremultiply(png_uint_32 component, png_uint_32 alpha,
+   png_uint_32 reciprocal/*from the above macro*/)
+{
+   /* The following gives 1.0 for an alpha of 0, which is fine, otherwise if 0/0
+    * is represented as some other value there is more likely to be a
+    * discontinuity which will probably damage compression when moving from a
+    * fully transparent area to a nearly transparent one.  (The assumption here
+    * is that opaque areas tend not to be 0 intensity.)
+    *
+    * There is a rounding problem here; if alpha is less than 128 it will end up
+    * as 0 when scaled to 8 bits.  To avoid introducing spurious colors into the
+    * output change for this too.
+    */
+   if (component >= alpha || alpha < 128)
+      return 255;
+
+   /* component<alpha, so component/alpha is less than one and
+    * component*reciprocal is less than 2^31.
+    */
+   else if (component > 0)
+   {
+      /* The test is that alpha/257 (rounded) is less than 255, the first value
+       * that becomes 255 is 65407.
+       * NOTE: this must agree with the PNG_DIV257 macro (which must, therefore,
+       * be exact!)  [Could also test reciprocal != 0]
+       */
+      if (alpha < 65407)
+      {
+         component *= reciprocal;
+         component += 64; /* round to nearest */
+         component >>= 7;
+      }
+
+      else
+         component *= 255;
+
+      /* Convert the component to sRGB. */
+      return (png_byte)PNG_sRGB_FROM_LINEAR(component);
+   }
+
+   else
+      return 0;
+}
+
+static int
+png_write_image_8bit(png_voidp argument)
+{
+   png_image_write_control *display = png_voidcast(png_image_write_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+
+   png_const_uint_16p input_row = png_voidcast(png_const_uint_16p,
+      display->first_row);
+   png_bytep output_row = png_voidcast(png_bytep, display->local_row);
+   png_uint_32 y = image->height;
+   const int channels = (image->format & PNG_FORMAT_FLAG_COLOR) ? 3 : 1;
+
+   if (image->format & PNG_FORMAT_FLAG_ALPHA)
+   {
+      png_bytep row_end;
+      int aindex;
+
+      if (image->format & PNG_FORMAT_FLAG_AFIRST)
+      {
+         aindex = -1;
+         ++input_row; /* To point to the first component */
+         ++output_row;
+      }
+
+      else
+         aindex = channels;
+
+      /* Use row_end in place of a loop counter: */
+      row_end = output_row + image->width * (channels+1);
+
+      while (y-- > 0)
+      {
+         png_const_uint_16p in_ptr = input_row;
+         png_bytep out_ptr = output_row;
+
+         while (out_ptr < row_end)
+         {
+            png_uint_16 alpha = in_ptr[aindex];
+            png_byte alphabyte = (png_byte)PNG_DIV257(alpha);
+            png_uint_32 reciprocal = 0;
+            int c;
+
+            /* Scale and write the alpha channel. */
+            out_ptr[aindex] = alphabyte;
+
+            if (alphabyte > 0 && alphabyte < 255)
+               reciprocal = UNP_RECIPROCAL(alpha);
+
+            c = channels;
+            do /* always at least one channel */
+               *out_ptr++ = png_unpremultiply(*in_ptr++, alpha, reciprocal);
+            while (--c > 0);
+
+            /* Skip to next component (skip the intervening alpha channel) */
+            ++in_ptr;
+            ++out_ptr;
+         } /* while out_ptr < row_end */
+
+         png_write_row(png_ptr, png_voidcast(png_const_bytep,
+            display->local_row));
+         input_row += display->row_bytes/(sizeof (png_uint_16));
+      } /* while y */
+   }
+
+   else
+   {
+      /* No alpha channel, so the row_end really is the end of the row and it
+       * is sufficient to loop over the components one by one.
+       */
+      png_bytep row_end = output_row + image->width * channels;
+
+      while (y-- > 0)
+      {
+         png_const_uint_16p in_ptr = input_row;
+         png_bytep out_ptr = output_row;
+
+         while (out_ptr < row_end)
+         {
+            png_uint_32 component = *in_ptr++;
+
+            component *= 255;
+            *out_ptr++ = (png_byte)PNG_sRGB_FROM_LINEAR(component);
+         }
+
+         png_write_row(png_ptr, output_row);
+         input_row += display->row_bytes/(sizeof (png_uint_16));
+      }
+   }
+
+   return 1;
+}
+
+static void
+png_image_set_PLTE(png_image_write_control *display)
+{
+   const png_imagep image = display->image;
+   const void *cmap = display->colormap;
+   const int entries = image->colormap_entries > 256 ? 256 :
+      (int)image->colormap_entries;
+
+   /* NOTE: the caller must check for cmap != NULL and entries != 0 */
+   const png_uint_32 format = image->format;
+   const int channels = PNG_IMAGE_SAMPLE_CHANNELS(format);
+
+#  ifdef PNG_FORMAT_BGR_SUPPORTED
+      const int afirst = (format & PNG_FORMAT_FLAG_AFIRST) != 0 &&
+         (format & PNG_FORMAT_FLAG_ALPHA) != 0;
+#  else
+#     define afirst 0
+#  endif
+
+#  ifdef PNG_FORMAT_BGR_SUPPORTED
+      const int bgr = (format & PNG_FORMAT_FLAG_BGR) ? 2 : 0;
+#  else
+#     define bgr 0
+#  endif
+
+   int i, num_trans;
+   png_color palette[256];
+   png_byte tRNS[256];
+
+   memset(tRNS, 255, (sizeof tRNS));
+   memset(palette, 0, (sizeof palette));
+
+   for (i=num_trans=0; i<entries; ++i)
+   {
+      /* This gets automatically converted to sRGB with reversal of the
+       * pre-multiplication if the color-map has an alpha channel.
+       */
+      if (format & PNG_FORMAT_FLAG_LINEAR)
+      {
+         png_const_uint_16p entry = png_voidcast(png_const_uint_16p, cmap);
+
+         entry += i * channels;
+
+         if (channels & 1) /* no alpha */
+         {
+            if (channels >= 3) /* RGB */
+            {
+               palette[i].blue = (png_byte)PNG_sRGB_FROM_LINEAR(255 *
+                  entry[(2 ^ bgr)]);
+               palette[i].green = (png_byte)PNG_sRGB_FROM_LINEAR(255 *
+                  entry[1]);
+               palette[i].red = (png_byte)PNG_sRGB_FROM_LINEAR(255 *
+                  entry[bgr]);
+            }
+
+            else /* Gray */
+               palette[i].blue = palette[i].red = palette[i].green =
+                  (png_byte)PNG_sRGB_FROM_LINEAR(255 * *entry);
+         }
+
+         else /* alpha */
+         {
+            png_uint_16 alpha = entry[afirst ? 0 : channels-1];
+            png_byte alphabyte = (png_byte)PNG_DIV257(alpha);
+            png_uint_32 reciprocal = 0;
+
+            /* Calculate a reciprocal, as in the png_write_image_8bit code above
+             * this is designed to produce a value scaled to 255*65535 when
+             * divided by 128 (i.e. asr 7).
+             */
+            if (alphabyte > 0 && alphabyte < 255)
+               reciprocal = (((0xffff*0xff)<<7)+(alpha>>1))/alpha;
+
+            tRNS[i] = alphabyte;
+            if (alphabyte < 255)
+               num_trans = i+1;
+
+            if (channels >= 3) /* RGB */
+            {
+               palette[i].blue = png_unpremultiply(entry[afirst + (2 ^ bgr)],
+                  alpha, reciprocal);
+               palette[i].green = png_unpremultiply(entry[afirst + 1], alpha,
+                  reciprocal);
+               palette[i].red = png_unpremultiply(entry[afirst + bgr], alpha,
+                  reciprocal);
+            }
+
+            else /* gray */
+               palette[i].blue = palette[i].red = palette[i].green =
+                  png_unpremultiply(entry[afirst], alpha, reciprocal);
+         }
+      }
+
+      else /* Color-map has sRGB values */
+      {
+         png_const_bytep entry = png_voidcast(png_const_bytep, cmap);
+
+         entry += i * channels;
+
+         switch (channels)
+         {
+            case 4:
+               tRNS[i] = entry[afirst ? 0 : 3];
+               if (tRNS[i] < 255)
+                  num_trans = i+1;
+               /* FALL THROUGH */
+            case 3:
+               palette[i].blue = entry[afirst + (2 ^ bgr)];
+               palette[i].green = entry[afirst + 1];
+               palette[i].red = entry[afirst + bgr];
+               break;
+
+            case 2:
+               tRNS[i] = entry[1 ^ afirst];
+               if (tRNS[i] < 255)
+                  num_trans = i+1;
+               /* FALL THROUGH */
+            case 1:
+               palette[i].blue = palette[i].red = palette[i].green =
+                  entry[afirst];
+               break;
+
+            default:
+               break;
+         }
+      }
+   }
+
+#  ifdef afirst
+#     undef afirst
+#  endif
+#  ifdef bgr
+#     undef bgr
+#  endif
+
+   png_set_PLTE(image->opaque->png_ptr, image->opaque->info_ptr, palette,
+      entries);
+
+   if (num_trans > 0)
+      png_set_tRNS(image->opaque->png_ptr, image->opaque->info_ptr, tRNS,
+         num_trans, NULL);
+
+   image->colormap_entries = entries;
+}
+
+static int
+png_image_write_main(png_voidp argument)
+{
+   png_image_write_control *display = png_voidcast(png_image_write_control*,
+      argument);
+   png_imagep image = display->image;
+   png_structrp png_ptr = image->opaque->png_ptr;
+   png_inforp info_ptr = image->opaque->info_ptr;
+   png_uint_32 format = image->format;
+
+   int colormap = (format & PNG_FORMAT_FLAG_COLORMAP) != 0;
+   int linear = !colormap && (format & PNG_FORMAT_FLAG_LINEAR) != 0; /* input */
+   int alpha = !colormap && (format & PNG_FORMAT_FLAG_ALPHA) != 0;
+   int write_16bit = linear && !colormap && !display->convert_to_8bit;
+
+#  ifdef PNG_BENIGN_ERRORS_SUPPORTED
+      /* Make sure we error out on any bad situation */
+      png_set_benign_errors(png_ptr, 0/*error*/);
+#  endif
+
+   /* Default the 'row_stride' parameter if required. */
+   if (display->row_stride == 0)
+      display->row_stride = PNG_IMAGE_ROW_STRIDE(*image);
+
+   /* Set the required transforms then write the rows in the correct order. */
+   if (format & PNG_FORMAT_FLAG_COLORMAP)
+   {
+      if (display->colormap != NULL && image->colormap_entries > 0)
+      {
+         png_uint_32 entries = image->colormap_entries;
+
+         png_set_IHDR(png_ptr, info_ptr, image->width, image->height,
+            entries > 16 ? 8 : (entries > 4 ? 4 : (entries > 2 ? 2 : 1)),
+            PNG_COLOR_TYPE_PALETTE, PNG_INTERLACE_NONE,
+            PNG_COMPRESSION_TYPE_BASE, PNG_FILTER_TYPE_BASE);
+
+         png_image_set_PLTE(display);
+      }
+
+      else
+         png_error(image->opaque->png_ptr,
+            "no color-map for color-mapped image");
+   }
+
+   else
+      png_set_IHDR(png_ptr, info_ptr, image->width, image->height,
+         write_16bit ? 16 : 8,
+         ((format & PNG_FORMAT_FLAG_COLOR) ? PNG_COLOR_MASK_COLOR : 0) +
+         ((format & PNG_FORMAT_FLAG_ALPHA) ? PNG_COLOR_MASK_ALPHA : 0),
+         PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_BASE, PNG_FILTER_TYPE_BASE);
+
+   /* Counter-intuitively the data transformations must be called *after*
+    * png_write_info, not before as in the read code, but the 'set' functions
+    * must still be called before.  Just set the color space information, never
+    * write an interlaced image.
+    */
+
+   if (write_16bit)
+   {
+      /* The gamma here is 1.0 (linear) and the cHRM chunk matches sRGB. */
+      png_set_gAMA_fixed(png_ptr, info_ptr, PNG_GAMMA_LINEAR);
+
+      if (!(image->flags & PNG_IMAGE_FLAG_COLORSPACE_NOT_sRGB))
+         png_set_cHRM_fixed(png_ptr, info_ptr,
+            /* color      x       y */
+            /* white */ 31270, 32900,
+            /* red   */ 64000, 33000,
+            /* green */ 30000, 60000,
+            /* blue  */ 15000,  6000
+         );
+   }
+
+   else if (!(image->flags & PNG_IMAGE_FLAG_COLORSPACE_NOT_sRGB))
+      png_set_sRGB(png_ptr, info_ptr, PNG_sRGB_INTENT_PERCEPTUAL);
+
+   /* Else writing an 8-bit file and the *colors* aren't sRGB, but the 8-bit
+    * space must still be gamma encoded.
+    */
+   else
+      png_set_gAMA_fixed(png_ptr, info_ptr, PNG_GAMMA_sRGB_INVERSE);
+
+   /* Write the file header. */
+   png_write_info(png_ptr, info_ptr);
+
+   /* Now set up the data transformations (*after* the header is written),
+    * remove the handled transformations from the 'format' flags for checking.
+    *
+    * First check for a little endian system if writing 16 bit files.
+    */
+   if (write_16bit)
+   {
+      PNG_CONST png_uint_16 le = 0x0001;
+
+      if (*(png_const_bytep)&le)
+         png_set_swap(png_ptr);
+   }
+
+#  ifdef PNG_SIMPLIFIED_WRITE_BGR_SUPPORTED
+      if (format & PNG_FORMAT_FLAG_BGR)
+      {
+         if (!colormap && (format & PNG_FORMAT_FLAG_COLOR) != 0)
+            png_set_bgr(png_ptr);
+         format &= ~PNG_FORMAT_FLAG_BGR;
+      }
+#  endif
+
+#  ifdef PNG_SIMPLIFIED_WRITE_AFIRST_SUPPORTED
+      if (format & PNG_FORMAT_FLAG_AFIRST)
+      {
+         if (!colormap && (format & PNG_FORMAT_FLAG_ALPHA) != 0)
+            png_set_swap_alpha(png_ptr);
+         format &= ~PNG_FORMAT_FLAG_AFIRST;
+      }
+#  endif
+
+   /* If there are 16 or fewer color-map entries we wrote a lower bit depth
+    * above, but the application data is still byte packed.
+    */
+   if (colormap && image->colormap_entries <= 16)
+      png_set_packing(png_ptr);
+
+   /* That should have handled all (both) the transforms. */
+   if ((format & ~(png_uint_32)(PNG_FORMAT_FLAG_COLOR | PNG_FORMAT_FLAG_LINEAR |
+         PNG_FORMAT_FLAG_ALPHA | PNG_FORMAT_FLAG_COLORMAP)) != 0)
+      png_error(png_ptr, "png_write_image: unsupported transformation");
+
+   {
+      png_const_bytep row = png_voidcast(png_const_bytep, display->buffer);
+      ptrdiff_t row_bytes = display->row_stride;
+
+      if (linear)
+         row_bytes *= (sizeof (png_uint_16));
+
+      if (row_bytes < 0)
+         row += (image->height-1) * (-row_bytes);
+
+      display->first_row = row;
+      display->row_bytes = row_bytes;
+   }
+
+   /* Apply 'fast' options if the flag is set. */
+   if ((image->flags & PNG_IMAGE_FLAG_FAST) != 0)
+   {
+      png_set_filter(png_ptr, PNG_FILTER_TYPE_BASE, PNG_NO_FILTERS);
+      /* NOTE: determined by experiment using pngstest, this reflects some
+       * balance between the time to write the image once and the time to read
+       * it about 50 times.  The speed-up in pngstest was about 10-20% of the
+       * total (user) time on a heavily loaded system.
+       */
+      png_set_compression_level(png_ptr, 3);
+   }
+
+   /* Check for the cases that currently require a pre-transform on the row
+    * before it is written.  This only applies when the input is 16-bit and
+    * either there is an alpha channel or it is converted to 8-bit.
+    */
+   if ((linear && alpha) || (!colormap && display->convert_to_8bit))
+   {
+      png_bytep row = png_voidcast(png_bytep, png_malloc(png_ptr,
+         png_get_rowbytes(png_ptr, info_ptr)));
+      int result;
+
+      display->local_row = row;
+      if (write_16bit)
+         result = png_safe_execute(image, png_write_image_16bit, display);
+      else
+         result = png_safe_execute(image, png_write_image_8bit, display);
+      display->local_row = NULL;
+
+      png_free(png_ptr, row);
+
+      /* Skip the 'write_end' on error: */
+      if (!result)
+         return 0;
+   }
+
+   /* Otherwise this is the case where the input is in a format currently
+    * supported by the rest of the libpng write code; call it directly.
+    */
+   else
+   {
+      png_const_bytep row = png_voidcast(png_const_bytep, display->first_row);
+      ptrdiff_t row_bytes = display->row_bytes;
+      png_uint_32 y = image->height;
+
+      while (y-- > 0)
+      {
+         png_write_row(png_ptr, row);
+         row += row_bytes;
+      }
+   }
+
+   png_write_end(png_ptr, info_ptr);
+   return 1;
+}
+
+int PNGAPI
+png_image_write_to_stdio(png_imagep image, FILE *file, int convert_to_8bit,
+   const void *buffer, png_int_32 row_stride, const void *colormap)
+{
+   /* Write the image to the given (FILE*). */
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      if (file != NULL)
+      {
+         if (png_image_write_init(image))
+         {
+            png_image_write_control display;
+            int result;
+
+            /* This is slightly evil, but png_init_io doesn't do anything other
+             * than this and we haven't changed the standard IO functions so
+             * this saves a 'safe' function.
+             */
+            image->opaque->png_ptr->io_ptr = file;
+
+            memset(&display, 0, (sizeof display));
+            display.image = image;
+            display.buffer = buffer;
+            display.row_stride = row_stride;
+            display.colormap = colormap;
+            display.convert_to_8bit = convert_to_8bit;
+
+            result = png_safe_execute(image, png_image_write_main, &display);
+            png_image_free(image);
+            return result;
+         }
+
+         else
+            return 0;
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_write_to_stdio: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_write_to_stdio: incorrect PNG_IMAGE_VERSION");
+
+   else
+      return 0;
+}
+
+int PNGAPI
+png_image_write_to_file(png_imagep image, const char *file_name,
+   int convert_to_8bit, const void *buffer, png_int_32 row_stride,
+   const void *colormap)
+{
+   /* Write the image to the named file. */
+   if (image != NULL && image->version == PNG_IMAGE_VERSION)
+   {
+      if (file_name != NULL)
+      {
+         FILE *fp = fopen(file_name, "wb");
+
+         if (fp != NULL)
+         {
+            if (png_image_write_to_stdio(image, fp, convert_to_8bit, buffer,
+               row_stride, colormap))
+            {
+               int error; /* from fflush/fclose */
+
+               /* Make sure the file is flushed correctly. */
+               if (fflush(fp) == 0 && ferror(fp) == 0)
+               {
+                  if (fclose(fp) == 0)
+                     return 1;
+
+                  error = errno; /* from fclose */
+               }
+
+               else
+               {
+                  error = errno; /* from fflush or ferror */
+                  (void)fclose(fp);
+               }
+
+               (void)remove(file_name);
+               /* The image has already been cleaned up; this is just used to
+                * set the error (because the original write succeeded).
+                */
+               return png_image_error(image, strerror(error));
+            }
+
+            else
+            {
+               /* Clean up: just the opened file. */
+               (void)fclose(fp);
+               (void)remove(file_name);
+               return 0;
+            }
+         }
+
+         else
+            return png_image_error(image, strerror(errno));
+      }
+
+      else
+         return png_image_error(image,
+            "png_image_write_to_file: invalid argument");
+   }
+
+   else if (image != NULL)
+      return png_image_error(image,
+         "png_image_write_to_file: incorrect PNG_IMAGE_VERSION");
+
+   else
+      return 0;
+}
+#endif /* PNG_STDIO_SUPPORTED */
+#endif /* SIMPLIFIED_WRITE */
+#endif /* PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngwtran.c b/ml/dlib/dlib/external/libpng/pngwtran.c
new file mode 100644
index 000000000..98703f8c8
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngwtran.c
@@ -0,0 +1,637 @@
+
+/* pngwtran.c - transforms the data in a row for PNG writers
+ *
+ * Last changed in libpng 1.6.0 [February 14, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_WRITE_SUPPORTED
+
+#ifdef PNG_WRITE_TRANSFORMS_SUPPORTED
+/* Transform the data according to the user's wishes.  The order of
+ * transformations is significant.
+ */
+void /* PRIVATE */
+png_do_write_transformations(png_structrp png_ptr, png_row_infop row_info)
+{
+   png_debug(1, "in png_do_write_transformations");
+
+   if (png_ptr == NULL)
+      return;
+
+#ifdef PNG_WRITE_USER_TRANSFORM_SUPPORTED
+   if (png_ptr->transformations & PNG_USER_TRANSFORM)
+      if (png_ptr->write_user_transform_fn != NULL)
+         (*(png_ptr->write_user_transform_fn)) /* User write transform
+                                                 function */
+             (png_ptr,  /* png_ptr */
+             row_info,  /* row_info: */
+                /*  png_uint_32 width;       width of row */
+                /*  png_size_t rowbytes;     number of bytes in row */
+                /*  png_byte color_type;     color type of pixels */
+                /*  png_byte bit_depth;      bit depth of samples */
+                /*  png_byte channels;       number of channels (1-4) */
+                /*  png_byte pixel_depth;    bits per pixel (depth*channels) */
+             png_ptr->row_buf + 1);      /* start of pixel data for row */
+#endif
+
+#ifdef PNG_WRITE_FILLER_SUPPORTED
+   if (png_ptr->transformations & PNG_FILLER)
+      png_do_strip_channel(row_info, png_ptr->row_buf + 1,
+         !(png_ptr->flags & PNG_FLAG_FILLER_AFTER));
+#endif
+
+#ifdef PNG_WRITE_PACKSWAP_SUPPORTED
+   if (png_ptr->transformations & PNG_PACKSWAP)
+      png_do_packswap(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_WRITE_PACK_SUPPORTED
+   if (png_ptr->transformations & PNG_PACK)
+      png_do_pack(row_info, png_ptr->row_buf + 1,
+          (png_uint_32)png_ptr->bit_depth);
+#endif
+
+#ifdef PNG_WRITE_SWAP_SUPPORTED
+   if (png_ptr->transformations & PNG_SWAP_BYTES)
+      png_do_swap(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_WRITE_SHIFT_SUPPORTED
+   if (png_ptr->transformations & PNG_SHIFT)
+      png_do_shift(row_info, png_ptr->row_buf + 1,
+          &(png_ptr->shift));
+#endif
+
+#ifdef PNG_WRITE_SWAP_ALPHA_SUPPORTED
+   if (png_ptr->transformations & PNG_SWAP_ALPHA)
+      png_do_write_swap_alpha(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_WRITE_INVERT_ALPHA_SUPPORTED
+   if (png_ptr->transformations & PNG_INVERT_ALPHA)
+      png_do_write_invert_alpha(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_WRITE_BGR_SUPPORTED
+   if (png_ptr->transformations & PNG_BGR)
+      png_do_bgr(row_info, png_ptr->row_buf + 1);
+#endif
+
+#ifdef PNG_WRITE_INVERT_SUPPORTED
+   if (png_ptr->transformations & PNG_INVERT_MONO)
+      png_do_invert(row_info, png_ptr->row_buf + 1);
+#endif
+}
+
+#ifdef PNG_WRITE_PACK_SUPPORTED
+/* Pack pixels into bytes.  Pass the true bit depth in bit_depth.  The
+ * row_info bit depth should be 8 (one pixel per byte).  The channels
+ * should be 1 (this only happens on grayscale and paletted images).
+ */
+void /* PRIVATE */
+png_do_pack(png_row_infop row_info, png_bytep row, png_uint_32 bit_depth)
+{
+   png_debug(1, "in png_do_pack");
+
+   if (row_info->bit_depth == 8 &&
+      row_info->channels == 1)
+   {
+      switch ((int)bit_depth)
+      {
+         case 1:
+         {
+            png_bytep sp, dp;
+            int mask, v;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            sp = row;
+            dp = row;
+            mask = 0x80;
+            v = 0;
+
+            for (i = 0; i < row_width; i++)
+            {
+               if (*sp != 0)
+                  v |= mask;
+
+               sp++;
+
+               if (mask > 1)
+                  mask >>= 1;
+
+               else
+               {
+                  mask = 0x80;
+                  *dp = (png_byte)v;
+                  dp++;
+                  v = 0;
+               }
+            }
+
+            if (mask != 0x80)
+               *dp = (png_byte)v;
+
+            break;
+         }
+
+         case 2:
+         {
+            png_bytep sp, dp;
+            int shift, v;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            sp = row;
+            dp = row;
+            shift = 6;
+            v = 0;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_byte value;
+
+               value = (png_byte)(*sp & 0x03);
+               v |= (value << shift);
+
+               if (shift == 0)
+               {
+                  shift = 6;
+                  *dp = (png_byte)v;
+                  dp++;
+                  v = 0;
+               }
+
+               else
+                  shift -= 2;
+
+               sp++;
+            }
+
+            if (shift != 6)
+               *dp = (png_byte)v;
+
+            break;
+         }
+
+         case 4:
+         {
+            png_bytep sp, dp;
+            int shift, v;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            sp = row;
+            dp = row;
+            shift = 4;
+            v = 0;
+
+            for (i = 0; i < row_width; i++)
+            {
+               png_byte value;
+
+               value = (png_byte)(*sp & 0x0f);
+               v |= (value << shift);
+
+               if (shift == 0)
+               {
+                  shift = 4;
+                  *dp = (png_byte)v;
+                  dp++;
+                  v = 0;
+               }
+
+               else
+                  shift -= 4;
+
+               sp++;
+            }
+
+            if (shift != 4)
+               *dp = (png_byte)v;
+
+            break;
+         }
+
+         default:
+            break;
+      }
+
+      row_info->bit_depth = (png_byte)bit_depth;
+      row_info->pixel_depth = (png_byte)(bit_depth * row_info->channels);
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth,
+          row_info->width);
+   }
+}
+#endif
+
+#ifdef PNG_WRITE_SHIFT_SUPPORTED
+/* Shift pixel values to take advantage of whole range.  Pass the
+ * true number of bits in bit_depth.  The row should be packed
+ * according to row_info->bit_depth.  Thus, if you had a row of
+ * bit depth 4, but the pixels only had values from 0 to 7, you
+ * would pass 3 as bit_depth, and this routine would translate the
+ * data to 0 to 15.
+ */
+void /* PRIVATE */
+png_do_shift(png_row_infop row_info, png_bytep row,
+    png_const_color_8p bit_depth)
+{
+   png_debug(1, "in png_do_shift");
+
+   if (row_info->color_type != PNG_COLOR_TYPE_PALETTE)
+   {
+      int shift_start[4], shift_dec[4];
+      int channels = 0;
+
+      if (row_info->color_type & PNG_COLOR_MASK_COLOR)
+      {
+         shift_start[channels] = row_info->bit_depth - bit_depth->red;
+         shift_dec[channels] = bit_depth->red;
+         channels++;
+
+         shift_start[channels] = row_info->bit_depth - bit_depth->green;
+         shift_dec[channels] = bit_depth->green;
+         channels++;
+
+         shift_start[channels] = row_info->bit_depth - bit_depth->blue;
+         shift_dec[channels] = bit_depth->blue;
+         channels++;
+      }
+
+      else
+      {
+         shift_start[channels] = row_info->bit_depth - bit_depth->gray;
+         shift_dec[channels] = bit_depth->gray;
+         channels++;
+      }
+
+      if (row_info->color_type & PNG_COLOR_MASK_ALPHA)
+      {
+         shift_start[channels] = row_info->bit_depth - bit_depth->alpha;
+         shift_dec[channels] = bit_depth->alpha;
+         channels++;
+      }
+
+      /* With low row depths, could only be grayscale, so one channel */
+      if (row_info->bit_depth < 8)
+      {
+         png_bytep bp = row;
+         png_size_t i;
+         unsigned int mask;
+         png_size_t row_bytes = row_info->rowbytes;
+
+         if (bit_depth->gray == 1 && row_info->bit_depth == 2)
+            mask = 0x55;
+
+         else if (row_info->bit_depth == 4 && bit_depth->gray == 3)
+            mask = 0x11;
+
+         else
+            mask = 0xff;
+
+         for (i = 0; i < row_bytes; i++, bp++)
+         {
+            int j;
+            unsigned int v, out;
+
+            v = *bp;
+            out = 0;
+
+            for (j = shift_start[0]; j > -shift_dec[0]; j -= shift_dec[0])
+            {
+               if (j > 0)
+                  out |= v << j;
+
+               else
+                  out |= (v >> (-j)) & mask;
+            }
+
+            *bp = (png_byte)(out & 0xff);
+         }
+      }
+
+      else if (row_info->bit_depth == 8)
+      {
+         png_bytep bp = row;
+         png_uint_32 i;
+         png_uint_32 istop = channels * row_info->width;
+
+         for (i = 0; i < istop; i++, bp++)
+         {
+
+            const unsigned int c = i%channels;
+            int j;
+            unsigned int v, out;
+
+            v = *bp;
+            out = 0;
+
+            for (j = shift_start[c]; j > -shift_dec[c]; j -= shift_dec[c])
+            {
+               if (j > 0)
+                  out |= v << j;
+
+               else
+                  out |= v >> (-j);
+            }
+
+            *bp = (png_byte)(out & 0xff);
+         }
+      }
+
+      else
+      {
+         png_bytep bp;
+         png_uint_32 i;
+         png_uint_32 istop = channels * row_info->width;
+
+         for (bp = row, i = 0; i < istop; i++)
+         {
+            const unsigned int c = i%channels;
+            int j;
+            unsigned int value, v;
+
+            v = png_get_uint_16(bp);
+            value = 0;
+
+            for (j = shift_start[c]; j > -shift_dec[c]; j -= shift_dec[c])
+            {
+               if (j > 0)
+                  value |= v << j;
+
+               else
+                  value |= v >> (-j);
+            }
+            *bp++ = (png_byte)((value >> 8) & 0xff);
+            *bp++ = (png_byte)(value & 0xff);
+         }
+      }
+   }
+}
+#endif
+
+#ifdef PNG_WRITE_SWAP_ALPHA_SUPPORTED
+void /* PRIVATE */
+png_do_write_swap_alpha(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_write_swap_alpha");
+
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This converts from ARGB to RGBA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               png_byte save = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = save;
+            }
+         }
+
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         else
+         {
+            /* This converts from AARRGGBB to RRGGBBAA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               png_byte save[2];
+               save[0] = *(sp++);
+               save[1] = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = save[0];
+               *(dp++) = save[1];
+            }
+         }
+#endif /* PNG_WRITE_16BIT_SUPPORTED */
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This converts from AG to GA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               png_byte save = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = save;
+            }
+         }
+
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         else
+         {
+            /* This converts from AAGG to GGAA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               png_byte save[2];
+               save[0] = *(sp++);
+               save[1] = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = save[0];
+               *(dp++) = save[1];
+            }
+         }
+#endif /* PNG_WRITE_16BIT_SUPPORTED */
+      }
+   }
+}
+#endif
+
+#ifdef PNG_WRITE_INVERT_ALPHA_SUPPORTED
+void /* PRIVATE */
+png_do_write_invert_alpha(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_write_invert_alpha");
+
+   {
+      if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This inverts the alpha channel in RGBA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               /* Does nothing
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               */
+               sp+=3; dp = sp;
+               *(dp++) = (png_byte)(255 - *(sp++));
+            }
+         }
+
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         else
+         {
+            /* This inverts the alpha channel in RRGGBBAA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               /* Does nothing
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               */
+               sp+=6; dp = sp;
+               *(dp++) = (png_byte)(255 - *(sp++));
+               *(dp++) = (png_byte)(255 - *(sp++));
+            }
+         }
+#endif /* PNG_WRITE_16BIT_SUPPORTED */
+      }
+
+      else if (row_info->color_type == PNG_COLOR_TYPE_GRAY_ALPHA)
+      {
+         if (row_info->bit_depth == 8)
+         {
+            /* This inverts the alpha channel in GA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               *(dp++) = *(sp++);
+               *(dp++) = (png_byte)(255 - *(sp++));
+            }
+         }
+
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         else
+         {
+            /* This inverts the alpha channel in GGAA */
+            png_bytep sp, dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            for (i = 0, sp = dp = row; i < row_width; i++)
+            {
+               /* Does nothing
+               *(dp++) = *(sp++);
+               *(dp++) = *(sp++);
+               */
+               sp+=2; dp = sp;
+               *(dp++) = (png_byte)(255 - *(sp++));
+               *(dp++) = (png_byte)(255 - *(sp++));
+            }
+         }
+#endif /* PNG_WRITE_16BIT_SUPPORTED */
+      }
+   }
+}
+#endif
+#endif /* PNG_WRITE_TRANSFORMS_SUPPORTED */
+
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+/* Undoes intrapixel differencing  */
+void /* PRIVATE */
+png_do_write_intrapixel(png_row_infop row_info, png_bytep row)
+{
+   png_debug(1, "in png_do_write_intrapixel");
+
+   if ((row_info->color_type & PNG_COLOR_MASK_COLOR))
+   {
+      int bytes_per_pixel;
+      png_uint_32 row_width = row_info->width;
+      if (row_info->bit_depth == 8)
+      {
+         png_bytep rp;
+         png_uint_32 i;
+
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+            bytes_per_pixel = 3;
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+            bytes_per_pixel = 4;
+
+         else
+            return;
+
+         for (i = 0, rp = row; i < row_width; i++, rp += bytes_per_pixel)
+         {
+            *(rp)     = (png_byte)((*rp       - *(rp + 1)) & 0xff);
+            *(rp + 2) = (png_byte)((*(rp + 2) - *(rp + 1)) & 0xff);
+         }
+      }
+
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+      else if (row_info->bit_depth == 16)
+      {
+         png_bytep rp;
+         png_uint_32 i;
+
+         if (row_info->color_type == PNG_COLOR_TYPE_RGB)
+            bytes_per_pixel = 6;
+
+         else if (row_info->color_type == PNG_COLOR_TYPE_RGB_ALPHA)
+            bytes_per_pixel = 8;
+
+         else
+            return;
+
+         for (i = 0, rp = row; i < row_width; i++, rp += bytes_per_pixel)
+         {
+            png_uint_32 s0   = (*(rp    ) << 8) | *(rp + 1);
+            png_uint_32 s1   = (*(rp + 2) << 8) | *(rp + 3);
+            png_uint_32 s2   = (*(rp + 4) << 8) | *(rp + 5);
+            png_uint_32 red  = (png_uint_32)((s0 - s1) & 0xffffL);
+            png_uint_32 blue = (png_uint_32)((s2 - s1) & 0xffffL);
+            *(rp    ) = (png_byte)((red >> 8) & 0xff);
+            *(rp + 1) = (png_byte)(red & 0xff);
+            *(rp + 4) = (png_byte)((blue >> 8) & 0xff);
+            *(rp + 5) = (png_byte)(blue & 0xff);
+         }
+      }
+#endif /* PNG_WRITE_16BIT_SUPPORTED */
+   }
+}
+#endif /* PNG_MNG_FEATURES_SUPPORTED */
+#endif /* PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/libpng/pngwutil.c b/ml/dlib/dlib/external/libpng/pngwutil.c
new file mode 100644
index 000000000..49e6a2d21
--- /dev/null
+++ b/ml/dlib/dlib/external/libpng/pngwutil.c
@@ -0,0 +1,3023 @@
+
+/* pngwutil.c - utilities to write a PNG file
+ *
+ * Last changed in libpng 1.6.2 [April 25, 2013]
+ * Copyright (c) 1998-2013 Glenn Randers-Pehrson
+ * (Version 0.96 Copyright (c) 1996, 1997 Andreas Dilger)
+ * (Version 0.88 Copyright (c) 1995, 1996 Guy Eric Schalnat, Group 42, Inc.)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "pngpriv.h"
+
+#ifdef PNG_WRITE_SUPPORTED
+
+#ifdef PNG_WRITE_INT_FUNCTIONS_SUPPORTED
+/* Place a 32-bit number into a buffer in PNG byte order.  We work
+ * with unsigned numbers for convenience, although one supported
+ * ancillary chunk uses signed (two's complement) numbers.
+ */
+void PNGAPI
+png_save_uint_32(png_bytep buf, png_uint_32 i)
+{
+   buf[0] = (png_byte)((i >> 24) & 0xff);
+   buf[1] = (png_byte)((i >> 16) & 0xff);
+   buf[2] = (png_byte)((i >> 8) & 0xff);
+   buf[3] = (png_byte)(i & 0xff);
+}
+
+/* Place a 16-bit number into a buffer in PNG byte order.
+ * The parameter is declared unsigned int, not png_uint_16,
+ * just to avoid potential problems on pre-ANSI C compilers.
+ */
+void PNGAPI
+png_save_uint_16(png_bytep buf, unsigned int i)
+{
+   buf[0] = (png_byte)((i >> 8) & 0xff);
+   buf[1] = (png_byte)(i & 0xff);
+}
+#endif
+
+/* Simple function to write the signature.  If we have already written
+ * the magic bytes of the signature, or more likely, the PNG stream is
+ * being embedded into another stream and doesn't need its own signature,
+ * we should call png_set_sig_bytes() to tell libpng how many of the
+ * bytes have already been written.
+ */
+void PNGAPI
+png_write_sig(png_structrp png_ptr)
+{
+   png_byte png_signature[8] = {137, 80, 78, 71, 13, 10, 26, 10};
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   /* Inform the I/O callback that the signature is being written */
+   png_ptr->io_state = PNG_IO_WRITING | PNG_IO_SIGNATURE;
+#endif
+
+   /* Write the rest of the 8 byte signature */
+   png_write_data(png_ptr, &png_signature[png_ptr->sig_bytes],
+      (png_size_t)(8 - png_ptr->sig_bytes));
+
+   if (png_ptr->sig_bytes < 3)
+      png_ptr->mode |= PNG_HAVE_PNG_SIGNATURE;
+}
+
+/* Write the start of a PNG chunk.  The type is the chunk type.
+ * The total_length is the sum of the lengths of all the data you will be
+ * passing in png_write_chunk_data().
+ */
+static void
+png_write_chunk_header(png_structrp png_ptr, png_uint_32 chunk_name,
+    png_uint_32 length)
+{
+   png_byte buf[8];
+
+#if defined(PNG_DEBUG) && (PNG_DEBUG > 0)
+   PNG_CSTRING_FROM_CHUNK(buf, chunk_name);
+   png_debug2(0, "Writing %s chunk, length = %lu", buf, (unsigned long)length);
+#endif
+
+   if (png_ptr == NULL)
+      return;
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   /* Inform the I/O callback that the chunk header is being written.
+    * PNG_IO_CHUNK_HDR requires a single I/O call.
+    */
+   png_ptr->io_state = PNG_IO_WRITING | PNG_IO_CHUNK_HDR;
+#endif
+
+   /* Write the length and the chunk name */
+   png_save_uint_32(buf, length);
+   png_save_uint_32(buf + 4, chunk_name);
+   png_write_data(png_ptr, buf, 8);
+
+   /* Put the chunk name into png_ptr->chunk_name */
+   png_ptr->chunk_name = chunk_name;
+
+   /* Reset the crc and run it over the chunk name */
+   png_reset_crc(png_ptr);
+
+   png_calculate_crc(png_ptr, buf + 4, 4);
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   /* Inform the I/O callback that chunk data will (possibly) be written.
+    * PNG_IO_CHUNK_DATA does NOT require a specific number of I/O calls.
+    */
+   png_ptr->io_state = PNG_IO_WRITING | PNG_IO_CHUNK_DATA;
+#endif
+}
+
+void PNGAPI
+png_write_chunk_start(png_structrp png_ptr, png_const_bytep chunk_string,
+    png_uint_32 length)
+{
+   png_write_chunk_header(png_ptr, PNG_CHUNK_FROM_STRING(chunk_string), length);
+}
+
+/* Write the data of a PNG chunk started with png_write_chunk_header().
+ * Note that multiple calls to this function are allowed, and that the
+ * sum of the lengths from these calls *must* add up to the total_length
+ * given to png_write_chunk_header().
+ */
+void PNGAPI
+png_write_chunk_data(png_structrp png_ptr, png_const_bytep data,
+    png_size_t length)
+{
+   /* Write the data, and run the CRC over it */
+   if (png_ptr == NULL)
+      return;
+
+   if (data != NULL && length > 0)
+   {
+      png_write_data(png_ptr, data, length);
+
+      /* Update the CRC after writing the data,
+       * in case that the user I/O routine alters it.
+       */
+      png_calculate_crc(png_ptr, data, length);
+   }
+}
+
+/* Finish a chunk started with png_write_chunk_header(). */
+void PNGAPI
+png_write_chunk_end(png_structrp png_ptr)
+{
+   png_byte buf[4];
+
+   if (png_ptr == NULL) return;
+
+#ifdef PNG_IO_STATE_SUPPORTED
+   /* Inform the I/O callback that the chunk CRC is being written.
+    * PNG_IO_CHUNK_CRC requires a single I/O function call.
+    */
+   png_ptr->io_state = PNG_IO_WRITING | PNG_IO_CHUNK_CRC;
+#endif
+
+   /* Write the crc in a single operation */
+   png_save_uint_32(buf, png_ptr->crc);
+
+   png_write_data(png_ptr, buf, (png_size_t)4);
+}
+
+/* Write a PNG chunk all at once.  The type is an array of ASCII characters
+ * representing the chunk name.  The array must be at least 4 bytes in
+ * length, and does not need to be null terminated.  To be safe, pass the
+ * pre-defined chunk names here, and if you need a new one, define it
+ * where the others are defined.  The length is the length of the data.
+ * All the data must be present.  If that is not possible, use the
+ * png_write_chunk_start(), png_write_chunk_data(), and png_write_chunk_end()
+ * functions instead.
+ */
+static void
+png_write_complete_chunk(png_structrp png_ptr, png_uint_32 chunk_name,
+   png_const_bytep data, png_size_t length)
+{
+   if (png_ptr == NULL)
+      return;
+
+   /* On 64 bit architectures 'length' may not fit in a png_uint_32. */
+   if (length > PNG_UINT_31_MAX)
+      png_error(png_ptr, "length exceeds PNG maxima");
+
+   png_write_chunk_header(png_ptr, chunk_name, (png_uint_32)length);
+   png_write_chunk_data(png_ptr, data, length);
+   png_write_chunk_end(png_ptr);
+}
+
+/* This is the API that calls the internal function above. */
+void PNGAPI
+png_write_chunk(png_structrp png_ptr, png_const_bytep chunk_string,
+   png_const_bytep data, png_size_t length)
+{
+   png_write_complete_chunk(png_ptr, PNG_CHUNK_FROM_STRING(chunk_string), data,
+      length);
+}
+
+/* This is used below to find the size of an image to pass to png_deflate_claim,
+ * so it only needs to be accurate if the size is less than 16384 bytes (the
+ * point at which a lower LZ window size can be used.)
+ */
+static png_alloc_size_t
+png_image_size(png_structrp png_ptr)
+{
+   /* Only return sizes up to the maximum of a png_uint_32, do this by limiting
+    * the width and height used to 15 bits.
+    */
+   png_uint_32 h = png_ptr->height;
+
+   if (png_ptr->rowbytes < 32768 && h < 32768)
+   {
+      if (png_ptr->interlaced)
+      {
+         /* Interlacing makes the image larger because of the replication of
+          * both the filter byte and the padding to a byte boundary.
+          */
+         png_uint_32 w = png_ptr->width;
+         unsigned int pd = png_ptr->pixel_depth;
+         png_alloc_size_t cb_base;
+         int pass;
+
+         for (cb_base=0, pass=0; pass<=6; ++pass)
+         {
+            png_uint_32 pw = PNG_PASS_COLS(w, pass);
+
+            if (pw > 0)
+               cb_base += (PNG_ROWBYTES(pd, pw)+1) * PNG_PASS_ROWS(h, pass);
+         }
+
+         return cb_base;
+      }
+
+      else
+         return (png_ptr->rowbytes+1) * h;
+   }
+
+   else
+      return 0xffffffffU;
+}
+
+#ifdef PNG_WRITE_OPTIMIZE_CMF_SUPPORTED
+   /* This is the code to hack the first two bytes of the deflate stream (the
+    * deflate header) to correct the windowBits value to match the actual data
+    * size.  Note that the second argument is the *uncompressed* size but the
+    * first argument is the *compressed* data (and it must be deflate
+    * compressed.)
+    */
+static void
+optimize_cmf(png_bytep data, png_alloc_size_t data_size)
+{
+   /* Optimize the CMF field in the zlib stream.  The resultant zlib stream is
+    * still compliant to the stream specification.
+    */
+   if (data_size <= 16384) /* else windowBits must be 15 */
+   {
+      unsigned int z_cmf = data[0];  /* zlib compression method and flags */
+
+      if ((z_cmf & 0x0f) == 8 && (z_cmf & 0xf0) <= 0x70)
+      {
+         unsigned int z_cinfo;
+         unsigned int half_z_window_size;
+
+         z_cinfo = z_cmf >> 4;
+         half_z_window_size = 1U << (z_cinfo + 7);
+
+         if (data_size <= half_z_window_size) /* else no change */
+         {
+            unsigned int tmp;
+
+            do
+            {
+               half_z_window_size >>= 1;
+               --z_cinfo;
+            }
+            while (z_cinfo > 0 && data_size <= half_z_window_size);
+
+            z_cmf = (z_cmf & 0x0f) | (z_cinfo << 4);
+
+            data[0] = (png_byte)z_cmf;
+            tmp = data[1] & 0xe0;
+            tmp += 0x1f - ((z_cmf << 8) + tmp) % 0x1f;
+            data[1] = (png_byte)tmp;
+         }
+      }
+   }
+}
+#else
+#  define optimize_cmf(dp,dl) ((void)0)
+#endif /* PNG_WRITE_OPTIMIZE_CMF_SUPPORTED */
+
+/* Initialize the compressor for the appropriate type of compression. */
+static int
+png_deflate_claim(png_structrp png_ptr, png_uint_32 owner,
+   png_alloc_size_t data_size)
+{
+   if (png_ptr->zowner != 0)
+   {
+      char msg[64];
+
+      PNG_STRING_FROM_CHUNK(msg, owner);
+      msg[4] = ':';
+      msg[5] = ' ';
+      PNG_STRING_FROM_CHUNK(msg+6, png_ptr->zowner);
+      /* So the message that results is "<chunk> using zstream"; this is an
+       * internal error, but is very useful for debugging.  i18n requirements
+       * are minimal.
+       */
+      (void)png_safecat(msg, (sizeof msg), 10, " using zstream");
+#     if PNG_LIBPNG_BUILD_BASE_TYPE >= PNG_LIBPNG_BUILD_RC
+         png_warning(png_ptr, msg);
+
+         /* Attempt sane error recovery */
+         if (png_ptr->zowner == png_IDAT) /* don't steal from IDAT */
+         {
+            png_ptr->zstream.msg = PNGZ_MSG_CAST("in use by IDAT");
+            return Z_STREAM_ERROR;
+         }
+
+         png_ptr->zowner = 0;
+#     else
+         png_error(png_ptr, msg);
+#     endif
+   }
+
+   {
+      int level = png_ptr->zlib_level;
+      int method = png_ptr->zlib_method;
+      int windowBits = png_ptr->zlib_window_bits;
+      int memLevel = png_ptr->zlib_mem_level;
+      int strategy; /* set below */
+      int ret; /* zlib return code */
+
+      if (owner == png_IDAT)
+      {
+         if (png_ptr->flags & PNG_FLAG_ZLIB_CUSTOM_STRATEGY)
+            strategy = png_ptr->zlib_strategy;
+
+         else if (png_ptr->do_filter != PNG_FILTER_NONE)
+            strategy = PNG_Z_DEFAULT_STRATEGY;
+
+         else
+            strategy = PNG_Z_DEFAULT_NOFILTER_STRATEGY;
+      }
+
+      else
+      {
+#        ifdef PNG_WRITE_CUSTOMIZE_ZTXT_COMPRESSION_SUPPORTED
+            level = png_ptr->zlib_text_level;
+            method = png_ptr->zlib_text_method;
+            windowBits = png_ptr->zlib_text_window_bits;
+            memLevel = png_ptr->zlib_text_mem_level;
+            strategy = png_ptr->zlib_text_strategy;
+#        else
+            /* If customization is not supported the values all come from the
+             * IDAT values except for the strategy, which is fixed to the
+             * default.  (This is the pre-1.6.0 behavior too, although it was
+             * implemented in a very different way.)
+             */
+            strategy = Z_DEFAULT_STRATEGY;
+#        endif
+      }
+
+      /* Adjust 'windowBits' down if larger than 'data_size'; to stop this
+       * happening just pass 32768 as the data_size parameter.  Notice that zlib
+       * requires an extra 262 bytes in the window in addition to the data to be
+       * able to see the whole of the data, so if data_size+262 takes us to the
+       * next windowBits size we need to fix up the value later.  (Because even
+       * though deflate needs the extra window, inflate does not!)
+       */
+      if (data_size <= 16384)
+      {
+         /* IMPLEMENTATION NOTE: this 'half_window_size' stuff is only here to
+          * work round a Microsoft Visual C misbehavior which, contrary to C-90,
+          * widens the result of the following shift to 64-bits if (and,
+          * apparently, only if) it is used in a test.
+          */
+         unsigned int half_window_size = 1U << (windowBits-1);
+
+         while (data_size + 262 <= half_window_size)
+         {
+            half_window_size >>= 1;
+            --windowBits;
+         }
+      }
+
+      /* Check against the previous initialized values, if any. */
+      if ((png_ptr->flags & PNG_FLAG_ZSTREAM_INITIALIZED) &&
+         (png_ptr->zlib_set_level != level ||
+         png_ptr->zlib_set_method != method ||
+         png_ptr->zlib_set_window_bits != windowBits ||
+         png_ptr->zlib_set_mem_level != memLevel ||
+         png_ptr->zlib_set_strategy != strategy))
+      {
+         if (deflateEnd(&png_ptr->zstream) != Z_OK)
+            png_warning(png_ptr, "deflateEnd failed (ignored)");
+
+         png_ptr->flags &= ~PNG_FLAG_ZSTREAM_INITIALIZED;
+      }
+
+      /* For safety clear out the input and output pointers (currently zlib
+       * doesn't use them on Init, but it might in the future).
+       */
+      png_ptr->zstream.next_in = NULL;
+      png_ptr->zstream.avail_in = 0;
+      png_ptr->zstream.next_out = NULL;
+      png_ptr->zstream.avail_out = 0;
+
+      /* Now initialize if required, setting the new parameters, otherwise just
+       * to a simple reset to the previous parameters.
+       */
+      if (png_ptr->flags & PNG_FLAG_ZSTREAM_INITIALIZED)
+         ret = deflateReset(&png_ptr->zstream);
+
+      else
+      {
+         ret = deflateInit2(&png_ptr->zstream, level, method, windowBits,
+            memLevel, strategy);
+
+         if (ret == Z_OK)
+            png_ptr->flags |= PNG_FLAG_ZSTREAM_INITIALIZED;
+      }
+
+      /* The return code is from either deflateReset or deflateInit2; they have
+       * pretty much the same set of error codes.
+       */
+      if (ret == Z_OK)
+         png_ptr->zowner = owner;
+
+      else
+         png_zstream_error(png_ptr, ret);
+
+      return ret;
+   }
+}
+
+/* Clean up (or trim) a linked list of compression buffers. */
+void /* PRIVATE */
+png_free_buffer_list(png_structrp png_ptr, png_compression_bufferp *listp)
+{
+   png_compression_bufferp list = *listp;
+
+   if (list != NULL)
+   {
+      *listp = NULL;
+
+      do
+      {
+         png_compression_bufferp next = list->next;
+
+         png_free(png_ptr, list);
+         list = next;
+      }
+      while (list != NULL);
+   }
+}
+
+#ifdef PNG_WRITE_COMPRESSED_TEXT_SUPPORTED
+/* This pair of functions encapsulates the operation of (a) compressing a
+ * text string, and (b) issuing it later as a series of chunk data writes.
+ * The compression_state structure is shared context for these functions
+ * set up by the caller to allow access to the relevant local variables.
+ *
+ * compression_buffer (new in 1.6.0) is just a linked list of zbuffer_size
+ * temporary buffers.  From 1.6.0 it is retained in png_struct so that it will
+ * be correctly freed in the event of a write error (previous implementations
+ * just leaked memory.)
+ */
+typedef struct
+{
+   png_const_bytep      input;        /* The uncompressed input data */
+   png_alloc_size_t     input_len;    /* Its length */
+   png_uint_32          output_len;   /* Final compressed length */
+   png_byte             output[1024]; /* First block of output */
+} compression_state;
+
+static void
+png_text_compress_init(compression_state *comp, png_const_bytep input,
+   png_alloc_size_t input_len)
+{
+   comp->input = input;
+   comp->input_len = input_len;
+   comp->output_len = 0;
+}
+
+/* Compress the data in the compression state input */
+static int
+png_text_compress(png_structrp png_ptr, png_uint_32 chunk_name,
+   compression_state *comp, png_uint_32 prefix_len)
+{
+   int ret;
+
+   /* To find the length of the output it is necessary to first compress the
+    * input, the result is buffered rather than using the two-pass algorithm
+    * that is used on the inflate side; deflate is assumed to be slower and a
+    * PNG writer is assumed to have more memory available than a PNG reader.
+    *
+    * IMPLEMENTATION NOTE: the zlib API deflateBound() can be used to find an
+    * upper limit on the output size, but it is always bigger than the input
+    * size so it is likely to be more efficient to use this linked-list
+    * approach.
+    */
+   ret = png_deflate_claim(png_ptr, chunk_name, comp->input_len);
+
+   if (ret != Z_OK)
+      return ret;
+
+   /* Set up the compression buffers, we need a loop here to avoid overflowing a
+    * uInt.  Use ZLIB_IO_MAX to limit the input.  The output is always limited
+    * by the output buffer size, so there is no need to check that.  Since this
+    * is ANSI-C we know that an 'int', hence a uInt, is always at least 16 bits
+    * in size.
+    */
+   {
+      png_compression_bufferp *end = &png_ptr->zbuffer_list;
+      png_alloc_size_t input_len = comp->input_len; /* may be zero! */
+      png_uint_32 output_len;
+
+      /* zlib updates these for us: */
+      png_ptr->zstream.next_in = PNGZ_INPUT_CAST(comp->input);
+      png_ptr->zstream.avail_in = 0; /* Set below */
+      png_ptr->zstream.next_out = comp->output;
+      png_ptr->zstream.avail_out = (sizeof comp->output);
+
+      output_len = png_ptr->zstream.avail_out;
+
+      do
+      {
+         uInt avail_in = ZLIB_IO_MAX;
+
+         if (avail_in > input_len)
+            avail_in = (uInt)input_len;
+
+         input_len -= avail_in;
+
+         png_ptr->zstream.avail_in = avail_in;
+
+         if (png_ptr->zstream.avail_out == 0)
+         {
+            png_compression_buffer *next;
+
+            /* Chunk data is limited to 2^31 bytes in length, so the prefix
+             * length must be counted here.
+             */
+            if (output_len + prefix_len > PNG_UINT_31_MAX)
+            {
+               ret = Z_MEM_ERROR;
+               break;
+            }
+
+            /* Need a new (malloc'ed) buffer, but there may be one present
+             * already.
+             */
+            next = *end;
+            if (next == NULL)
+            {
+               next = png_voidcast(png_compression_bufferp, png_malloc_base
+                  (png_ptr, PNG_COMPRESSION_BUFFER_SIZE(png_ptr)));
+
+               if (next == NULL)
+               {
+                  ret = Z_MEM_ERROR;
+                  break;
+               }
+
+               /* Link in this buffer (so that it will be freed later) */
+               next->next = NULL;
+               *end = next;
+            }
+
+            png_ptr->zstream.next_out = next->output;
+            png_ptr->zstream.avail_out = png_ptr->zbuffer_size;
+            output_len += png_ptr->zstream.avail_out;
+
+            /* Move 'end' to the next buffer pointer. */
+            end = &next->next;
+         }
+
+         /* Compress the data */
+         ret = deflate(&png_ptr->zstream,
+            input_len > 0 ? Z_NO_FLUSH : Z_FINISH);
+
+         /* Claw back input data that was not consumed (because avail_in is
+          * reset above every time round the loop).
+          */
+         input_len += png_ptr->zstream.avail_in;
+         png_ptr->zstream.avail_in = 0; /* safety */
+      }
+      while (ret == Z_OK);
+
+      /* There may be some space left in the last output buffer, this needs to
+       * be subtracted from output_len.
+       */
+      output_len -= png_ptr->zstream.avail_out;
+      png_ptr->zstream.avail_out = 0; /* safety */
+      comp->output_len = output_len;
+
+      /* Now double check the output length, put in a custom message if it is
+       * too long.  Otherwise ensure the z_stream::msg pointer is set to
+       * something.
+       */
+      if (output_len + prefix_len >= PNG_UINT_31_MAX)
+      {
+         png_ptr->zstream.msg = PNGZ_MSG_CAST("compressed data too long");
+         ret = Z_MEM_ERROR;
+      }
+
+      else
+         png_zstream_error(png_ptr, ret);
+
+      /* Reset zlib for another zTXt/iTXt or image data */
+      png_ptr->zowner = 0;
+
+      /* The only success case is Z_STREAM_END, input_len must be 0, if not this
+       * is an internal error.
+       */
+      if (ret == Z_STREAM_END && input_len == 0)
+      {
+         /* Fix up the deflate header, if required */
+         optimize_cmf(comp->output, comp->input_len);
+
+         /* But Z_OK is returned, not Z_STREAM_END; this allows the claim
+          * function above to return Z_STREAM_END on an error (though it never
+          * does in the current versions of zlib.)
+          */
+         return Z_OK;
+      }
+
+      else
+         return ret;
+   }
+}
+
+/* Ship the compressed text out via chunk writes */
+static void
+png_write_compressed_data_out(png_structrp png_ptr, compression_state *comp)
+{
+   png_uint_32 output_len = comp->output_len;
+   png_const_bytep output = comp->output;
+   png_uint_32 avail = (sizeof comp->output);
+   png_compression_buffer *next = png_ptr->zbuffer_list;
+
+   for (;;)
+   {
+      if (avail > output_len)
+         avail = output_len;
+
+      png_write_chunk_data(png_ptr, output, avail);
+
+      output_len -= avail;
+
+      if (output_len == 0 || next == NULL)
+         break;
+
+      avail = png_ptr->zbuffer_size;
+      output = next->output;
+      next = next->next;
+   }
+
+   /* This is an internal error; 'next' must have been NULL! */
+   if (output_len > 0)
+      png_error(png_ptr, "error writing ancillary chunked compressed data");
+}
+#endif /* PNG_WRITE_COMPRESSED_TEXT_SUPPORTED */
+
+#if defined(PNG_WRITE_TEXT_SUPPORTED) || defined(PNG_WRITE_pCAL_SUPPORTED) || \
+    defined(PNG_WRITE_iCCP_SUPPORTED) || defined(PNG_WRITE_sPLT_SUPPORTED)
+/* Check that the tEXt or zTXt keyword is valid per PNG 1.0 specification,
+ * and if invalid, correct the keyword rather than discarding the entire
+ * chunk.  The PNG 1.0 specification requires keywords 1-79 characters in
+ * length, forbids leading or trailing whitespace, multiple internal spaces,
+ * and the non-break space (0x80) from ISO 8859-1.  Returns keyword length.
+ *
+ * The 'new_key' buffer must be 80 characters in size (for the keyword plus a
+ * trailing '\0').  If this routine returns 0 then there was no keyword, or a
+ * valid one could not be generated, and the caller must png_error.
+ */
+static png_uint_32
+png_check_keyword(png_structrp png_ptr, png_const_charp key, png_bytep new_key)
+{
+   png_const_charp orig_key = key;
+   png_uint_32 key_len = 0;
+   int bad_character = 0;
+   int space = 1;
+
+   png_debug(1, "in png_check_keyword");
+
+   if (key == NULL)
+   {
+      *new_key = 0;
+      return 0;
+   }
+
+   while (*key && key_len < 79)
+   {
+      png_byte ch = (png_byte)(0xff & *key++);
+
+      if ((ch > 32 && ch <= 126) || (ch >= 161 /*&& ch <= 255*/))
+         *new_key++ = ch, ++key_len, space = 0;
+
+      else if (!space)
+      {
+         /* A space or an invalid character when one wasn't seen immediately
+          * before; output just a space.
+          */
+         *new_key++ = 32, ++key_len, space = 1;
+
+         /* If the character was not a space then it is invalid. */
+         if (ch != 32)
+            bad_character = ch;
+      }
+
+      else if (!bad_character)
+         bad_character = ch; /* just skip it, record the first error */
+   }
+
+   if (key_len > 0 && space) /* trailing space */
+   {
+      --key_len, --new_key;
+      if (!bad_character)
+         bad_character = 32;
+   }
+
+   /* Terminate the keyword */
+   *new_key = 0;
+
+   if (key_len == 0)
+      return 0;
+
+   /* Try to only output one warning per keyword: */
+   if (*key) /* keyword too long */
+      png_warning(png_ptr, "keyword truncated");
+
+   else if (bad_character)
+   {
+      PNG_WARNING_PARAMETERS(p)
+
+      png_warning_parameter(p, 1, orig_key);
+      png_warning_parameter_signed(p, 2, PNG_NUMBER_FORMAT_02x, bad_character);
+
+      png_formatted_warning(png_ptr, p, "keyword \"@1\": bad character '0x@2'");
+   }
+
+   return key_len;
+}
+#endif
+
+/* Write the IHDR chunk, and update the png_struct with the necessary
+ * information.  Note that the rest of this code depends upon this
+ * information being correct.
+ */
+void /* PRIVATE */
+png_write_IHDR(png_structrp png_ptr, png_uint_32 width, png_uint_32 height,
+    int bit_depth, int color_type, int compression_type, int filter_type,
+    int interlace_type)
+{
+   png_byte buf[13]; /* Buffer to store the IHDR info */
+
+   png_debug(1, "in png_write_IHDR");
+
+   /* Check that we have valid input data from the application info */
+   switch (color_type)
+   {
+      case PNG_COLOR_TYPE_GRAY:
+         switch (bit_depth)
+         {
+            case 1:
+            case 2:
+            case 4:
+            case 8:
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+            case 16:
+#endif
+               png_ptr->channels = 1; break;
+
+            default:
+               png_error(png_ptr,
+                   "Invalid bit depth for grayscale image");
+         }
+         break;
+
+      case PNG_COLOR_TYPE_RGB:
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         if (bit_depth != 8 && bit_depth != 16)
+#else
+         if (bit_depth != 8)
+#endif
+            png_error(png_ptr, "Invalid bit depth for RGB image");
+
+         png_ptr->channels = 3;
+         break;
+
+      case PNG_COLOR_TYPE_PALETTE:
+         switch (bit_depth)
+         {
+            case 1:
+            case 2:
+            case 4:
+            case 8:
+               png_ptr->channels = 1;
+               break;
+
+            default:
+               png_error(png_ptr, "Invalid bit depth for paletted image");
+         }
+         break;
+
+      case PNG_COLOR_TYPE_GRAY_ALPHA:
+         if (bit_depth != 8 && bit_depth != 16)
+            png_error(png_ptr, "Invalid bit depth for grayscale+alpha image");
+
+         png_ptr->channels = 2;
+         break;
+
+      case PNG_COLOR_TYPE_RGB_ALPHA:
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+         if (bit_depth != 8 && bit_depth != 16)
+#else
+         if (bit_depth != 8)
+#endif
+            png_error(png_ptr, "Invalid bit depth for RGBA image");
+
+         png_ptr->channels = 4;
+         break;
+
+      default:
+         png_error(png_ptr, "Invalid image color type specified");
+   }
+
+   if (compression_type != PNG_COMPRESSION_TYPE_BASE)
+   {
+      png_warning(png_ptr, "Invalid compression type specified");
+      compression_type = PNG_COMPRESSION_TYPE_BASE;
+   }
+
+   /* Write filter_method 64 (intrapixel differencing) only if
+    * 1. Libpng was compiled with PNG_MNG_FEATURES_SUPPORTED and
+    * 2. Libpng did not write a PNG signature (this filter_method is only
+    *    used in PNG datastreams that are embedded in MNG datastreams) and
+    * 3. The application called png_permit_mng_features with a mask that
+    *    included PNG_FLAG_MNG_FILTER_64 and
+    * 4. The filter_method is 64 and
+    * 5. The color_type is RGB or RGBA
+    */
+   if (
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+       !((png_ptr->mng_features_permitted & PNG_FLAG_MNG_FILTER_64) &&
+       ((png_ptr->mode&PNG_HAVE_PNG_SIGNATURE) == 0) &&
+       (color_type == PNG_COLOR_TYPE_RGB ||
+        color_type == PNG_COLOR_TYPE_RGB_ALPHA) &&
+       (filter_type == PNG_INTRAPIXEL_DIFFERENCING)) &&
+#endif
+       filter_type != PNG_FILTER_TYPE_BASE)
+   {
+      png_warning(png_ptr, "Invalid filter type specified");
+      filter_type = PNG_FILTER_TYPE_BASE;
+   }
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   if (interlace_type != PNG_INTERLACE_NONE &&
+       interlace_type != PNG_INTERLACE_ADAM7)
+   {
+      png_warning(png_ptr, "Invalid interlace type specified");
+      interlace_type = PNG_INTERLACE_ADAM7;
+   }
+#else
+   interlace_type=PNG_INTERLACE_NONE;
+#endif
+
+   /* Save the relevent information */
+   png_ptr->bit_depth = (png_byte)bit_depth;
+   png_ptr->color_type = (png_byte)color_type;
+   png_ptr->interlaced = (png_byte)interlace_type;
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+   png_ptr->filter_type = (png_byte)filter_type;
+#endif
+   png_ptr->compression_type = (png_byte)compression_type;
+   png_ptr->width = width;
+   png_ptr->height = height;
+
+   png_ptr->pixel_depth = (png_byte)(bit_depth * png_ptr->channels);
+   png_ptr->rowbytes = PNG_ROWBYTES(png_ptr->pixel_depth, width);
+   /* Set the usr info, so any transformations can modify it */
+   png_ptr->usr_width = png_ptr->width;
+   png_ptr->usr_bit_depth = png_ptr->bit_depth;
+   png_ptr->usr_channels = png_ptr->channels;
+
+   /* Pack the header information into the buffer */
+   png_save_uint_32(buf, width);
+   png_save_uint_32(buf + 4, height);
+   buf[8] = (png_byte)bit_depth;
+   buf[9] = (png_byte)color_type;
+   buf[10] = (png_byte)compression_type;
+   buf[11] = (png_byte)filter_type;
+   buf[12] = (png_byte)interlace_type;
+
+   /* Write the chunk */
+   png_write_complete_chunk(png_ptr, png_IHDR, buf, (png_size_t)13);
+
+   if (!(png_ptr->do_filter))
+   {
+      if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE ||
+          png_ptr->bit_depth < 8)
+         png_ptr->do_filter = PNG_FILTER_NONE;
+
+      else
+         png_ptr->do_filter = PNG_ALL_FILTERS;
+   }
+
+   png_ptr->mode = PNG_HAVE_IHDR; /* not READY_FOR_ZTXT */
+}
+
+/* Write the palette.  We are careful not to trust png_color to be in the
+ * correct order for PNG, so people can redefine it to any convenient
+ * structure.
+ */
+void /* PRIVATE */
+png_write_PLTE(png_structrp png_ptr, png_const_colorp palette,
+    png_uint_32 num_pal)
+{
+   png_uint_32 i;
+   png_const_colorp pal_ptr;
+   png_byte buf[3];
+
+   png_debug(1, "in png_write_PLTE");
+
+   if ((
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+       !(png_ptr->mng_features_permitted & PNG_FLAG_MNG_EMPTY_PLTE) &&
+#endif
+       num_pal == 0) || num_pal > 256)
+   {
+      if (png_ptr->color_type == PNG_COLOR_TYPE_PALETTE)
+      {
+         png_error(png_ptr, "Invalid number of colors in palette");
+      }
+
+      else
+      {
+         png_warning(png_ptr, "Invalid number of colors in palette");
+         return;
+      }
+   }
+
+   if (!(png_ptr->color_type&PNG_COLOR_MASK_COLOR))
+   {
+      png_warning(png_ptr,
+          "Ignoring request to write a PLTE chunk in grayscale PNG");
+
+      return;
+   }
+
+   png_ptr->num_palette = (png_uint_16)num_pal;
+   png_debug1(3, "num_palette = %d", png_ptr->num_palette);
+
+   png_write_chunk_header(png_ptr, png_PLTE, (png_uint_32)(num_pal * 3));
+#ifdef PNG_POINTER_INDEXING_SUPPORTED
+
+   for (i = 0, pal_ptr = palette; i < num_pal; i++, pal_ptr++)
+   {
+      buf[0] = pal_ptr->red;
+      buf[1] = pal_ptr->green;
+      buf[2] = pal_ptr->blue;
+      png_write_chunk_data(png_ptr, buf, (png_size_t)3);
+   }
+
+#else
+   /* This is a little slower but some buggy compilers need to do this
+    * instead
+    */
+   pal_ptr=palette;
+
+   for (i = 0; i < num_pal; i++)
+   {
+      buf[0] = pal_ptr[i].red;
+      buf[1] = pal_ptr[i].green;
+      buf[2] = pal_ptr[i].blue;
+      png_write_chunk_data(png_ptr, buf, (png_size_t)3);
+   }
+
+#endif
+   png_write_chunk_end(png_ptr);
+   png_ptr->mode |= PNG_HAVE_PLTE;
+}
+
+/* This is similar to png_text_compress, above, except that it does not require
+ * all of the data at once and, instead of buffering the compressed result,
+ * writes it as IDAT chunks.  Unlike png_text_compress it *can* png_error out
+ * because it calls the write interface.  As a result it does its own error
+ * reporting and does not return an error code.  In the event of error it will
+ * just call png_error.  The input data length may exceed 32-bits.  The 'flush'
+ * parameter is exactly the same as that to deflate, with the following
+ * meanings:
+ *
+ * Z_NO_FLUSH: normal incremental output of compressed data
+ * Z_SYNC_FLUSH: do a SYNC_FLUSH, used by png_write_flush
+ * Z_FINISH: this is the end of the input, do a Z_FINISH and clean up
+ *
+ * The routine manages the acquire and release of the png_ptr->zstream by
+ * checking and (at the end) clearing png_ptr->zowner, it does some sanity
+ * checks on the 'mode' flags while doing this.
+ */
+void /* PRIVATE */
+png_compress_IDAT(png_structrp png_ptr, png_const_bytep input,
+   png_alloc_size_t input_len, int flush)
+{
+   if (png_ptr->zowner != png_IDAT)
+   {
+      /* First time.   Ensure we have a temporary buffer for compression and
+       * trim the buffer list if it has more than one entry to free memory.
+       * If 'WRITE_COMPRESSED_TEXT' is not set the list will never have been
+       * created at this point, but the check here is quick and safe.
+       */
+      if (png_ptr->zbuffer_list == NULL)
+      {
+         png_ptr->zbuffer_list = png_voidcast(png_compression_bufferp,
+            png_malloc(png_ptr, PNG_COMPRESSION_BUFFER_SIZE(png_ptr)));
+         png_ptr->zbuffer_list->next = NULL;
+      }
+
+      else
+         png_free_buffer_list(png_ptr, &png_ptr->zbuffer_list->next);
+
+      /* It is a terminal error if we can't claim the zstream. */
+      if (png_deflate_claim(png_ptr, png_IDAT, png_image_size(png_ptr)) != Z_OK)
+         png_error(png_ptr, png_ptr->zstream.msg);
+
+      /* The output state is maintained in png_ptr->zstream, so it must be
+       * initialized here after the claim.
+       */
+      png_ptr->zstream.next_out = png_ptr->zbuffer_list->output;
+      png_ptr->zstream.avail_out = png_ptr->zbuffer_size;
+   }
+
+   /* Now loop reading and writing until all the input is consumed or an error
+    * terminates the operation.  The _out values are maintained across calls to
+    * this function, but the input must be reset each time.
+    */
+   png_ptr->zstream.next_in = PNGZ_INPUT_CAST(input);
+   png_ptr->zstream.avail_in = 0; /* set below */
+   for (;;)
+   {
+      int ret;
+
+      /* INPUT: from the row data */
+      uInt avail = ZLIB_IO_MAX;
+
+      if (avail > input_len)
+         avail = (uInt)input_len; /* safe because of the check */
+
+      png_ptr->zstream.avail_in = avail;
+      input_len -= avail;
+
+      ret = deflate(&png_ptr->zstream, input_len > 0 ? Z_NO_FLUSH : flush);
+
+      /* Include as-yet unconsumed input */
+      input_len += png_ptr->zstream.avail_in;
+      png_ptr->zstream.avail_in = 0;
+
+      /* OUTPUT: write complete IDAT chunks when avail_out drops to zero, note
+       * that these two zstream fields are preserved across the calls, therefore
+       * there is no need to set these up on entry to the loop.
+       */
+      if (png_ptr->zstream.avail_out == 0)
+      {
+         png_bytep data = png_ptr->zbuffer_list->output;
+         uInt size = png_ptr->zbuffer_size;
+
+         /* Write an IDAT containing the data then reset the buffer.  The
+          * first IDAT may need deflate header optimization.
+          */
+#        ifdef PNG_WRITE_OPTIMIZE_CMF_SUPPORTED
+            if (!(png_ptr->mode & PNG_HAVE_IDAT) &&
+               png_ptr->compression_type == PNG_COMPRESSION_TYPE_BASE)
+               optimize_cmf(data, png_image_size(png_ptr));
+#        endif
+
+         png_write_complete_chunk(png_ptr, png_IDAT, data, size);
+         png_ptr->mode |= PNG_HAVE_IDAT;
+
+         png_ptr->zstream.next_out = data;
+         png_ptr->zstream.avail_out = size;
+
+         /* For SYNC_FLUSH or FINISH it is essential to keep calling zlib with
+          * the same flush parameter until it has finished output, for NO_FLUSH
+          * it doesn't matter.
+          */
+         if (ret == Z_OK && flush != Z_NO_FLUSH)
+            continue;
+      }
+
+      /* The order of these checks doesn't matter much; it just effect which
+       * possible error might be detected if multiple things go wrong at once.
+       */
+      if (ret == Z_OK) /* most likely return code! */
+      {
+         /* If all the input has been consumed then just return.  If Z_FINISH
+          * was used as the flush parameter something has gone wrong if we get
+          * here.
+          */
+         if (input_len == 0)
+         {
+            if (flush == Z_FINISH)
+               png_error(png_ptr, "Z_OK on Z_FINISH with output space");
+
+            return;
+         }
+      }
+
+      else if (ret == Z_STREAM_END && flush == Z_FINISH)
+      {
+         /* This is the end of the IDAT data; any pending output must be
+          * flushed.  For small PNG files we may still be at the beginning.
+          */
+         png_bytep data = png_ptr->zbuffer_list->output;
+         uInt size = png_ptr->zbuffer_size - png_ptr->zstream.avail_out;
+
+#        ifdef PNG_WRITE_OPTIMIZE_CMF_SUPPORTED
+            if (!(png_ptr->mode & PNG_HAVE_IDAT) &&
+               png_ptr->compression_type == PNG_COMPRESSION_TYPE_BASE)
+               optimize_cmf(data, png_image_size(png_ptr));
+#        endif
+
+         png_write_complete_chunk(png_ptr, png_IDAT, data, size);
+         png_ptr->zstream.avail_out = 0;
+         png_ptr->zstream.next_out = NULL;
+         png_ptr->mode |= PNG_HAVE_IDAT | PNG_AFTER_IDAT;
+
+         png_ptr->zowner = 0; /* Release the stream */
+         return;
+      }
+
+      else
+      {
+         /* This is an error condition. */
+         png_zstream_error(png_ptr, ret);
+         png_error(png_ptr, png_ptr->zstream.msg);
+      }
+   }
+}
+
+/* Write an IEND chunk */
+void /* PRIVATE */
+png_write_IEND(png_structrp png_ptr)
+{
+   png_debug(1, "in png_write_IEND");
+
+   png_write_complete_chunk(png_ptr, png_IEND, NULL, (png_size_t)0);
+   png_ptr->mode |= PNG_HAVE_IEND;
+}
+
+#ifdef PNG_WRITE_gAMA_SUPPORTED
+/* Write a gAMA chunk */
+void /* PRIVATE */
+png_write_gAMA_fixed(png_structrp png_ptr, png_fixed_point file_gamma)
+{
+   png_byte buf[4];
+
+   png_debug(1, "in png_write_gAMA");
+
+   /* file_gamma is saved in 1/100,000ths */
+   png_save_uint_32(buf, (png_uint_32)file_gamma);
+   png_write_complete_chunk(png_ptr, png_gAMA, buf, (png_size_t)4);
+}
+#endif
+
+#ifdef PNG_WRITE_sRGB_SUPPORTED
+/* Write a sRGB chunk */
+void /* PRIVATE */
+png_write_sRGB(png_structrp png_ptr, int srgb_intent)
+{
+   png_byte buf[1];
+
+   png_debug(1, "in png_write_sRGB");
+
+   if (srgb_intent >= PNG_sRGB_INTENT_LAST)
+      png_warning(png_ptr,
+          "Invalid sRGB rendering intent specified");
+
+   buf[0]=(png_byte)srgb_intent;
+   png_write_complete_chunk(png_ptr, png_sRGB, buf, (png_size_t)1);
+}
+#endif
+
+#ifdef PNG_WRITE_iCCP_SUPPORTED
+/* Write an iCCP chunk */
+void /* PRIVATE */
+png_write_iCCP(png_structrp png_ptr, png_const_charp name,
+    png_const_bytep profile)
+{
+   png_uint_32 name_len;
+   png_uint_32 profile_len;
+   png_byte new_name[81]; /* 1 byte for the compression byte */
+   compression_state comp;
+
+   png_debug(1, "in png_write_iCCP");
+
+   /* These are all internal problems: the profile should have been checked
+    * before when it was stored.
+    */
+   if (profile == NULL)
+      png_error(png_ptr, "No profile for iCCP chunk"); /* internal error */
+
+   profile_len = png_get_uint_32(profile);
+
+   if (profile_len < 132)
+      png_error(png_ptr, "ICC profile too short");
+
+   if (profile_len & 0x03)
+      png_error(png_ptr, "ICC profile length invalid (not a multiple of 4)");
+
+   {
+      png_uint_32 embedded_profile_len = png_get_uint_32(profile);
+
+      if (profile_len != embedded_profile_len)
+         png_error(png_ptr, "Profile length does not match profile");
+   }
+
+   name_len = png_check_keyword(png_ptr, name, new_name);
+
+   if (name_len == 0)
+      png_error(png_ptr, "iCCP: invalid keyword");
+
+   new_name[++name_len] = PNG_COMPRESSION_TYPE_BASE;
+
+   /* Make sure we include the NULL after the name and the compression type */
+   ++name_len;
+
+   png_text_compress_init(&comp, profile, profile_len);
+
+   /* Allow for keyword terminator and compression byte */
+   if (png_text_compress(png_ptr, png_iCCP, &comp, name_len) != Z_OK)
+      png_error(png_ptr, png_ptr->zstream.msg);
+
+   png_write_chunk_header(png_ptr, png_iCCP, name_len + comp.output_len);
+
+   png_write_chunk_data(png_ptr, new_name, name_len);
+
+   png_write_compressed_data_out(png_ptr, &comp);
+
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_sPLT_SUPPORTED
+/* Write a sPLT chunk */
+void /* PRIVATE */
+png_write_sPLT(png_structrp png_ptr, png_const_sPLT_tp spalette)
+{
+   png_uint_32 name_len;
+   png_byte new_name[80];
+   png_byte entrybuf[10];
+   png_size_t entry_size = (spalette->depth == 8 ? 6 : 10);
+   png_size_t palette_size = entry_size * spalette->nentries;
+   png_sPLT_entryp ep;
+#ifndef PNG_POINTER_INDEXING_SUPPORTED
+   int i;
+#endif
+
+   png_debug(1, "in png_write_sPLT");
+
+   name_len = png_check_keyword(png_ptr, spalette->name, new_name);
+
+   if (name_len == 0)
+      png_error(png_ptr, "sPLT: invalid keyword");
+
+   /* Make sure we include the NULL after the name */
+   png_write_chunk_header(png_ptr, png_sPLT,
+       (png_uint_32)(name_len + 2 + palette_size));
+
+   png_write_chunk_data(png_ptr, (png_bytep)new_name,
+       (png_size_t)(name_len + 1));
+
+   png_write_chunk_data(png_ptr, &spalette->depth, (png_size_t)1);
+
+   /* Loop through each palette entry, writing appropriately */
+#ifdef PNG_POINTER_INDEXING_SUPPORTED
+   for (ep = spalette->entries; ep<spalette->entries + spalette->nentries; ep++)
+   {
+      if (spalette->depth == 8)
+      {
+         entrybuf[0] = (png_byte)ep->red;
+         entrybuf[1] = (png_byte)ep->green;
+         entrybuf[2] = (png_byte)ep->blue;
+         entrybuf[3] = (png_byte)ep->alpha;
+         png_save_uint_16(entrybuf + 4, ep->frequency);
+      }
+
+      else
+      {
+         png_save_uint_16(entrybuf + 0, ep->red);
+         png_save_uint_16(entrybuf + 2, ep->green);
+         png_save_uint_16(entrybuf + 4, ep->blue);
+         png_save_uint_16(entrybuf + 6, ep->alpha);
+         png_save_uint_16(entrybuf + 8, ep->frequency);
+      }
+
+      png_write_chunk_data(png_ptr, entrybuf, entry_size);
+   }
+#else
+   ep=spalette->entries;
+   for (i = 0; i>spalette->nentries; i++)
+   {
+      if (spalette->depth == 8)
+      {
+         entrybuf[0] = (png_byte)ep[i].red;
+         entrybuf[1] = (png_byte)ep[i].green;
+         entrybuf[2] = (png_byte)ep[i].blue;
+         entrybuf[3] = (png_byte)ep[i].alpha;
+         png_save_uint_16(entrybuf + 4, ep[i].frequency);
+      }
+
+      else
+      {
+         png_save_uint_16(entrybuf + 0, ep[i].red);
+         png_save_uint_16(entrybuf + 2, ep[i].green);
+         png_save_uint_16(entrybuf + 4, ep[i].blue);
+         png_save_uint_16(entrybuf + 6, ep[i].alpha);
+         png_save_uint_16(entrybuf + 8, ep[i].frequency);
+      }
+
+      png_write_chunk_data(png_ptr, entrybuf, entry_size);
+   }
+#endif
+
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_sBIT_SUPPORTED
+/* Write the sBIT chunk */
+void /* PRIVATE */
+png_write_sBIT(png_structrp png_ptr, png_const_color_8p sbit, int color_type)
+{
+   png_byte buf[4];
+   png_size_t size;
+
+   png_debug(1, "in png_write_sBIT");
+
+   /* Make sure we don't depend upon the order of PNG_COLOR_8 */
+   if (color_type & PNG_COLOR_MASK_COLOR)
+   {
+      png_byte maxbits;
+
+      maxbits = (png_byte)(color_type==PNG_COLOR_TYPE_PALETTE ? 8 :
+          png_ptr->usr_bit_depth);
+
+      if (sbit->red == 0 || sbit->red > maxbits ||
+          sbit->green == 0 || sbit->green > maxbits ||
+          sbit->blue == 0 || sbit->blue > maxbits)
+      {
+         png_warning(png_ptr, "Invalid sBIT depth specified");
+         return;
+      }
+
+      buf[0] = sbit->red;
+      buf[1] = sbit->green;
+      buf[2] = sbit->blue;
+      size = 3;
+   }
+
+   else
+   {
+      if (sbit->gray == 0 || sbit->gray > png_ptr->usr_bit_depth)
+      {
+         png_warning(png_ptr, "Invalid sBIT depth specified");
+         return;
+      }
+
+      buf[0] = sbit->gray;
+      size = 1;
+   }
+
+   if (color_type & PNG_COLOR_MASK_ALPHA)
+   {
+      if (sbit->alpha == 0 || sbit->alpha > png_ptr->usr_bit_depth)
+      {
+         png_warning(png_ptr, "Invalid sBIT depth specified");
+         return;
+      }
+
+      buf[size++] = sbit->alpha;
+   }
+
+   png_write_complete_chunk(png_ptr, png_sBIT, buf, size);
+}
+#endif
+
+#ifdef PNG_WRITE_cHRM_SUPPORTED
+/* Write the cHRM chunk */
+void /* PRIVATE */
+png_write_cHRM_fixed(png_structrp png_ptr, const png_xy *xy)
+{
+   png_byte buf[32];
+
+   png_debug(1, "in png_write_cHRM");
+
+   /* Each value is saved in 1/100,000ths */
+   png_save_int_32(buf,      xy->whitex);
+   png_save_int_32(buf +  4, xy->whitey);
+
+   png_save_int_32(buf +  8, xy->redx);
+   png_save_int_32(buf + 12, xy->redy);
+
+   png_save_int_32(buf + 16, xy->greenx);
+   png_save_int_32(buf + 20, xy->greeny);
+
+   png_save_int_32(buf + 24, xy->bluex);
+   png_save_int_32(buf + 28, xy->bluey);
+
+   png_write_complete_chunk(png_ptr, png_cHRM, buf, 32);
+}
+#endif
+
+#ifdef PNG_WRITE_tRNS_SUPPORTED
+/* Write the tRNS chunk */
+void /* PRIVATE */
+png_write_tRNS(png_structrp png_ptr, png_const_bytep trans_alpha,
+    png_const_color_16p tran, int num_trans, int color_type)
+{
+   png_byte buf[6];
+
+   png_debug(1, "in png_write_tRNS");
+
+   if (color_type == PNG_COLOR_TYPE_PALETTE)
+   {
+      if (num_trans <= 0 || num_trans > (int)png_ptr->num_palette)
+      {
+         png_app_warning(png_ptr,
+             "Invalid number of transparent colors specified");
+         return;
+      }
+
+      /* Write the chunk out as it is */
+      png_write_complete_chunk(png_ptr, png_tRNS, trans_alpha,
+         (png_size_t)num_trans);
+   }
+
+   else if (color_type == PNG_COLOR_TYPE_GRAY)
+   {
+      /* One 16 bit value */
+      if (tran->gray >= (1 << png_ptr->bit_depth))
+      {
+         png_app_warning(png_ptr,
+             "Ignoring attempt to write tRNS chunk out-of-range for bit_depth");
+
+         return;
+      }
+
+      png_save_uint_16(buf, tran->gray);
+      png_write_complete_chunk(png_ptr, png_tRNS, buf, (png_size_t)2);
+   }
+
+   else if (color_type == PNG_COLOR_TYPE_RGB)
+   {
+      /* Three 16 bit values */
+      png_save_uint_16(buf, tran->red);
+      png_save_uint_16(buf + 2, tran->green);
+      png_save_uint_16(buf + 4, tran->blue);
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+      if (png_ptr->bit_depth == 8 && (buf[0] | buf[2] | buf[4]))
+#else
+      if (buf[0] | buf[2] | buf[4])
+#endif
+      {
+         png_app_warning(png_ptr,
+           "Ignoring attempt to write 16-bit tRNS chunk when bit_depth is 8");
+         return;
+      }
+
+      png_write_complete_chunk(png_ptr, png_tRNS, buf, (png_size_t)6);
+   }
+
+   else
+   {
+      png_app_warning(png_ptr, "Can't write tRNS with an alpha channel");
+   }
+}
+#endif
+
+#ifdef PNG_WRITE_bKGD_SUPPORTED
+/* Write the background chunk */
+void /* PRIVATE */
+png_write_bKGD(png_structrp png_ptr, png_const_color_16p back, int color_type)
+{
+   png_byte buf[6];
+
+   png_debug(1, "in png_write_bKGD");
+
+   if (color_type == PNG_COLOR_TYPE_PALETTE)
+   {
+      if (
+#ifdef PNG_MNG_FEATURES_SUPPORTED
+          (png_ptr->num_palette ||
+          (!(png_ptr->mng_features_permitted & PNG_FLAG_MNG_EMPTY_PLTE))) &&
+#endif
+         back->index >= png_ptr->num_palette)
+      {
+         png_warning(png_ptr, "Invalid background palette index");
+         return;
+      }
+
+      buf[0] = back->index;
+      png_write_complete_chunk(png_ptr, png_bKGD, buf, (png_size_t)1);
+   }
+
+   else if (color_type & PNG_COLOR_MASK_COLOR)
+   {
+      png_save_uint_16(buf, back->red);
+      png_save_uint_16(buf + 2, back->green);
+      png_save_uint_16(buf + 4, back->blue);
+#ifdef PNG_WRITE_16BIT_SUPPORTED
+      if (png_ptr->bit_depth == 8 && (buf[0] | buf[2] | buf[4]))
+#else
+      if (buf[0] | buf[2] | buf[4])
+#endif
+      {
+         png_warning(png_ptr,
+             "Ignoring attempt to write 16-bit bKGD chunk when bit_depth is 8");
+
+         return;
+      }
+
+      png_write_complete_chunk(png_ptr, png_bKGD, buf, (png_size_t)6);
+   }
+
+   else
+   {
+      if (back->gray >= (1 << png_ptr->bit_depth))
+      {
+         png_warning(png_ptr,
+             "Ignoring attempt to write bKGD chunk out-of-range for bit_depth");
+
+         return;
+      }
+
+      png_save_uint_16(buf, back->gray);
+      png_write_complete_chunk(png_ptr, png_bKGD, buf, (png_size_t)2);
+   }
+}
+#endif
+
+#ifdef PNG_WRITE_hIST_SUPPORTED
+/* Write the histogram */
+void /* PRIVATE */
+png_write_hIST(png_structrp png_ptr, png_const_uint_16p hist, int num_hist)
+{
+   int i;
+   png_byte buf[3];
+
+   png_debug(1, "in png_write_hIST");
+
+   if (num_hist > (int)png_ptr->num_palette)
+   {
+      png_debug2(3, "num_hist = %d, num_palette = %d", num_hist,
+          png_ptr->num_palette);
+
+      png_warning(png_ptr, "Invalid number of histogram entries specified");
+      return;
+   }
+
+   png_write_chunk_header(png_ptr, png_hIST, (png_uint_32)(num_hist * 2));
+
+   for (i = 0; i < num_hist; i++)
+   {
+      png_save_uint_16(buf, hist[i]);
+      png_write_chunk_data(png_ptr, buf, (png_size_t)2);
+   }
+
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_tEXt_SUPPORTED
+/* Write a tEXt chunk */
+void /* PRIVATE */
+png_write_tEXt(png_structrp png_ptr, png_const_charp key, png_const_charp text,
+    png_size_t text_len)
+{
+   png_uint_32 key_len;
+   png_byte new_key[80];
+
+   png_debug(1, "in png_write_tEXt");
+
+   key_len = png_check_keyword(png_ptr, key, new_key);
+
+   if (key_len == 0)
+      png_error(png_ptr, "tEXt: invalid keyword");
+
+   if (text == NULL || *text == '\0')
+      text_len = 0;
+
+   else
+      text_len = strlen(text);
+
+   if (text_len > PNG_UINT_31_MAX - (key_len+1))
+      png_error(png_ptr, "tEXt: text too long");
+
+   /* Make sure we include the 0 after the key */
+   png_write_chunk_header(png_ptr, png_tEXt,
+       (png_uint_32)/*checked above*/(key_len + text_len + 1));
+   /*
+    * We leave it to the application to meet PNG-1.0 requirements on the
+    * contents of the text.  PNG-1.0 through PNG-1.2 discourage the use of
+    * any non-Latin-1 characters except for NEWLINE.  ISO PNG will forbid them.
+    * The NUL character is forbidden by PNG-1.0 through PNG-1.2 and ISO PNG.
+    */
+   png_write_chunk_data(png_ptr, new_key, key_len + 1);
+
+   if (text_len)
+      png_write_chunk_data(png_ptr, (png_const_bytep)text, text_len);
+
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_zTXt_SUPPORTED
+/* Write a compressed text chunk */
+void /* PRIVATE */
+png_write_zTXt(png_structrp png_ptr, png_const_charp key, png_const_charp text,
+    png_size_t text_len, int compression)
+{
+   png_uint_32 key_len;
+   png_byte new_key[81];
+   compression_state comp;
+
+   png_debug(1, "in png_write_zTXt");
+   PNG_UNUSED(text_len) /* Always use strlen */
+
+   if (compression == PNG_TEXT_COMPRESSION_NONE)
+   {
+      png_write_tEXt(png_ptr, key, text, 0);
+      return;
+   }
+
+   if (compression != PNG_TEXT_COMPRESSION_zTXt)
+      png_error(png_ptr, "zTXt: invalid compression type");
+
+   key_len = png_check_keyword(png_ptr, key, new_key);
+
+   if (key_len == 0)
+      png_error(png_ptr, "zTXt: invalid keyword");
+
+   /* Add the compression method and 1 for the keyword separator. */
+   new_key[++key_len] = PNG_COMPRESSION_TYPE_BASE;
+   ++key_len;
+
+   /* Compute the compressed data; do it now for the length */
+   png_text_compress_init(&comp, (png_const_bytep)text,
+      text == NULL ? 0 : strlen(text));
+
+   if (png_text_compress(png_ptr, png_zTXt, &comp, key_len) != Z_OK)
+      png_error(png_ptr, png_ptr->zstream.msg);
+
+   /* Write start of chunk */
+   png_write_chunk_header(png_ptr, png_zTXt, key_len + comp.output_len);
+
+   /* Write key */
+   png_write_chunk_data(png_ptr, new_key, key_len);
+
+   /* Write the compressed data */
+   png_write_compressed_data_out(png_ptr, &comp);
+
+   /* Close the chunk */
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_iTXt_SUPPORTED
+/* Write an iTXt chunk */
+void /* PRIVATE */
+png_write_iTXt(png_structrp png_ptr, int compression, png_const_charp key,
+    png_const_charp lang, png_const_charp lang_key, png_const_charp text)
+{
+   png_uint_32 key_len, prefix_len;
+   png_size_t lang_len, lang_key_len;
+   png_byte new_key[82];
+   compression_state comp;
+
+   png_debug(1, "in png_write_iTXt");
+
+   key_len = png_check_keyword(png_ptr, key, new_key);
+
+   if (key_len == 0)
+      png_error(png_ptr, "iTXt: invalid keyword");
+
+   /* Set the compression flag */
+   switch (compression)
+   {
+      case PNG_ITXT_COMPRESSION_NONE:
+      case PNG_TEXT_COMPRESSION_NONE:
+         compression = new_key[++key_len] = 0; /* no compression */
+         break;
+
+      case PNG_TEXT_COMPRESSION_zTXt:
+      case PNG_ITXT_COMPRESSION_zTXt:
+         compression = new_key[++key_len] = 1; /* compressed */
+         break;
+
+      default:
+         png_error(png_ptr, "iTXt: invalid compression");
+   }
+
+   new_key[++key_len] = PNG_COMPRESSION_TYPE_BASE;
+   ++key_len; /* for the keywod separator */
+
+   /* We leave it to the application to meet PNG-1.0 requirements on the
+    * contents of the text.  PNG-1.0 through PNG-1.2 discourage the use of
+    * any non-Latin-1 characters except for NEWLINE.  ISO PNG, however,
+    * specifies that the text is UTF-8 and this really doesn't require any
+    * checking.
+    *
+    * The NUL character is forbidden by PNG-1.0 through PNG-1.2 and ISO PNG.
+    *
+    * TODO: validate the language tag correctly (see the spec.)
+    */
+   if (lang == NULL) lang = ""; /* empty language is valid */
+   lang_len = strlen(lang)+1;
+   if (lang_key == NULL) lang_key = ""; /* may be empty */
+   lang_key_len = strlen(lang_key)+1;
+   if (text == NULL) text = ""; /* may be empty */
+
+   prefix_len = key_len;
+   if (lang_len > PNG_UINT_31_MAX-prefix_len)
+      prefix_len = PNG_UINT_31_MAX;
+   else
+      prefix_len = (png_uint_32)(prefix_len + lang_len);
+
+   if (lang_key_len > PNG_UINT_31_MAX-prefix_len)
+      prefix_len = PNG_UINT_31_MAX;
+   else
+      prefix_len = (png_uint_32)(prefix_len + lang_key_len);
+
+   png_text_compress_init(&comp, (png_const_bytep)text, strlen(text));
+
+   if (compression)
+   {
+      if (png_text_compress(png_ptr, png_iTXt, &comp, prefix_len) != Z_OK)
+         png_error(png_ptr, png_ptr->zstream.msg);
+   }
+
+   else
+   {
+      if (comp.input_len > PNG_UINT_31_MAX-prefix_len)
+         png_error(png_ptr, "iTXt: uncompressed text too long");
+
+      /* So the string will fit in a chunk: */
+      comp.output_len = (png_uint_32)/*SAFE*/comp.input_len;
+   }
+
+   png_write_chunk_header(png_ptr, png_iTXt, comp.output_len + prefix_len);
+
+   png_write_chunk_data(png_ptr, new_key, key_len);
+
+   png_write_chunk_data(png_ptr, (png_const_bytep)lang, lang_len);
+
+   png_write_chunk_data(png_ptr, (png_const_bytep)lang_key, lang_key_len);
+
+   if (compression)
+      png_write_compressed_data_out(png_ptr, &comp);
+
+   else
+      png_write_chunk_data(png_ptr, (png_const_bytep)text, comp.input_len);
+
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_oFFs_SUPPORTED
+/* Write the oFFs chunk */
+void /* PRIVATE */
+png_write_oFFs(png_structrp png_ptr, png_int_32 x_offset, png_int_32 y_offset,
+    int unit_type)
+{
+   png_byte buf[9];
+
+   png_debug(1, "in png_write_oFFs");
+
+   if (unit_type >= PNG_OFFSET_LAST)
+      png_warning(png_ptr, "Unrecognized unit type for oFFs chunk");
+
+   png_save_int_32(buf, x_offset);
+   png_save_int_32(buf + 4, y_offset);
+   buf[8] = (png_byte)unit_type;
+
+   png_write_complete_chunk(png_ptr, png_oFFs, buf, (png_size_t)9);
+}
+#endif
+#ifdef PNG_WRITE_pCAL_SUPPORTED
+/* Write the pCAL chunk (described in the PNG extensions document) */
+void /* PRIVATE */
+png_write_pCAL(png_structrp png_ptr, png_charp purpose, png_int_32 X0,
+    png_int_32 X1, int type, int nparams, png_const_charp units,
+    png_charpp params)
+{
+   png_uint_32 purpose_len;
+   png_size_t units_len, total_len;
+   png_size_tp params_len;
+   png_byte buf[10];
+   png_byte new_purpose[80];
+   int i;
+
+   png_debug1(1, "in png_write_pCAL (%d parameters)", nparams);
+
+   if (type >= PNG_EQUATION_LAST)
+      png_error(png_ptr, "Unrecognized equation type for pCAL chunk");
+
+   purpose_len = png_check_keyword(png_ptr, purpose, new_purpose);
+
+   if (purpose_len == 0)
+      png_error(png_ptr, "pCAL: invalid keyword");
+
+   ++purpose_len; /* terminator */
+
+   png_debug1(3, "pCAL purpose length = %d", (int)purpose_len);
+   units_len = strlen(units) + (nparams == 0 ? 0 : 1);
+   png_debug1(3, "pCAL units length = %d", (int)units_len);
+   total_len = purpose_len + units_len + 10;
+
+   params_len = (png_size_tp)png_malloc(png_ptr,
+       (png_alloc_size_t)(nparams * (sizeof (png_size_t))));
+
+   /* Find the length of each parameter, making sure we don't count the
+    * null terminator for the last parameter.
+    */
+   for (i = 0; i < nparams; i++)
+   {
+      params_len[i] = strlen(params[i]) + (i == nparams - 1 ? 0 : 1);
+      png_debug2(3, "pCAL parameter %d length = %lu", i,
+          (unsigned long)params_len[i]);
+      total_len += params_len[i];
+   }
+
+   png_debug1(3, "pCAL total length = %d", (int)total_len);
+   png_write_chunk_header(png_ptr, png_pCAL, (png_uint_32)total_len);
+   png_write_chunk_data(png_ptr, new_purpose, purpose_len);
+   png_save_int_32(buf, X0);
+   png_save_int_32(buf + 4, X1);
+   buf[8] = (png_byte)type;
+   buf[9] = (png_byte)nparams;
+   png_write_chunk_data(png_ptr, buf, (png_size_t)10);
+   png_write_chunk_data(png_ptr, (png_const_bytep)units, (png_size_t)units_len);
+
+   for (i = 0; i < nparams; i++)
+   {
+      png_write_chunk_data(png_ptr, (png_const_bytep)params[i], params_len[i]);
+   }
+
+   png_free(png_ptr, params_len);
+   png_write_chunk_end(png_ptr);
+}
+#endif
+
+#ifdef PNG_WRITE_sCAL_SUPPORTED
+/* Write the sCAL chunk */
+void /* PRIVATE */
+png_write_sCAL_s(png_structrp png_ptr, int unit, png_const_charp width,
+    png_const_charp height)
+{
+   png_byte buf[64];
+   png_size_t wlen, hlen, total_len;
+
+   png_debug(1, "in png_write_sCAL_s");
+
+   wlen = strlen(width);
+   hlen = strlen(height);
+   total_len = wlen + hlen + 2;
+
+   if (total_len > 64)
+   {
+      png_warning(png_ptr, "Can't write sCAL (buffer too small)");
+      return;
+   }
+
+   buf[0] = (png_byte)unit;
+   memcpy(buf + 1, width, wlen + 1);      /* Append the '\0' here */
+   memcpy(buf + wlen + 2, height, hlen);  /* Do NOT append the '\0' here */
+
+   png_debug1(3, "sCAL total length = %u", (unsigned int)total_len);
+   png_write_complete_chunk(png_ptr, png_sCAL, buf, total_len);
+}
+#endif
+
+#ifdef PNG_WRITE_pHYs_SUPPORTED
+/* Write the pHYs chunk */
+void /* PRIVATE */
+png_write_pHYs(png_structrp png_ptr, png_uint_32 x_pixels_per_unit,
+    png_uint_32 y_pixels_per_unit,
+    int unit_type)
+{
+   png_byte buf[9];
+
+   png_debug(1, "in png_write_pHYs");
+
+   if (unit_type >= PNG_RESOLUTION_LAST)
+      png_warning(png_ptr, "Unrecognized unit type for pHYs chunk");
+
+   png_save_uint_32(buf, x_pixels_per_unit);
+   png_save_uint_32(buf + 4, y_pixels_per_unit);
+   buf[8] = (png_byte)unit_type;
+
+   png_write_complete_chunk(png_ptr, png_pHYs, buf, (png_size_t)9);
+}
+#endif
+
+#ifdef PNG_WRITE_tIME_SUPPORTED
+/* Write the tIME chunk.  Use either png_convert_from_struct_tm()
+ * or png_convert_from_time_t(), or fill in the structure yourself.
+ */
+void /* PRIVATE */
+png_write_tIME(png_structrp png_ptr, png_const_timep mod_time)
+{
+   png_byte buf[7];
+
+   png_debug(1, "in png_write_tIME");
+
+   if (mod_time->month  > 12 || mod_time->month  < 1 ||
+       mod_time->day    > 31 || mod_time->day    < 1 ||
+       mod_time->hour   > 23 || mod_time->second > 60)
+   {
+      png_warning(png_ptr, "Invalid time specified for tIME chunk");
+      return;
+   }
+
+   png_save_uint_16(buf, mod_time->year);
+   buf[2] = mod_time->month;
+   buf[3] = mod_time->day;
+   buf[4] = mod_time->hour;
+   buf[5] = mod_time->minute;
+   buf[6] = mod_time->second;
+
+   png_write_complete_chunk(png_ptr, png_tIME, buf, (png_size_t)7);
+}
+#endif
+
+/* Initializes the row writing capability of libpng */
+void /* PRIVATE */
+png_write_start_row(png_structrp png_ptr)
+{
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   /* Start of interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_ystart[7] = {0, 0, 4, 0, 2, 0, 1};
+
+   /* Offset to next interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_yinc[7] = {8, 8, 8, 4, 4, 2, 2};
+#endif
+
+   png_alloc_size_t buf_size;
+   int usr_pixel_depth;
+
+   png_debug(1, "in png_write_start_row");
+
+   usr_pixel_depth = png_ptr->usr_channels * png_ptr->usr_bit_depth;
+   buf_size = PNG_ROWBYTES(usr_pixel_depth, png_ptr->width) + 1;
+
+   /* 1.5.6: added to allow checking in the row write code. */
+   png_ptr->transformed_pixel_depth = png_ptr->pixel_depth;
+   png_ptr->maximum_pixel_depth = (png_byte)usr_pixel_depth;
+
+   /* Set up row buffer */
+   png_ptr->row_buf = (png_bytep)png_malloc(png_ptr, buf_size);
+
+   png_ptr->row_buf[0] = PNG_FILTER_VALUE_NONE;
+
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+   /* Set up filtering buffer, if using this filter */
+   if (png_ptr->do_filter & PNG_FILTER_SUB)
+   {
+      png_ptr->sub_row = (png_bytep)png_malloc(png_ptr, png_ptr->rowbytes + 1);
+
+      png_ptr->sub_row[0] = PNG_FILTER_VALUE_SUB;
+   }
+
+   /* We only need to keep the previous row if we are using one of these. */
+   if (png_ptr->do_filter & (PNG_FILTER_AVG | PNG_FILTER_UP | PNG_FILTER_PAETH))
+   {
+      /* Set up previous row buffer */
+      png_ptr->prev_row = (png_bytep)png_calloc(png_ptr, buf_size);
+
+      if (png_ptr->do_filter & PNG_FILTER_UP)
+      {
+         png_ptr->up_row = (png_bytep)png_malloc(png_ptr,
+            png_ptr->rowbytes + 1);
+
+         png_ptr->up_row[0] = PNG_FILTER_VALUE_UP;
+      }
+
+      if (png_ptr->do_filter & PNG_FILTER_AVG)
+      {
+         png_ptr->avg_row = (png_bytep)png_malloc(png_ptr,
+             png_ptr->rowbytes + 1);
+
+         png_ptr->avg_row[0] = PNG_FILTER_VALUE_AVG;
+      }
+
+      if (png_ptr->do_filter & PNG_FILTER_PAETH)
+      {
+         png_ptr->paeth_row = (png_bytep)png_malloc(png_ptr,
+             png_ptr->rowbytes + 1);
+
+         png_ptr->paeth_row[0] = PNG_FILTER_VALUE_PAETH;
+      }
+   }
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* If interlaced, we need to set up width and height of pass */
+   if (png_ptr->interlaced)
+   {
+      if (!(png_ptr->transformations & PNG_INTERLACE))
+      {
+         png_ptr->num_rows = (png_ptr->height + png_pass_yinc[0] - 1 -
+             png_pass_ystart[0]) / png_pass_yinc[0];
+
+         png_ptr->usr_width = (png_ptr->width + png_pass_inc[0] - 1 -
+             png_pass_start[0]) / png_pass_inc[0];
+      }
+
+      else
+      {
+         png_ptr->num_rows = png_ptr->height;
+         png_ptr->usr_width = png_ptr->width;
+      }
+   }
+
+   else
+#endif
+   {
+      png_ptr->num_rows = png_ptr->height;
+      png_ptr->usr_width = png_ptr->width;
+   }
+}
+
+/* Internal use only.  Called when finished processing a row of data. */
+void /* PRIVATE */
+png_write_finish_row(png_structrp png_ptr)
+{
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   /* Start of interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_ystart[7] = {0, 0, 4, 0, 2, 0, 1};
+
+   /* Offset to next interlace block in the y direction */
+   static PNG_CONST png_byte png_pass_yinc[7] = {8, 8, 8, 4, 4, 2, 2};
+#endif
+
+   png_debug(1, "in png_write_finish_row");
+
+   /* Next row */
+   png_ptr->row_number++;
+
+   /* See if we are done */
+   if (png_ptr->row_number < png_ptr->num_rows)
+      return;
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+   /* If interlaced, go to next pass */
+   if (png_ptr->interlaced)
+   {
+      png_ptr->row_number = 0;
+      if (png_ptr->transformations & PNG_INTERLACE)
+      {
+         png_ptr->pass++;
+      }
+
+      else
+      {
+         /* Loop until we find a non-zero width or height pass */
+         do
+         {
+            png_ptr->pass++;
+
+            if (png_ptr->pass >= 7)
+               break;
+
+            png_ptr->usr_width = (png_ptr->width +
+                png_pass_inc[png_ptr->pass] - 1 -
+                png_pass_start[png_ptr->pass]) /
+                png_pass_inc[png_ptr->pass];
+
+            png_ptr->num_rows = (png_ptr->height +
+                png_pass_yinc[png_ptr->pass] - 1 -
+                png_pass_ystart[png_ptr->pass]) /
+                png_pass_yinc[png_ptr->pass];
+
+            if (png_ptr->transformations & PNG_INTERLACE)
+               break;
+
+         } while (png_ptr->usr_width == 0 || png_ptr->num_rows == 0);
+
+      }
+
+      /* Reset the row above the image for the next pass */
+      if (png_ptr->pass < 7)
+      {
+         if (png_ptr->prev_row != NULL)
+            memset(png_ptr->prev_row, 0,
+                (png_size_t)(PNG_ROWBYTES(png_ptr->usr_channels*
+                png_ptr->usr_bit_depth, png_ptr->width)) + 1);
+
+         return;
+      }
+   }
+#endif
+
+   /* If we get here, we've just written the last row, so we need
+      to flush the compressor */
+   png_compress_IDAT(png_ptr, NULL, 0, Z_FINISH);
+}
+
+#ifdef PNG_WRITE_INTERLACING_SUPPORTED
+/* Pick out the correct pixels for the interlace pass.
+ * The basic idea here is to go through the row with a source
+ * pointer and a destination pointer (sp and dp), and copy the
+ * correct pixels for the pass.  As the row gets compacted,
+ * sp will always be >= dp, so we should never overwrite anything.
+ * See the default: case for the easiest code to understand.
+ */
+void /* PRIVATE */
+png_do_write_interlace(png_row_infop row_info, png_bytep row, int pass)
+{
+   /* Arrays to facilitate easy interlacing - use pass (0 - 6) as index */
+
+   /* Start of interlace block */
+   static PNG_CONST png_byte png_pass_start[7] = {0, 4, 0, 2, 0, 1, 0};
+
+   /* Offset to next interlace block */
+   static PNG_CONST png_byte  png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
+
+   png_debug(1, "in png_do_write_interlace");
+
+   /* We don't have to do anything on the last pass (6) */
+   if (pass < 6)
+   {
+      /* Each pixel depth is handled separately */
+      switch (row_info->pixel_depth)
+      {
+         case 1:
+         {
+            png_bytep sp;
+            png_bytep dp;
+            int shift;
+            int d;
+            int value;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            dp = row;
+            d = 0;
+            shift = 7;
+
+            for (i = png_pass_start[pass]; i < row_width;
+               i += png_pass_inc[pass])
+            {
+               sp = row + (png_size_t)(i >> 3);
+               value = (int)(*sp >> (7 - (int)(i & 0x07))) & 0x01;
+               d |= (value << shift);
+
+               if (shift == 0)
+               {
+                  shift = 7;
+                  *dp++ = (png_byte)d;
+                  d = 0;
+               }
+
+               else
+                  shift--;
+
+            }
+            if (shift != 7)
+               *dp = (png_byte)d;
+
+            break;
+         }
+
+         case 2:
+         {
+            png_bytep sp;
+            png_bytep dp;
+            int shift;
+            int d;
+            int value;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            dp = row;
+            shift = 6;
+            d = 0;
+
+            for (i = png_pass_start[pass]; i < row_width;
+               i += png_pass_inc[pass])
+            {
+               sp = row + (png_size_t)(i >> 2);
+               value = (*sp >> ((3 - (int)(i & 0x03)) << 1)) & 0x03;
+               d |= (value << shift);
+
+               if (shift == 0)
+               {
+                  shift = 6;
+                  *dp++ = (png_byte)d;
+                  d = 0;
+               }
+
+               else
+                  shift -= 2;
+            }
+            if (shift != 6)
+               *dp = (png_byte)d;
+
+            break;
+         }
+
+         case 4:
+         {
+            png_bytep sp;
+            png_bytep dp;
+            int shift;
+            int d;
+            int value;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+
+            dp = row;
+            shift = 4;
+            d = 0;
+            for (i = png_pass_start[pass]; i < row_width;
+                i += png_pass_inc[pass])
+            {
+               sp = row + (png_size_t)(i >> 1);
+               value = (*sp >> ((1 - (int)(i & 0x01)) << 2)) & 0x0f;
+               d |= (value << shift);
+
+               if (shift == 0)
+               {
+                  shift = 4;
+                  *dp++ = (png_byte)d;
+                  d = 0;
+               }
+
+               else
+                  shift -= 4;
+            }
+            if (shift != 4)
+               *dp = (png_byte)d;
+
+            break;
+         }
+
+         default:
+         {
+            png_bytep sp;
+            png_bytep dp;
+            png_uint_32 i;
+            png_uint_32 row_width = row_info->width;
+            png_size_t pixel_bytes;
+
+            /* Start at the beginning */
+            dp = row;
+
+            /* Find out how many bytes each pixel takes up */
+            pixel_bytes = (row_info->pixel_depth >> 3);
+
+            /* Loop through the row, only looking at the pixels that matter */
+            for (i = png_pass_start[pass]; i < row_width;
+               i += png_pass_inc[pass])
+            {
+               /* Find out where the original pixel is */
+               sp = row + (png_size_t)i * pixel_bytes;
+
+               /* Move the pixel */
+               if (dp != sp)
+                  memcpy(dp, sp, pixel_bytes);
+
+               /* Next pixel */
+               dp += pixel_bytes;
+            }
+            break;
+         }
+      }
+      /* Set new row width */
+      row_info->width = (row_info->width +
+          png_pass_inc[pass] - 1 -
+          png_pass_start[pass]) /
+          png_pass_inc[pass];
+
+      row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth,
+          row_info->width);
+   }
+}
+#endif
+
+/* This filters the row, chooses which filter to use, if it has not already
+ * been specified by the application, and then writes the row out with the
+ * chosen filter.
+ */
+static void png_write_filtered_row(png_structrp png_ptr, png_bytep filtered_row,
+   png_size_t row_bytes);
+
+#define PNG_MAXSUM (((png_uint_32)(-1)) >> 1)
+#define PNG_HISHIFT 10
+#define PNG_LOMASK ((png_uint_32)0xffffL)
+#define PNG_HIMASK ((png_uint_32)(~PNG_LOMASK >> PNG_HISHIFT))
+void /* PRIVATE */
+png_write_find_filter(png_structrp png_ptr, png_row_infop row_info)
+{
+   png_bytep best_row;
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+   png_bytep prev_row, row_buf;
+   png_uint_32 mins, bpp;
+   png_byte filter_to_do = png_ptr->do_filter;
+   png_size_t row_bytes = row_info->rowbytes;
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+   int num_p_filters = png_ptr->num_prev_filters;
+#endif
+
+   png_debug(1, "in png_write_find_filter");
+
+#ifndef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+  if (png_ptr->row_number == 0 && filter_to_do == PNG_ALL_FILTERS)
+  {
+     /* These will never be selected so we need not test them. */
+     filter_to_do &= ~(PNG_FILTER_UP | PNG_FILTER_PAETH);
+  }
+#endif
+
+   /* Find out how many bytes offset each pixel is */
+   bpp = (row_info->pixel_depth + 7) >> 3;
+
+   prev_row = png_ptr->prev_row;
+#endif
+   best_row = png_ptr->row_buf;
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+   row_buf = best_row;
+   mins = PNG_MAXSUM;
+
+   /* The prediction method we use is to find which method provides the
+    * smallest value when summing the absolute values of the distances
+    * from zero, using anything >= 128 as negative numbers.  This is known
+    * as the "minimum sum of absolute differences" heuristic.  Other
+    * heuristics are the "weighted minimum sum of absolute differences"
+    * (experimental and can in theory improve compression), and the "zlib
+    * predictive" method (not implemented yet), which does test compressions
+    * of lines using different filter methods, and then chooses the
+    * (series of) filter(s) that give minimum compressed data size (VERY
+    * computationally expensive).
+    *
+    * GRR 980525:  consider also
+    *
+    *   (1) minimum sum of absolute differences from running average (i.e.,
+    *       keep running sum of non-absolute differences & count of bytes)
+    *       [track dispersion, too?  restart average if dispersion too large?]
+    *
+    *  (1b) minimum sum of absolute differences from sliding average, probably
+    *       with window size <= deflate window (usually 32K)
+    *
+    *   (2) minimum sum of squared differences from zero or running average
+    *       (i.e., ~ root-mean-square approach)
+    */
+
+
+   /* We don't need to test the 'no filter' case if this is the only filter
+    * that has been chosen, as it doesn't actually do anything to the data.
+    */
+   if ((filter_to_do & PNG_FILTER_NONE) && filter_to_do != PNG_FILTER_NONE)
+   {
+      png_bytep rp;
+      png_uint_32 sum = 0;
+      png_size_t i;
+      int v;
+
+      for (i = 0, rp = row_buf + 1; i < row_bytes; i++, rp++)
+      {
+         v = *rp;
+         sum += (v < 128) ? v : 256 - v;
+      }
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         png_uint_32 sumhi, sumlo;
+         int j;
+         sumlo = sum & PNG_LOMASK;
+         sumhi = (sum >> PNG_HISHIFT) & PNG_HIMASK; /* Gives us some footroom */
+
+         /* Reduce the sum if we match any of the previous rows */
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_NONE)
+            {
+               sumlo = (sumlo * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               sumhi = (sumhi * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         /* Factor in the cost of this filter (this is here for completeness,
+          * but it makes no sense to have a "cost" for the NONE filter, as
+          * it has the minimum possible computational cost - none).
+          */
+         sumlo = (sumlo * png_ptr->filter_costs[PNG_FILTER_VALUE_NONE]) >>
+             PNG_COST_SHIFT;
+
+         sumhi = (sumhi * png_ptr->filter_costs[PNG_FILTER_VALUE_NONE]) >>
+             PNG_COST_SHIFT;
+
+         if (sumhi > PNG_HIMASK)
+            sum = PNG_MAXSUM;
+
+         else
+            sum = (sumhi << PNG_HISHIFT) + sumlo;
+      }
+#endif
+      mins = sum;
+   }
+
+   /* Sub filter */
+   if (filter_to_do == PNG_FILTER_SUB)
+   /* It's the only filter so no testing is needed */
+   {
+      png_bytep rp, lp, dp;
+      png_size_t i;
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->sub_row + 1; i < bpp;
+           i++, rp++, dp++)
+      {
+         *dp = *rp;
+      }
+
+      for (lp = row_buf + 1; i < row_bytes;
+         i++, rp++, lp++, dp++)
+      {
+         *dp = (png_byte)(((int)*rp - (int)*lp) & 0xff);
+      }
+
+      best_row = png_ptr->sub_row;
+   }
+
+   else if (filter_to_do & PNG_FILTER_SUB)
+   {
+      png_bytep rp, dp, lp;
+      png_uint_32 sum = 0, lmins = mins;
+      png_size_t i;
+      int v;
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      /* We temporarily increase the "minimum sum" by the factor we
+       * would reduce the sum of this filter, so that we can do the
+       * early exit comparison without scaling the sum each time.
+       */
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 lmhi, lmlo;
+         lmlo = lmins & PNG_LOMASK;
+         lmhi = (lmins >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_SUB)
+            {
+               lmlo = (lmlo * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               lmhi = (lmhi * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         lmlo = (lmlo * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_SUB]) >>
+             PNG_COST_SHIFT;
+
+         lmhi = (lmhi * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_SUB]) >>
+             PNG_COST_SHIFT;
+
+         if (lmhi > PNG_HIMASK)
+            lmins = PNG_MAXSUM;
+
+         else
+            lmins = (lmhi << PNG_HISHIFT) + lmlo;
+      }
+#endif
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->sub_row + 1; i < bpp;
+           i++, rp++, dp++)
+      {
+         v = *dp = *rp;
+
+         sum += (v < 128) ? v : 256 - v;
+      }
+
+      for (lp = row_buf + 1; i < row_bytes;
+         i++, rp++, lp++, dp++)
+      {
+         v = *dp = (png_byte)(((int)*rp - (int)*lp) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+
+         if (sum > lmins)  /* We are already worse, don't continue. */
+            break;
+      }
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 sumhi, sumlo;
+         sumlo = sum & PNG_LOMASK;
+         sumhi = (sum >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_SUB)
+            {
+               sumlo = (sumlo * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               sumhi = (sumhi * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         sumlo = (sumlo * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_SUB]) >>
+             PNG_COST_SHIFT;
+
+         sumhi = (sumhi * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_SUB]) >>
+             PNG_COST_SHIFT;
+
+         if (sumhi > PNG_HIMASK)
+            sum = PNG_MAXSUM;
+
+         else
+            sum = (sumhi << PNG_HISHIFT) + sumlo;
+      }
+#endif
+
+      if (sum < mins)
+      {
+         mins = sum;
+         best_row = png_ptr->sub_row;
+      }
+   }
+
+   /* Up filter */
+   if (filter_to_do == PNG_FILTER_UP)
+   {
+      png_bytep rp, dp, pp;
+      png_size_t i;
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->up_row + 1,
+          pp = prev_row + 1; i < row_bytes;
+          i++, rp++, pp++, dp++)
+      {
+         *dp = (png_byte)(((int)*rp - (int)*pp) & 0xff);
+      }
+
+      best_row = png_ptr->up_row;
+   }
+
+   else if (filter_to_do & PNG_FILTER_UP)
+   {
+      png_bytep rp, dp, pp;
+      png_uint_32 sum = 0, lmins = mins;
+      png_size_t i;
+      int v;
+
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 lmhi, lmlo;
+         lmlo = lmins & PNG_LOMASK;
+         lmhi = (lmins >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_UP)
+            {
+               lmlo = (lmlo * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               lmhi = (lmhi * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         lmlo = (lmlo * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_UP]) >>
+             PNG_COST_SHIFT;
+
+         lmhi = (lmhi * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_UP]) >>
+             PNG_COST_SHIFT;
+
+         if (lmhi > PNG_HIMASK)
+            lmins = PNG_MAXSUM;
+
+         else
+            lmins = (lmhi << PNG_HISHIFT) + lmlo;
+      }
+#endif
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->up_row + 1,
+          pp = prev_row + 1; i < row_bytes; i++)
+      {
+         v = *dp++ = (png_byte)(((int)*rp++ - (int)*pp++) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+
+         if (sum > lmins)  /* We are already worse, don't continue. */
+            break;
+      }
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 sumhi, sumlo;
+         sumlo = sum & PNG_LOMASK;
+         sumhi = (sum >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_UP)
+            {
+               sumlo = (sumlo * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               sumhi = (sumhi * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         sumlo = (sumlo * png_ptr->filter_costs[PNG_FILTER_VALUE_UP]) >>
+             PNG_COST_SHIFT;
+
+         sumhi = (sumhi * png_ptr->filter_costs[PNG_FILTER_VALUE_UP]) >>
+             PNG_COST_SHIFT;
+
+         if (sumhi > PNG_HIMASK)
+            sum = PNG_MAXSUM;
+
+         else
+            sum = (sumhi << PNG_HISHIFT) + sumlo;
+      }
+#endif
+
+      if (sum < mins)
+      {
+         mins = sum;
+         best_row = png_ptr->up_row;
+      }
+   }
+
+   /* Avg filter */
+   if (filter_to_do == PNG_FILTER_AVG)
+   {
+      png_bytep rp, dp, pp, lp;
+      png_uint_32 i;
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->avg_row + 1,
+           pp = prev_row + 1; i < bpp; i++)
+      {
+         *dp++ = (png_byte)(((int)*rp++ - ((int)*pp++ / 2)) & 0xff);
+      }
+
+      for (lp = row_buf + 1; i < row_bytes; i++)
+      {
+         *dp++ = (png_byte)(((int)*rp++ - (((int)*pp++ + (int)*lp++) / 2))
+                 & 0xff);
+      }
+      best_row = png_ptr->avg_row;
+   }
+
+   else if (filter_to_do & PNG_FILTER_AVG)
+   {
+      png_bytep rp, dp, pp, lp;
+      png_uint_32 sum = 0, lmins = mins;
+      png_size_t i;
+      int v;
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 lmhi, lmlo;
+         lmlo = lmins & PNG_LOMASK;
+         lmhi = (lmins >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_AVG)
+            {
+               lmlo = (lmlo * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               lmhi = (lmhi * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         lmlo = (lmlo * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_AVG]) >>
+             PNG_COST_SHIFT;
+
+         lmhi = (lmhi * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_AVG]) >>
+             PNG_COST_SHIFT;
+
+         if (lmhi > PNG_HIMASK)
+            lmins = PNG_MAXSUM;
+
+         else
+            lmins = (lmhi << PNG_HISHIFT) + lmlo;
+      }
+#endif
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->avg_row + 1,
+           pp = prev_row + 1; i < bpp; i++)
+      {
+         v = *dp++ = (png_byte)(((int)*rp++ - ((int)*pp++ / 2)) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+      }
+
+      for (lp = row_buf + 1; i < row_bytes; i++)
+      {
+         v = *dp++ =
+             (png_byte)(((int)*rp++ - (((int)*pp++ + (int)*lp++) / 2)) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+
+         if (sum > lmins)  /* We are already worse, don't continue. */
+            break;
+      }
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 sumhi, sumlo;
+         sumlo = sum & PNG_LOMASK;
+         sumhi = (sum >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_NONE)
+            {
+               sumlo = (sumlo * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               sumhi = (sumhi * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         sumlo = (sumlo * png_ptr->filter_costs[PNG_FILTER_VALUE_AVG]) >>
+             PNG_COST_SHIFT;
+
+         sumhi = (sumhi * png_ptr->filter_costs[PNG_FILTER_VALUE_AVG]) >>
+             PNG_COST_SHIFT;
+
+         if (sumhi > PNG_HIMASK)
+            sum = PNG_MAXSUM;
+
+         else
+            sum = (sumhi << PNG_HISHIFT) + sumlo;
+      }
+#endif
+
+      if (sum < mins)
+      {
+         mins = sum;
+         best_row = png_ptr->avg_row;
+      }
+   }
+
+   /* Paeth filter */
+   if (filter_to_do == PNG_FILTER_PAETH)
+   {
+      png_bytep rp, dp, pp, cp, lp;
+      png_size_t i;
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->paeth_row + 1,
+          pp = prev_row + 1; i < bpp; i++)
+      {
+         *dp++ = (png_byte)(((int)*rp++ - (int)*pp++) & 0xff);
+      }
+
+      for (lp = row_buf + 1, cp = prev_row + 1; i < row_bytes; i++)
+      {
+         int a, b, c, pa, pb, pc, p;
+
+         b = *pp++;
+         c = *cp++;
+         a = *lp++;
+
+         p = b - c;
+         pc = a - c;
+
+#ifdef PNG_USE_ABS
+         pa = abs(p);
+         pb = abs(pc);
+         pc = abs(p + pc);
+#else
+         pa = p < 0 ? -p : p;
+         pb = pc < 0 ? -pc : pc;
+         pc = (p + pc) < 0 ? -(p + pc) : p + pc;
+#endif
+
+         p = (pa <= pb && pa <=pc) ? a : (pb <= pc) ? b : c;
+
+         *dp++ = (png_byte)(((int)*rp++ - p) & 0xff);
+      }
+      best_row = png_ptr->paeth_row;
+   }
+
+   else if (filter_to_do & PNG_FILTER_PAETH)
+   {
+      png_bytep rp, dp, pp, cp, lp;
+      png_uint_32 sum = 0, lmins = mins;
+      png_size_t i;
+      int v;
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 lmhi, lmlo;
+         lmlo = lmins & PNG_LOMASK;
+         lmhi = (lmins >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_PAETH)
+            {
+               lmlo = (lmlo * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               lmhi = (lmhi * png_ptr->inv_filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         lmlo = (lmlo * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_PAETH]) >>
+             PNG_COST_SHIFT;
+
+         lmhi = (lmhi * png_ptr->inv_filter_costs[PNG_FILTER_VALUE_PAETH]) >>
+             PNG_COST_SHIFT;
+
+         if (lmhi > PNG_HIMASK)
+            lmins = PNG_MAXSUM;
+
+         else
+            lmins = (lmhi << PNG_HISHIFT) + lmlo;
+      }
+#endif
+
+      for (i = 0, rp = row_buf + 1, dp = png_ptr->paeth_row + 1,
+          pp = prev_row + 1; i < bpp; i++)
+      {
+         v = *dp++ = (png_byte)(((int)*rp++ - (int)*pp++) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+      }
+
+      for (lp = row_buf + 1, cp = prev_row + 1; i < row_bytes; i++)
+      {
+         int a, b, c, pa, pb, pc, p;
+
+         b = *pp++;
+         c = *cp++;
+         a = *lp++;
+
+#ifndef PNG_SLOW_PAETH
+         p = b - c;
+         pc = a - c;
+#ifdef PNG_USE_ABS
+         pa = abs(p);
+         pb = abs(pc);
+         pc = abs(p + pc);
+#else
+         pa = p < 0 ? -p : p;
+         pb = pc < 0 ? -pc : pc;
+         pc = (p + pc) < 0 ? -(p + pc) : p + pc;
+#endif
+         p = (pa <= pb && pa <=pc) ? a : (pb <= pc) ? b : c;
+#else /* PNG_SLOW_PAETH */
+         p = a + b - c;
+         pa = abs(p - a);
+         pb = abs(p - b);
+         pc = abs(p - c);
+
+         if (pa <= pb && pa <= pc)
+            p = a;
+
+         else if (pb <= pc)
+            p = b;
+
+         else
+            p = c;
+#endif /* PNG_SLOW_PAETH */
+
+         v = *dp++ = (png_byte)(((int)*rp++ - p) & 0xff);
+
+         sum += (v < 128) ? v : 256 - v;
+
+         if (sum > lmins)  /* We are already worse, don't continue. */
+            break;
+      }
+
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+      if (png_ptr->heuristic_method == PNG_FILTER_HEURISTIC_WEIGHTED)
+      {
+         int j;
+         png_uint_32 sumhi, sumlo;
+         sumlo = sum & PNG_LOMASK;
+         sumhi = (sum >> PNG_HISHIFT) & PNG_HIMASK;
+
+         for (j = 0; j < num_p_filters; j++)
+         {
+            if (png_ptr->prev_filters[j] == PNG_FILTER_VALUE_PAETH)
+            {
+               sumlo = (sumlo * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+
+               sumhi = (sumhi * png_ptr->filter_weights[j]) >>
+                   PNG_WEIGHT_SHIFT;
+            }
+         }
+
+         sumlo = (sumlo * png_ptr->filter_costs[PNG_FILTER_VALUE_PAETH]) >>
+             PNG_COST_SHIFT;
+
+         sumhi = (sumhi * png_ptr->filter_costs[PNG_FILTER_VALUE_PAETH]) >>
+             PNG_COST_SHIFT;
+
+         if (sumhi > PNG_HIMASK)
+            sum = PNG_MAXSUM;
+
+         else
+            sum = (sumhi << PNG_HISHIFT) + sumlo;
+      }
+#endif
+
+      if (sum < mins)
+      {
+         best_row = png_ptr->paeth_row;
+      }
+   }
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+
+   /* Do the actual writing of the filtered row data from the chosen filter. */
+   png_write_filtered_row(png_ptr, best_row, row_info->rowbytes+1);
+
+#ifdef PNG_WRITE_FILTER_SUPPORTED
+#ifdef PNG_WRITE_WEIGHTED_FILTER_SUPPORTED
+   /* Save the type of filter we picked this time for future calculations */
+   if (png_ptr->num_prev_filters > 0)
+   {
+      int j;
+
+      for (j = 1; j < num_p_filters; j++)
+      {
+         png_ptr->prev_filters[j] = png_ptr->prev_filters[j - 1];
+      }
+
+      png_ptr->prev_filters[j] = best_row[0];
+   }
+#endif
+#endif /* PNG_WRITE_FILTER_SUPPORTED */
+}
+
+
+/* Do the actual writing of a previously filtered row. */
+static void
+png_write_filtered_row(png_structrp png_ptr, png_bytep filtered_row,
+   png_size_t full_row_length/*includes filter byte*/)
+{
+   png_debug(1, "in png_write_filtered_row");
+
+   png_debug1(2, "filter = %d", filtered_row[0]);
+
+   png_compress_IDAT(png_ptr, filtered_row, full_row_length, Z_NO_FLUSH);
+
+   /* Swap the current and previous rows */
+   if (png_ptr->prev_row != NULL)
+   {
+      png_bytep tptr;
+
+      tptr = png_ptr->prev_row;
+      png_ptr->prev_row = png_ptr->row_buf;
+      png_ptr->row_buf = tptr;
+   }
+
+   /* Finish row - updates counters and flushes zlib if last row */
+   png_write_finish_row(png_ptr);
+
+#ifdef PNG_WRITE_FLUSH_SUPPORTED
+   png_ptr->flush_rows++;
+
+   if (png_ptr->flush_dist > 0 &&
+       png_ptr->flush_rows >= png_ptr->flush_dist)
+   {
+      png_write_flush(png_ptr);
+   }
+#endif
+}
+#endif /* PNG_WRITE_SUPPORTED */
diff --git a/ml/dlib/dlib/external/pybind11/CMakeLists.txt b/ml/dlib/dlib/external/pybind11/CMakeLists.txt
new file mode 100644
index 000000000..4280ba742
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/CMakeLists.txt
@@ -0,0 +1,155 @@
+# CMakeLists.txt -- Build system for the pybind11 modules
+#
+# Copyright (c) 2015 Wenzel Jakob <wenzel@inf.ethz.ch>
+#
+# All rights reserved. Use of this source code is governed by a
+# BSD-style license that can be found in the LICENSE file.
+
+cmake_minimum_required(VERSION 2.8.12)
+
+if (POLICY CMP0048)
+  # cmake warns if loaded from a min-3.0-required parent dir, so silence the warning:
+  cmake_policy(SET CMP0048 NEW)
+endif()
+
+# CMake versions < 3.4.0 do not support try_compile/pthread checks without C as active language.
+if(CMAKE_VERSION VERSION_LESS 3.4.0)
+  project(pybind11)
+else()
+  project(pybind11 CXX)
+endif()
+
+# Check if pybind11 is being used directly or via add_subdirectory
+set(PYBIND11_MASTER_PROJECT OFF)
+if (CMAKE_CURRENT_SOURCE_DIR STREQUAL CMAKE_SOURCE_DIR)
+  set(PYBIND11_MASTER_PROJECT ON)
+endif()
+
+option(PYBIND11_INSTALL "Install pybind11 header files?" ${PYBIND11_MASTER_PROJECT})
+option(PYBIND11_TEST    "Build pybind11 test suite?"     ${PYBIND11_MASTER_PROJECT})
+
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/tools")
+
+include(pybind11Tools)
+
+# Cache variables so pybind11_add_module can be used in parent projects
+set(PYBIND11_INCLUDE_DIR "${CMAKE_CURRENT_LIST_DIR}/include" CACHE INTERNAL "")
+set(PYTHON_INCLUDE_DIRS ${PYTHON_INCLUDE_DIRS} CACHE INTERNAL "")
+set(PYTHON_LIBRARIES ${PYTHON_LIBRARIES} CACHE INTERNAL "")
+set(PYTHON_MODULE_PREFIX ${PYTHON_MODULE_PREFIX} CACHE INTERNAL "")
+set(PYTHON_MODULE_EXTENSION ${PYTHON_MODULE_EXTENSION} CACHE INTERNAL "")
+
+# NB: when adding a header don't forget to also add it to setup.py
+set(PYBIND11_HEADERS
+  include/pybind11/detail/class.h
+  include/pybind11/detail/common.h
+  include/pybind11/detail/descr.h
+  include/pybind11/detail/init.h
+  include/pybind11/detail/internals.h
+  include/pybind11/detail/typeid.h
+  include/pybind11/attr.h
+  include/pybind11/buffer_info.h
+  include/pybind11/cast.h
+  include/pybind11/chrono.h
+  include/pybind11/common.h
+  include/pybind11/complex.h
+  include/pybind11/options.h
+  include/pybind11/eigen.h
+  include/pybind11/embed.h
+  include/pybind11/eval.h
+  include/pybind11/functional.h
+  include/pybind11/numpy.h
+  include/pybind11/operators.h
+  include/pybind11/pybind11.h
+  include/pybind11/pytypes.h
+  include/pybind11/stl.h
+  include/pybind11/stl_bind.h
+)
+string(REPLACE "include/" "${CMAKE_CURRENT_SOURCE_DIR}/include/"
+       PYBIND11_HEADERS "${PYBIND11_HEADERS}")
+
+if (PYBIND11_TEST)
+  add_subdirectory(tests)
+endif()
+
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+# extract project version from source
+file(STRINGS "${PYBIND11_INCLUDE_DIR}/pybind11/detail/common.h" pybind11_version_defines
+     REGEX "#define PYBIND11_VERSION_(MAJOR|MINOR|PATCH) ")
+foreach(ver ${pybind11_version_defines})
+  if (ver MATCHES "#define PYBIND11_VERSION_(MAJOR|MINOR|PATCH) +([^ ]+)$")
+    set(PYBIND11_VERSION_${CMAKE_MATCH_1} "${CMAKE_MATCH_2}" CACHE INTERNAL "")
+  endif()
+endforeach()
+set(${PROJECT_NAME}_VERSION ${PYBIND11_VERSION_MAJOR}.${PYBIND11_VERSION_MINOR}.${PYBIND11_VERSION_PATCH})
+message(STATUS "pybind11 v${${PROJECT_NAME}_VERSION}")
+
+option (USE_PYTHON_INCLUDE_DIR "Install pybind11 headers in Python include directory instead of default installation prefix" OFF)
+if (USE_PYTHON_INCLUDE_DIR)
+    file(RELATIVE_PATH CMAKE_INSTALL_INCLUDEDIR ${CMAKE_INSTALL_PREFIX} ${PYTHON_INCLUDE_DIRS})
+endif()
+
+if(NOT (CMAKE_VERSION VERSION_LESS 3.0))  # CMake >= 3.0
+  # Build an interface library target:
+  add_library(pybind11 INTERFACE)
+  add_library(pybind11::pybind11 ALIAS pybind11)  # to match exported target
+  target_include_directories(pybind11 INTERFACE $<BUILD_INTERFACE:${PYBIND11_INCLUDE_DIR}>
+                                                $<BUILD_INTERFACE:${PYTHON_INCLUDE_DIRS}>
+                                                $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
+  target_compile_options(pybind11 INTERFACE $<BUILD_INTERFACE:${PYBIND11_CPP_STANDARD}>)
+
+  add_library(module INTERFACE)
+  add_library(pybind11::module ALIAS module)
+  if(NOT MSVC)
+    target_compile_options(module INTERFACE -fvisibility=hidden)
+  endif()
+  target_link_libraries(module INTERFACE pybind11::pybind11)
+  if(WIN32 OR CYGWIN)
+    target_link_libraries(module INTERFACE $<BUILD_INTERFACE:${PYTHON_LIBRARIES}>)
+  elseif(APPLE)
+    target_link_libraries(module INTERFACE "-undefined dynamic_lookup")
+  endif()
+
+  add_library(embed INTERFACE)
+  add_library(pybind11::embed ALIAS embed)
+  target_link_libraries(embed INTERFACE pybind11::pybind11 $<BUILD_INTERFACE:${PYTHON_LIBRARIES}>)
+endif()
+
+if (PYBIND11_INSTALL)
+  install(DIRECTORY ${PYBIND11_INCLUDE_DIR}/pybind11 DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+  # GNUInstallDirs "DATADIR" wrong here; CMake search path wants "share".
+  set(PYBIND11_CMAKECONFIG_INSTALL_DIR "share/cmake/${PROJECT_NAME}" CACHE STRING "install path for pybind11Config.cmake")
+
+  configure_package_config_file(tools/${PROJECT_NAME}Config.cmake.in
+                                "${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}Config.cmake"
+                                INSTALL_DESTINATION ${PYBIND11_CMAKECONFIG_INSTALL_DIR})
+  # Remove CMAKE_SIZEOF_VOID_P from ConfigVersion.cmake since the library does
+  # not depend on architecture specific settings or libraries.
+  set(_PYBIND11_CMAKE_SIZEOF_VOID_P ${CMAKE_SIZEOF_VOID_P})
+  unset(CMAKE_SIZEOF_VOID_P)
+  write_basic_package_version_file(${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}ConfigVersion.cmake
+                                   VERSION ${${PROJECT_NAME}_VERSION}
+                                   COMPATIBILITY AnyNewerVersion)
+  set(CMAKE_SIZEOF_VOID_P ${_PYBIND11_CMAKE_SIZEOF_VOID_P})
+  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}Config.cmake
+                ${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}ConfigVersion.cmake
+                tools/FindPythonLibsNew.cmake
+                tools/pybind11Tools.cmake
+          DESTINATION ${PYBIND11_CMAKECONFIG_INSTALL_DIR})
+
+  if(NOT (CMAKE_VERSION VERSION_LESS 3.0))
+    if(NOT PYBIND11_EXPORT_NAME)
+      set(PYBIND11_EXPORT_NAME "${PROJECT_NAME}Targets")
+    endif()
+
+    install(TARGETS pybind11 module embed
+            EXPORT "${PYBIND11_EXPORT_NAME}")
+    if(PYBIND11_MASTER_PROJECT)
+      install(EXPORT "${PYBIND11_EXPORT_NAME}"
+              NAMESPACE "${PROJECT_NAME}::"
+              DESTINATION ${PYBIND11_CMAKECONFIG_INSTALL_DIR})
+    endif()
+  endif()
+endif()
diff --git a/ml/dlib/dlib/external/pybind11/CONTRIBUTING.md b/ml/dlib/dlib/external/pybind11/CONTRIBUTING.md
new file mode 100644
index 000000000..375735f6c
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/CONTRIBUTING.md
@@ -0,0 +1,47 @@
+Thank you for your interest in this project! Please refer to the following
+sections on how to contribute code and bug reports.
+
+### Reporting bugs
+
+At the moment, this project is run in the spare time of a single person
+([Wenzel Jakob](http://rgl.epfl.ch/people/wjakob)) with very limited resources
+for issue tracker tickets. Thus, before submitting a question or bug report,
+please take a moment of your time and ensure that your issue isn't already
+discussed in the project documentation provided at
+[http://pybind11.readthedocs.org/en/latest](http://pybind11.readthedocs.org/en/latest).
+
+Assuming that you have identified a previously unknown problem or an important
+question, it's essential that you submit a self-contained and minimal piece of
+code that reproduces the problem. In other words: no external dependencies,
+isolate the function(s) that cause breakage, submit matched and complete C++
+and Python snippets that can be easily compiled and run on my end.
+
+## Pull requests
+Contributions are submitted, reviewed, and accepted using Github pull requests.
+Please refer to [this
+article](https://help.github.com/articles/using-pull-requests) for details and
+adhere to the following rules to make the process as smooth as possible:
+
+* Make a new branch for every feature you're working on.
+* Make small and clean pull requests that are easy to review but make sure they
+  do add value by themselves.
+* Add tests for any new functionality and run the test suite (``make pytest``)
+  to ensure that no existing features break.
+* This project has a strong focus on providing general solutions using a
+  minimal amount of code, thus small pull requests are greatly preferred.
+
+### Licensing of contributions
+
+pybind11 is provided under a BSD-style license that can be found in the
+``LICENSE`` file. By using, distributing, or contributing to this project, you
+agree to the terms and conditions of this license.
+
+You are under no obligation whatsoever to provide any bug fixes, patches, or
+upgrades to the features, functionality or performance of the source code
+("Enhancements") to anyone; however, if you choose to make your Enhancements
+available either publicly, or directly to the author of this software, without
+imposing a separate written license agreement for such Enhancements, then you
+hereby grant the following license: a non-exclusive, royalty-free perpetual
+license to install, use, modify, prepare derivative works, incorporate into
+other computer software, distribute, and sublicense such enhancements or
+derivative works thereof, in binary and source code form.
diff --git a/ml/dlib/dlib/external/pybind11/LICENSE b/ml/dlib/dlib/external/pybind11/LICENSE
new file mode 100644
index 000000000..6f15578cc
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/LICENSE
@@ -0,0 +1,29 @@
+Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>, All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors
+   may be used to endorse or promote products derived from this software
+   without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+Please also refer to the file CONTRIBUTING.md, which clarifies licensing of
+external contributions to this project including patches, pull requests, etc.
diff --git a/ml/dlib/dlib/external/pybind11/README.md b/ml/dlib/dlib/external/pybind11/README.md
new file mode 100644
index 000000000..447788240
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/README.md
@@ -0,0 +1,129 @@
+![pybind11 logo](https://github.com/pybind/pybind11/raw/master/docs/pybind11-logo.png)
+
+# pybind11 — Seamless operability between C++11 and Python
+
+[![Documentation Status](https://readthedocs.org/projects/pybind11/badge/?version=master)](http://pybind11.readthedocs.org/en/master/?badge=master)
+[![Documentation Status](https://readthedocs.org/projects/pybind11/badge/?version=stable)](http://pybind11.readthedocs.org/en/stable/?badge=stable)
+[![Gitter chat](https://img.shields.io/gitter/room/gitterHQ/gitter.svg)](https://gitter.im/pybind/Lobby)
+[![Build Status](https://travis-ci.org/pybind/pybind11.svg?branch=master)](https://travis-ci.org/pybind/pybind11)
+[![Build status](https://ci.appveyor.com/api/projects/status/riaj54pn4h08xy40?svg=true)](https://ci.appveyor.com/project/wjakob/pybind11)
+
+**pybind11** is a lightweight header-only library that exposes C++ types in Python
+and vice versa, mainly to create Python bindings of existing C++ code. Its
+goals and syntax are similar to the excellent
+[Boost.Python](http://www.boost.org/doc/libs/1_58_0/libs/python/doc/) library
+by David Abrahams: to minimize boilerplate code in traditional extension
+modules by inferring type information using compile-time introspection.
+
+The main issue with Boost.Python—and the reason for creating such a similar
+project—is Boost. Boost is an enormously large and complex suite of utility
+libraries that works with almost every C++ compiler in existence. This
+compatibility has its cost: arcane template tricks and workarounds are
+necessary to support the oldest and buggiest of compiler specimens. Now that
+C++11-compatible compilers are widely available, this heavy machinery has
+become an excessively large and unnecessary dependency.
+
+Think of this library as a tiny self-contained version of Boost.Python with
+everything stripped away that isn't relevant for binding generation. Without
+comments, the core header files only require ~4K lines of code and depend on
+Python (2.7 or 3.x, or PyPy2.7 >= 5.7) and the C++ standard library. This
+compact implementation was possible thanks to some of the new C++11 language
+features (specifically: tuples, lambda functions and variadic templates). Since
+its creation, this library has grown beyond Boost.Python in many ways, leading
+to dramatically simpler binding code in many common situations.
+
+Tutorial and reference documentation is provided at
+[http://pybind11.readthedocs.org/en/master](http://pybind11.readthedocs.org/en/master).
+A PDF version of the manual is available
+[here](https://media.readthedocs.org/pdf/pybind11/master/pybind11.pdf).
+
+## Core features
+pybind11 can map the following core C++ features to Python
+
+- Functions accepting and returning custom data structures per value, reference, or pointer
+- Instance methods and static methods
+- Overloaded functions
+- Instance attributes and static attributes
+- Arbitrary exception types
+- Enumerations
+- Callbacks
+- Iterators and ranges
+- Custom operators
+- Single and multiple inheritance
+- STL data structures
+- Iterators and ranges
+- Smart pointers with reference counting like ``std::shared_ptr``
+- Internal references with correct reference counting
+- C++ classes with virtual (and pure virtual) methods can be extended in Python
+
+## Goodies
+In addition to the core functionality, pybind11 provides some extra goodies:
+
+- Python 2.7, 3.x, and PyPy (PyPy2.7 >= 5.7) are supported with an
+  implementation-agnostic interface.
+
+- It is possible to bind C++11 lambda functions with captured variables. The
+  lambda capture data is stored inside the resulting Python function object.
+
+- pybind11 uses C++11 move constructors and move assignment operators whenever
+  possible to efficiently transfer custom data types.
+
+- It's easy to expose the internal storage of custom data types through
+  Pythons' buffer protocols. This is handy e.g. for fast conversion between
+  C++ matrix classes like Eigen and NumPy without expensive copy operations.
+
+- pybind11 can automatically vectorize functions so that they are transparently
+  applied to all entries of one or more NumPy array arguments.
+
+- Python's slice-based access and assignment operations can be supported with
+  just a few lines of code.
+
+- Everything is contained in just a few header files; there is no need to link
+  against any additional libraries.
+
+- Binaries are generally smaller by a factor of at least 2 compared to
+  equivalent bindings generated by Boost.Python. A recent pybind11 conversion
+  of PyRosetta, an enormous Boost.Python binding project,
+  [reported](http://graylab.jhu.edu/RosettaCon2016/PyRosetta-4.pdf) a binary
+  size reduction of **5.4x** and compile time reduction by **5.8x**.
+
+- When supported by the compiler, two new C++14 features (relaxed constexpr and
+  return value deduction) are used to precompute function signatures at compile
+  time, leading to smaller binaries.
+
+- With little extra effort, C++ types can be pickled and unpickled similar to
+  regular Python objects.
+
+## Supported compilers
+
+1. Clang/LLVM 3.3 or newer (for Apple Xcode's clang, this is 5.0.0 or newer)
+2. GCC 4.8 or newer
+3. Microsoft Visual Studio 2015 Update 3 or newer
+4. Intel C++ compiler 16 or newer (15 with a [workaround](https://github.com/pybind/pybind11/issues/276))
+5. Cygwin/GCC (tested on 2.5.1)
+
+## About
+
+This project was created by [Wenzel Jakob](http://rgl.epfl.ch/people/wjakob).
+Significant features and/or improvements to the code were contributed by
+Jonas Adler,
+Sylvain Corlay,
+Trent Houliston,
+Axel Huebl,
+@hulucc,
+Sergey Lyskov
+Johan Mabille,
+Tomasz Miąsko,
+Dean Moldovan,
+Ben Pritchard,
+Jason Rhinelander,
+Boris Schäling,
+Pim Schellart,
+Ivan Smirnov, and
+Patrick Stewart.
+
+### License
+
+pybind11 is provided under a BSD-style license that can be found in the
+``LICENSE`` file. By using, distributing, or contributing to this project,
+you agree to the terms and conditions of this license.
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/attr.h b/ml/dlib/dlib/external/pybind11/include/pybind11/attr.h
new file mode 100644
index 000000000..dce875a6b
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/attr.h
@@ -0,0 +1,489 @@
+/*
+    pybind11/attr.h: Infrastructure for processing custom
+    type and function attributes
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "cast.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+/// \addtogroup annotations
+/// @{
+
+/// Annotation for methods
+struct is_method { handle class_; is_method(const handle &c) : class_(c) { } };
+
+/// Annotation for operators
+struct is_operator { };
+
+/// Annotation for parent scope
+struct scope { handle value; scope(const handle &s) : value(s) { } };
+
+/// Annotation for documentation
+struct doc { const char *value; doc(const char *value) : value(value) { } };
+
+/// Annotation for function names
+struct name { const char *value; name(const char *value) : value(value) { } };
+
+/// Annotation indicating that a function is an overload associated with a given "sibling"
+struct sibling { handle value; sibling(const handle &value) : value(value.ptr()) { } };
+
+/// Annotation indicating that a class derives from another given type
+template <typename T> struct base {
+    PYBIND11_DEPRECATED("base<T>() was deprecated in favor of specifying 'T' as a template argument to class_")
+    base() { }
+};
+
+/// Keep patient alive while nurse lives
+template <size_t Nurse, size_t Patient> struct keep_alive { };
+
+/// Annotation indicating that a class is involved in a multiple inheritance relationship
+struct multiple_inheritance { };
+
+/// Annotation which enables dynamic attributes, i.e. adds `__dict__` to a class
+struct dynamic_attr { };
+
+/// Annotation which enables the buffer protocol for a type
+struct buffer_protocol { };
+
+/// Annotation which requests that a special metaclass is created for a type
+struct metaclass {
+    handle value;
+
+    PYBIND11_DEPRECATED("py::metaclass() is no longer required. It's turned on by default now.")
+    metaclass() {}
+
+    /// Override pybind11's default metaclass
+    explicit metaclass(handle value) : value(value) { }
+};
+
+/// Annotation that marks a class as local to the module:
+struct module_local { const bool value; constexpr module_local(bool v = true) : value(v) { } };
+
+/// Annotation to mark enums as an arithmetic type
+struct arithmetic { };
+
+/** \rst
+    A call policy which places one or more guard variables (``Ts...``) around the function call.
+
+    For example, this definition:
+
+    .. code-block:: cpp
+
+        m.def("foo", foo, py::call_guard<T>());
+
+    is equivalent to the following pseudocode:
+
+    .. code-block:: cpp
+
+        m.def("foo", [](args...) {
+            T scope_guard;
+            return foo(args...); // forwarded arguments
+        });
+ \endrst */
+template <typename... Ts> struct call_guard;
+
+template <> struct call_guard<> { using type = detail::void_type; };
+
+template <typename T>
+struct call_guard<T> {
+    static_assert(std::is_default_constructible<T>::value,
+                  "The guard type must be default constructible");
+
+    using type = T;
+};
+
+template <typename T, typename... Ts>
+struct call_guard<T, Ts...> {
+    struct type {
+        T guard{}; // Compose multiple guard types with left-to-right default-constructor order
+        typename call_guard<Ts...>::type next{};
+    };
+};
+
+/// @} annotations
+
+NAMESPACE_BEGIN(detail)
+/* Forward declarations */
+enum op_id : int;
+enum op_type : int;
+struct undefined_t;
+template <op_id id, op_type ot, typename L = undefined_t, typename R = undefined_t> struct op_;
+inline void keep_alive_impl(size_t Nurse, size_t Patient, function_call &call, handle ret);
+
+/// Internal data structure which holds metadata about a keyword argument
+struct argument_record {
+    const char *name;  ///< Argument name
+    const char *descr; ///< Human-readable version of the argument value
+    handle value;      ///< Associated Python object
+    bool convert : 1;  ///< True if the argument is allowed to convert when loading
+    bool none : 1;     ///< True if None is allowed when loading
+
+    argument_record(const char *name, const char *descr, handle value, bool convert, bool none)
+        : name(name), descr(descr), value(value), convert(convert), none(none) { }
+};
+
+/// Internal data structure which holds metadata about a bound function (signature, overloads, etc.)
+struct function_record {
+    function_record()
+        : is_constructor(false), is_new_style_constructor(false), is_stateless(false),
+          is_operator(false), has_args(false), has_kwargs(false), is_method(false) { }
+
+    /// Function name
+    char *name = nullptr; /* why no C++ strings? They generate heavier code.. */
+
+    // User-specified documentation string
+    char *doc = nullptr;
+
+    /// Human-readable version of the function signature
+    char *signature = nullptr;
+
+    /// List of registered keyword arguments
+    std::vector<argument_record> args;
+
+    /// Pointer to lambda function which converts arguments and performs the actual call
+    handle (*impl) (function_call &) = nullptr;
+
+    /// Storage for the wrapped function pointer and captured data, if any
+    void *data[3] = { };
+
+    /// Pointer to custom destructor for 'data' (if needed)
+    void (*free_data) (function_record *ptr) = nullptr;
+
+    /// Return value policy associated with this function
+    return_value_policy policy = return_value_policy::automatic;
+
+    /// True if name == '__init__'
+    bool is_constructor : 1;
+
+    /// True if this is a new-style `__init__` defined in `detail/init.h`
+    bool is_new_style_constructor : 1;
+
+    /// True if this is a stateless function pointer
+    bool is_stateless : 1;
+
+    /// True if this is an operator (__add__), etc.
+    bool is_operator : 1;
+
+    /// True if the function has a '*args' argument
+    bool has_args : 1;
+
+    /// True if the function has a '**kwargs' argument
+    bool has_kwargs : 1;
+
+    /// True if this is a method
+    bool is_method : 1;
+
+    /// Number of arguments (including py::args and/or py::kwargs, if present)
+    std::uint16_t nargs;
+
+    /// Python method object
+    PyMethodDef *def = nullptr;
+
+    /// Python handle to the parent scope (a class or a module)
+    handle scope;
+
+    /// Python handle to the sibling function representing an overload chain
+    handle sibling;
+
+    /// Pointer to next overload
+    function_record *next = nullptr;
+};
+
+/// Special data structure which (temporarily) holds metadata about a bound class
+struct type_record {
+    PYBIND11_NOINLINE type_record()
+        : multiple_inheritance(false), dynamic_attr(false), buffer_protocol(false), module_local(false) { }
+
+    /// Handle to the parent scope
+    handle scope;
+
+    /// Name of the class
+    const char *name = nullptr;
+
+    // Pointer to RTTI type_info data structure
+    const std::type_info *type = nullptr;
+
+    /// How large is the underlying C++ type?
+    size_t type_size = 0;
+
+    /// How large is the type's holder?
+    size_t holder_size = 0;
+
+    /// The global operator new can be overridden with a class-specific variant
+    void *(*operator_new)(size_t) = ::operator new;
+
+    /// Function pointer to class_<..>::init_instance
+    void (*init_instance)(instance *, const void *) = nullptr;
+
+    /// Function pointer to class_<..>::dealloc
+    void (*dealloc)(detail::value_and_holder &) = nullptr;
+
+    /// List of base classes of the newly created type
+    list bases;
+
+    /// Optional docstring
+    const char *doc = nullptr;
+
+    /// Custom metaclass (optional)
+    handle metaclass;
+
+    /// Multiple inheritance marker
+    bool multiple_inheritance : 1;
+
+    /// Does the class manage a __dict__?
+    bool dynamic_attr : 1;
+
+    /// Does the class implement the buffer protocol?
+    bool buffer_protocol : 1;
+
+    /// Is the default (unique_ptr) holder type used?
+    bool default_holder : 1;
+
+    /// Is the class definition local to the module shared object?
+    bool module_local : 1;
+
+    PYBIND11_NOINLINE void add_base(const std::type_info &base, void *(*caster)(void *)) {
+        auto base_info = detail::get_type_info(base, false);
+        if (!base_info) {
+            std::string tname(base.name());
+            detail::clean_type_id(tname);
+            pybind11_fail("generic_type: type \"" + std::string(name) +
+                          "\" referenced unknown base type \"" + tname + "\"");
+        }
+
+        if (default_holder != base_info->default_holder) {
+            std::string tname(base.name());
+            detail::clean_type_id(tname);
+            pybind11_fail("generic_type: type \"" + std::string(name) + "\" " +
+                    (default_holder ? "does not have" : "has") +
+                    " a non-default holder type while its base \"" + tname + "\" " +
+                    (base_info->default_holder ? "does not" : "does"));
+        }
+
+        bases.append((PyObject *) base_info->type);
+
+        if (base_info->type->tp_dictoffset != 0)
+            dynamic_attr = true;
+
+        if (caster)
+            base_info->implicit_casts.emplace_back(type, caster);
+    }
+};
+
+inline function_call::function_call(function_record &f, handle p) :
+        func(f), parent(p) {
+    args.reserve(f.nargs);
+    args_convert.reserve(f.nargs);
+}
+
+/// Tag for a new-style `__init__` defined in `detail/init.h`
+struct is_new_style_constructor { };
+
+/**
+ * Partial template specializations to process custom attributes provided to
+ * cpp_function_ and class_. These are either used to initialize the respective
+ * fields in the type_record and function_record data structures or executed at
+ * runtime to deal with custom call policies (e.g. keep_alive).
+ */
+template <typename T, typename SFINAE = void> struct process_attribute;
+
+template <typename T> struct process_attribute_default {
+    /// Default implementation: do nothing
+    static void init(const T &, function_record *) { }
+    static void init(const T &, type_record *) { }
+    static void precall(function_call &) { }
+    static void postcall(function_call &, handle) { }
+};
+
+/// Process an attribute specifying the function's name
+template <> struct process_attribute<name> : process_attribute_default<name> {
+    static void init(const name &n, function_record *r) { r->name = const_cast<char *>(n.value); }
+};
+
+/// Process an attribute specifying the function's docstring
+template <> struct process_attribute<doc> : process_attribute_default<doc> {
+    static void init(const doc &n, function_record *r) { r->doc = const_cast<char *>(n.value); }
+};
+
+/// Process an attribute specifying the function's docstring (provided as a C-style string)
+template <> struct process_attribute<const char *> : process_attribute_default<const char *> {
+    static void init(const char *d, function_record *r) { r->doc = const_cast<char *>(d); }
+    static void init(const char *d, type_record *r) { r->doc = const_cast<char *>(d); }
+};
+template <> struct process_attribute<char *> : process_attribute<const char *> { };
+
+/// Process an attribute indicating the function's return value policy
+template <> struct process_attribute<return_value_policy> : process_attribute_default<return_value_policy> {
+    static void init(const return_value_policy &p, function_record *r) { r->policy = p; }
+};
+
+/// Process an attribute which indicates that this is an overloaded function associated with a given sibling
+template <> struct process_attribute<sibling> : process_attribute_default<sibling> {
+    static void init(const sibling &s, function_record *r) { r->sibling = s.value; }
+};
+
+/// Process an attribute which indicates that this function is a method
+template <> struct process_attribute<is_method> : process_attribute_default<is_method> {
+    static void init(const is_method &s, function_record *r) { r->is_method = true; r->scope = s.class_; }
+};
+
+/// Process an attribute which indicates the parent scope of a method
+template <> struct process_attribute<scope> : process_attribute_default<scope> {
+    static void init(const scope &s, function_record *r) { r->scope = s.value; }
+};
+
+/// Process an attribute which indicates that this function is an operator
+template <> struct process_attribute<is_operator> : process_attribute_default<is_operator> {
+    static void init(const is_operator &, function_record *r) { r->is_operator = true; }
+};
+
+template <> struct process_attribute<is_new_style_constructor> : process_attribute_default<is_new_style_constructor> {
+    static void init(const is_new_style_constructor &, function_record *r) { r->is_new_style_constructor = true; }
+};
+
+/// Process a keyword argument attribute (*without* a default value)
+template <> struct process_attribute<arg> : process_attribute_default<arg> {
+    static void init(const arg &a, function_record *r) {
+        if (r->is_method && r->args.empty())
+            r->args.emplace_back("self", nullptr, handle(), true /*convert*/, false /*none not allowed*/);
+        r->args.emplace_back(a.name, nullptr, handle(), !a.flag_noconvert, a.flag_none);
+    }
+};
+
+/// Process a keyword argument attribute (*with* a default value)
+template <> struct process_attribute<arg_v> : process_attribute_default<arg_v> {
+    static void init(const arg_v &a, function_record *r) {
+        if (r->is_method && r->args.empty())
+            r->args.emplace_back("self", nullptr /*descr*/, handle() /*parent*/, true /*convert*/, false /*none not allowed*/);
+
+        if (!a.value) {
+#if !defined(NDEBUG)
+            std::string descr("'");
+            if (a.name) descr += std::string(a.name) + ": ";
+            descr += a.type + "'";
+            if (r->is_method) {
+                if (r->name)
+                    descr += " in method '" + (std::string) str(r->scope) + "." + (std::string) r->name + "'";
+                else
+                    descr += " in method of '" + (std::string) str(r->scope) + "'";
+            } else if (r->name) {
+                descr += " in function '" + (std::string) r->name + "'";
+            }
+            pybind11_fail("arg(): could not convert default argument "
+                          + descr + " into a Python object (type not registered yet?)");
+#else
+            pybind11_fail("arg(): could not convert default argument "
+                          "into a Python object (type not registered yet?). "
+                          "Compile in debug mode for more information.");
+#endif
+        }
+        r->args.emplace_back(a.name, a.descr, a.value.inc_ref(), !a.flag_noconvert, a.flag_none);
+    }
+};
+
+/// Process a parent class attribute.  Single inheritance only (class_ itself already guarantees that)
+template <typename T>
+struct process_attribute<T, enable_if_t<is_pyobject<T>::value>> : process_attribute_default<handle> {
+    static void init(const handle &h, type_record *r) { r->bases.append(h); }
+};
+
+/// Process a parent class attribute (deprecated, does not support multiple inheritance)
+template <typename T>
+struct process_attribute<base<T>> : process_attribute_default<base<T>> {
+    static void init(const base<T> &, type_record *r) { r->add_base(typeid(T), nullptr); }
+};
+
+/// Process a multiple inheritance attribute
+template <>
+struct process_attribute<multiple_inheritance> : process_attribute_default<multiple_inheritance> {
+    static void init(const multiple_inheritance &, type_record *r) { r->multiple_inheritance = true; }
+};
+
+template <>
+struct process_attribute<dynamic_attr> : process_attribute_default<dynamic_attr> {
+    static void init(const dynamic_attr &, type_record *r) { r->dynamic_attr = true; }
+};
+
+template <>
+struct process_attribute<buffer_protocol> : process_attribute_default<buffer_protocol> {
+    static void init(const buffer_protocol &, type_record *r) { r->buffer_protocol = true; }
+};
+
+template <>
+struct process_attribute<metaclass> : process_attribute_default<metaclass> {
+    static void init(const metaclass &m, type_record *r) { r->metaclass = m.value; }
+};
+
+template <>
+struct process_attribute<module_local> : process_attribute_default<module_local> {
+    static void init(const module_local &l, type_record *r) { r->module_local = l.value; }
+};
+
+/// Process an 'arithmetic' attribute for enums (does nothing here)
+template <>
+struct process_attribute<arithmetic> : process_attribute_default<arithmetic> {};
+
+template <typename... Ts>
+struct process_attribute<call_guard<Ts...>> : process_attribute_default<call_guard<Ts...>> { };
+
+/**
+ * Process a keep_alive call policy -- invokes keep_alive_impl during the
+ * pre-call handler if both Nurse, Patient != 0 and use the post-call handler
+ * otherwise
+ */
+template <size_t Nurse, size_t Patient> struct process_attribute<keep_alive<Nurse, Patient>> : public process_attribute_default<keep_alive<Nurse, Patient>> {
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N != 0 && P != 0, int> = 0>
+    static void precall(function_call &call) { keep_alive_impl(Nurse, Patient, call, handle()); }
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N != 0 && P != 0, int> = 0>
+    static void postcall(function_call &, handle) { }
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N == 0 || P == 0, int> = 0>
+    static void precall(function_call &) { }
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N == 0 || P == 0, int> = 0>
+    static void postcall(function_call &call, handle ret) { keep_alive_impl(Nurse, Patient, call, ret); }
+};
+
+/// Recursively iterate over variadic template arguments
+template <typename... Args> struct process_attributes {
+    static void init(const Args&... args, function_record *r) {
+        int unused[] = { 0, (process_attribute<typename std::decay<Args>::type>::init(args, r), 0) ... };
+        ignore_unused(unused);
+    }
+    static void init(const Args&... args, type_record *r) {
+        int unused[] = { 0, (process_attribute<typename std::decay<Args>::type>::init(args, r), 0) ... };
+        ignore_unused(unused);
+    }
+    static void precall(function_call &call) {
+        int unused[] = { 0, (process_attribute<typename std::decay<Args>::type>::precall(call), 0) ... };
+        ignore_unused(unused);
+    }
+    static void postcall(function_call &call, handle fn_ret) {
+        int unused[] = { 0, (process_attribute<typename std::decay<Args>::type>::postcall(call, fn_ret), 0) ... };
+        ignore_unused(unused);
+    }
+};
+
+template <typename T>
+using is_call_guard = is_instantiation<call_guard, T>;
+
+/// Extract the ``type`` from the first `call_guard` in `Extras...` (or `void_type` if none found)
+template <typename... Extra>
+using extract_guard_t = typename exactly_one_t<is_call_guard, call_guard<>, Extra...>::type;
+
+/// Check the number of named arguments at compile time
+template <typename... Extra,
+          size_t named = constexpr_sum(std::is_base_of<arg, Extra>::value...),
+          size_t self  = constexpr_sum(std::is_same<is_method, Extra>::value...)>
+constexpr bool expected_num_args(size_t nargs, bool has_args, bool has_kwargs) {
+    return named == 0 || (self + named + has_args + has_kwargs) == nargs;
+}
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/buffer_info.h b/ml/dlib/dlib/external/pybind11/include/pybind11/buffer_info.h
new file mode 100644
index 000000000..9f072fa73
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/buffer_info.h
@@ -0,0 +1,108 @@
+/*
+    pybind11/buffer_info.h: Python buffer object interface
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+/// Information record describing a Python buffer object
+struct buffer_info {
+    void *ptr = nullptr;          // Pointer to the underlying storage
+    ssize_t itemsize = 0;         // Size of individual items in bytes
+    ssize_t size = 0;             // Total number of entries
+    std::string format;           // For homogeneous buffers, this should be set to format_descriptor<T>::format()
+    ssize_t ndim = 0;             // Number of dimensions
+    std::vector<ssize_t> shape;   // Shape of the tensor (1 entry per dimension)
+    std::vector<ssize_t> strides; // Number of entries between adjacent entries (for each per dimension)
+
+    buffer_info() { }
+
+    buffer_info(void *ptr, ssize_t itemsize, const std::string &format, ssize_t ndim,
+                detail::any_container<ssize_t> shape_in, detail::any_container<ssize_t> strides_in)
+    : ptr(ptr), itemsize(itemsize), size(1), format(format), ndim(ndim),
+      shape(std::move(shape_in)), strides(std::move(strides_in)) {
+        if (ndim != (ssize_t) shape.size() || ndim != (ssize_t) strides.size())
+            pybind11_fail("buffer_info: ndim doesn't match shape and/or strides length");
+        for (size_t i = 0; i < (size_t) ndim; ++i)
+            size *= shape[i];
+    }
+
+    template <typename T>
+    buffer_info(T *ptr, detail::any_container<ssize_t> shape_in, detail::any_container<ssize_t> strides_in)
+    : buffer_info(private_ctr_tag(), ptr, sizeof(T), format_descriptor<T>::format(), static_cast<ssize_t>(shape_in->size()), std::move(shape_in), std::move(strides_in)) { }
+
+    buffer_info(void *ptr, ssize_t itemsize, const std::string &format, ssize_t size)
+    : buffer_info(ptr, itemsize, format, 1, {size}, {itemsize}) { }
+
+    template <typename T>
+    buffer_info(T *ptr, ssize_t size)
+    : buffer_info(ptr, sizeof(T), format_descriptor<T>::format(), size) { }
+
+    explicit buffer_info(Py_buffer *view, bool ownview = true)
+    : buffer_info(view->buf, view->itemsize, view->format, view->ndim,
+            {view->shape, view->shape + view->ndim}, {view->strides, view->strides + view->ndim}) {
+        this->view = view;
+        this->ownview = ownview;
+    }
+
+    buffer_info(const buffer_info &) = delete;
+    buffer_info& operator=(const buffer_info &) = delete;
+
+    buffer_info(buffer_info &&other) {
+        (*this) = std::move(other);
+    }
+
+    buffer_info& operator=(buffer_info &&rhs) {
+        ptr = rhs.ptr;
+        itemsize = rhs.itemsize;
+        size = rhs.size;
+        format = std::move(rhs.format);
+        ndim = rhs.ndim;
+        shape = std::move(rhs.shape);
+        strides = std::move(rhs.strides);
+        std::swap(view, rhs.view);
+        std::swap(ownview, rhs.ownview);
+        return *this;
+    }
+
+    ~buffer_info() {
+        if (view && ownview) { PyBuffer_Release(view); delete view; }
+    }
+
+private:
+    struct private_ctr_tag { };
+
+    buffer_info(private_ctr_tag, void *ptr, ssize_t itemsize, const std::string &format, ssize_t ndim,
+                detail::any_container<ssize_t> &&shape_in, detail::any_container<ssize_t> &&strides_in)
+    : buffer_info(ptr, itemsize, format, ndim, std::move(shape_in), std::move(strides_in)) { }
+
+    Py_buffer *view = nullptr;
+    bool ownview = false;
+};
+
+NAMESPACE_BEGIN(detail)
+
+template <typename T, typename SFINAE = void> struct compare_buffer_info {
+    static bool compare(const buffer_info& b) {
+        return b.format == format_descriptor<T>::format() && b.itemsize == (ssize_t) sizeof(T);
+    }
+};
+
+template <typename T> struct compare_buffer_info<T, detail::enable_if_t<std::is_integral<T>::value>> {
+    static bool compare(const buffer_info& b) {
+        return (size_t) b.itemsize == sizeof(T) && (b.format == format_descriptor<T>::value ||
+            ((sizeof(T) == sizeof(long)) && b.format == (std::is_unsigned<T>::value ? "L" : "l")) ||
+            ((sizeof(T) == sizeof(size_t)) && b.format == (std::is_unsigned<T>::value ? "N" : "n")));
+    }
+};
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/cast.h b/ml/dlib/dlib/external/pybind11/include/pybind11/cast.h
new file mode 100644
index 000000000..a722a9e81
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/cast.h
@@ -0,0 +1,2063 @@
+/*
+    pybind11/cast.h: Partial template specializations to cast between
+    C++ and Python types
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pytypes.h"
+#include "detail/typeid.h"
+#include "detail/descr.h"
+#include "detail/internals.h"
+#include <array>
+#include <limits>
+#include <tuple>
+
+#if defined(PYBIND11_CPP17)
+#  if defined(__has_include)
+#    if __has_include(<string_view>)
+#      define PYBIND11_HAS_STRING_VIEW
+#    endif
+#  elif defined(_MSC_VER)
+#    define PYBIND11_HAS_STRING_VIEW
+#  endif
+#endif
+#ifdef PYBIND11_HAS_STRING_VIEW
+#include <string_view>
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/// A life support system for temporary objects created by `type_caster::load()`.
+/// Adding a patient will keep it alive up until the enclosing function returns.
+class loader_life_support {
+public:
+    /// A new patient frame is created when a function is entered
+    loader_life_support() {
+        get_internals().loader_patient_stack.push_back(nullptr);
+    }
+
+    /// ... and destroyed after it returns
+    ~loader_life_support() {
+        auto &stack = get_internals().loader_patient_stack;
+        if (stack.empty())
+            pybind11_fail("loader_life_support: internal error");
+
+        auto ptr = stack.back();
+        stack.pop_back();
+        Py_CLEAR(ptr);
+
+        // A heuristic to reduce the stack's capacity (e.g. after long recursive calls)
+        if (stack.capacity() > 16 && stack.size() != 0 && stack.capacity() / stack.size() > 2)
+            stack.shrink_to_fit();
+    }
+
+    /// This can only be used inside a pybind11-bound function, either by `argument_loader`
+    /// at argument preparation time or by `py::cast()` at execution time.
+    PYBIND11_NOINLINE static void add_patient(handle h) {
+        auto &stack = get_internals().loader_patient_stack;
+        if (stack.empty())
+            throw cast_error("When called outside a bound function, py::cast() cannot "
+                             "do Python -> C++ conversions which require the creation "
+                             "of temporary values");
+
+        auto &list_ptr = stack.back();
+        if (list_ptr == nullptr) {
+            list_ptr = PyList_New(1);
+            if (!list_ptr)
+                pybind11_fail("loader_life_support: error allocating list");
+            PyList_SET_ITEM(list_ptr, 0, h.inc_ref().ptr());
+        } else {
+            auto result = PyList_Append(list_ptr, h.ptr());
+            if (result == -1)
+                pybind11_fail("loader_life_support: error adding patient");
+        }
+    }
+};
+
+// Gets the cache entry for the given type, creating it if necessary.  The return value is the pair
+// returned by emplace, i.e. an iterator for the entry and a bool set to `true` if the entry was
+// just created.
+inline std::pair<decltype(internals::registered_types_py)::iterator, bool> all_type_info_get_cache(PyTypeObject *type);
+
+// Populates a just-created cache entry.
+PYBIND11_NOINLINE inline void all_type_info_populate(PyTypeObject *t, std::vector<type_info *> &bases) {
+    std::vector<PyTypeObject *> check;
+    for (handle parent : reinterpret_borrow<tuple>(t->tp_bases))
+        check.push_back((PyTypeObject *) parent.ptr());
+
+    auto const &type_dict = get_internals().registered_types_py;
+    for (size_t i = 0; i < check.size(); i++) {
+        auto type = check[i];
+        // Ignore Python2 old-style class super type:
+        if (!PyType_Check((PyObject *) type)) continue;
+
+        // Check `type` in the current set of registered python types:
+        auto it = type_dict.find(type);
+        if (it != type_dict.end()) {
+            // We found a cache entry for it, so it's either pybind-registered or has pre-computed
+            // pybind bases, but we have to make sure we haven't already seen the type(s) before: we
+            // want to follow Python/virtual C++ rules that there should only be one instance of a
+            // common base.
+            for (auto *tinfo : it->second) {
+                // NB: Could use a second set here, rather than doing a linear search, but since
+                // having a large number of immediate pybind11-registered types seems fairly
+                // unlikely, that probably isn't worthwhile.
+                bool found = false;
+                for (auto *known : bases) {
+                    if (known == tinfo) { found = true; break; }
+                }
+                if (!found) bases.push_back(tinfo);
+            }
+        }
+        else if (type->tp_bases) {
+            // It's some python type, so keep follow its bases classes to look for one or more
+            // registered types
+            if (i + 1 == check.size()) {
+                // When we're at the end, we can pop off the current element to avoid growing
+                // `check` when adding just one base (which is typical--i.e. when there is no
+                // multiple inheritance)
+                check.pop_back();
+                i--;
+            }
+            for (handle parent : reinterpret_borrow<tuple>(type->tp_bases))
+                check.push_back((PyTypeObject *) parent.ptr());
+        }
+    }
+}
+
+/**
+ * Extracts vector of type_info pointers of pybind-registered roots of the given Python type.  Will
+ * be just 1 pybind type for the Python type of a pybind-registered class, or for any Python-side
+ * derived class that uses single inheritance.  Will contain as many types as required for a Python
+ * class that uses multiple inheritance to inherit (directly or indirectly) from multiple
+ * pybind-registered classes.  Will be empty if neither the type nor any base classes are
+ * pybind-registered.
+ *
+ * The value is cached for the lifetime of the Python type.
+ */
+inline const std::vector<detail::type_info *> &all_type_info(PyTypeObject *type) {
+    auto ins = all_type_info_get_cache(type);
+    if (ins.second)
+        // New cache entry: populate it
+        all_type_info_populate(type, ins.first->second);
+
+    return ins.first->second;
+}
+
+/**
+ * Gets a single pybind11 type info for a python type.  Returns nullptr if neither the type nor any
+ * ancestors are pybind11-registered.  Throws an exception if there are multiple bases--use
+ * `all_type_info` instead if you want to support multiple bases.
+ */
+PYBIND11_NOINLINE inline detail::type_info* get_type_info(PyTypeObject *type) {
+    auto &bases = all_type_info(type);
+    if (bases.size() == 0)
+        return nullptr;
+    if (bases.size() > 1)
+        pybind11_fail("pybind11::detail::get_type_info: type has multiple pybind11-registered bases");
+    return bases.front();
+}
+
+inline detail::type_info *get_local_type_info(const std::type_index &tp) {
+    auto &locals = registered_local_types_cpp();
+    auto it = locals.find(tp);
+    if (it != locals.end())
+        return it->second;
+    return nullptr;
+}
+
+inline detail::type_info *get_global_type_info(const std::type_index &tp) {
+    auto &types = get_internals().registered_types_cpp;
+    auto it = types.find(tp);
+    if (it != types.end())
+        return it->second;
+    return nullptr;
+}
+
+/// Return the type info for a given C++ type; on lookup failure can either throw or return nullptr.
+PYBIND11_NOINLINE inline detail::type_info *get_type_info(const std::type_index &tp,
+                                                          bool throw_if_missing = false) {
+    if (auto ltype = get_local_type_info(tp))
+        return ltype;
+    if (auto gtype = get_global_type_info(tp))
+        return gtype;
+
+    if (throw_if_missing) {
+        std::string tname = tp.name();
+        detail::clean_type_id(tname);
+        pybind11_fail("pybind11::detail::get_type_info: unable to find type info for \"" + tname + "\"");
+    }
+    return nullptr;
+}
+
+PYBIND11_NOINLINE inline handle get_type_handle(const std::type_info &tp, bool throw_if_missing) {
+    detail::type_info *type_info = get_type_info(tp, throw_if_missing);
+    return handle(type_info ? ((PyObject *) type_info->type) : nullptr);
+}
+
+struct value_and_holder {
+    instance *inst;
+    size_t index;
+    const detail::type_info *type;
+    void **vh;
+
+    // Main constructor for a found value/holder:
+    value_and_holder(instance *i, const detail::type_info *type, size_t vpos, size_t index) :
+        inst{i}, index{index}, type{type},
+        vh{inst->simple_layout ? inst->simple_value_holder : &inst->nonsimple.values_and_holders[vpos]}
+    {}
+
+    // Default constructor (used to signal a value-and-holder not found by get_value_and_holder())
+    value_and_holder() : inst{nullptr} {}
+
+    // Used for past-the-end iterator
+    value_and_holder(size_t index) : index{index} {}
+
+    template <typename V = void> V *&value_ptr() const {
+        return reinterpret_cast<V *&>(vh[0]);
+    }
+    // True if this `value_and_holder` has a non-null value pointer
+    explicit operator bool() const { return value_ptr(); }
+
+    template <typename H> H &holder() const {
+        return reinterpret_cast<H &>(vh[1]);
+    }
+    bool holder_constructed() const {
+        return inst->simple_layout
+            ? inst->simple_holder_constructed
+            : inst->nonsimple.status[index] & instance::status_holder_constructed;
+    }
+    void set_holder_constructed(bool v = true) {
+        if (inst->simple_layout)
+            inst->simple_holder_constructed = v;
+        else if (v)
+            inst->nonsimple.status[index] |= instance::status_holder_constructed;
+        else
+            inst->nonsimple.status[index] &= (uint8_t) ~instance::status_holder_constructed;
+    }
+    bool instance_registered() const {
+        return inst->simple_layout
+            ? inst->simple_instance_registered
+            : inst->nonsimple.status[index] & instance::status_instance_registered;
+    }
+    void set_instance_registered(bool v = true) {
+        if (inst->simple_layout)
+            inst->simple_instance_registered = v;
+        else if (v)
+            inst->nonsimple.status[index] |= instance::status_instance_registered;
+        else
+            inst->nonsimple.status[index] &= (uint8_t) ~instance::status_instance_registered;
+    }
+};
+
+// Container for accessing and iterating over an instance's values/holders
+struct values_and_holders {
+private:
+    instance *inst;
+    using type_vec = std::vector<detail::type_info *>;
+    const type_vec &tinfo;
+
+public:
+    values_and_holders(instance *inst) : inst{inst}, tinfo(all_type_info(Py_TYPE(inst))) {}
+
+    struct iterator {
+    private:
+        instance *inst;
+        const type_vec *types;
+        value_and_holder curr;
+        friend struct values_and_holders;
+        iterator(instance *inst, const type_vec *tinfo)
+            : inst{inst}, types{tinfo},
+            curr(inst /* instance */,
+                 types->empty() ? nullptr : (*types)[0] /* type info */,
+                 0, /* vpos: (non-simple types only): the first vptr comes first */
+                 0 /* index */)
+        {}
+        // Past-the-end iterator:
+        iterator(size_t end) : curr(end) {}
+    public:
+        bool operator==(const iterator &other) { return curr.index == other.curr.index; }
+        bool operator!=(const iterator &other) { return curr.index != other.curr.index; }
+        iterator &operator++() {
+            if (!inst->simple_layout)
+                curr.vh += 1 + (*types)[curr.index]->holder_size_in_ptrs;
+            ++curr.index;
+            curr.type = curr.index < types->size() ? (*types)[curr.index] : nullptr;
+            return *this;
+        }
+        value_and_holder &operator*() { return curr; }
+        value_and_holder *operator->() { return &curr; }
+    };
+
+    iterator begin() { return iterator(inst, &tinfo); }
+    iterator end() { return iterator(tinfo.size()); }
+
+    iterator find(const type_info *find_type) {
+        auto it = begin(), endit = end();
+        while (it != endit && it->type != find_type) ++it;
+        return it;
+    }
+
+    size_t size() { return tinfo.size(); }
+};
+
+/**
+ * Extracts C++ value and holder pointer references from an instance (which may contain multiple
+ * values/holders for python-side multiple inheritance) that match the given type.  Throws an error
+ * if the given type (or ValueType, if omitted) is not a pybind11 base of the given instance.  If
+ * `find_type` is omitted (or explicitly specified as nullptr) the first value/holder are returned,
+ * regardless of type (and the resulting .type will be nullptr).
+ *
+ * The returned object should be short-lived: in particular, it must not outlive the called-upon
+ * instance.
+ */
+PYBIND11_NOINLINE inline value_and_holder instance::get_value_and_holder(const type_info *find_type /*= nullptr default in common.h*/, bool throw_if_missing /*= true in common.h*/) {
+    // Optimize common case:
+    if (!find_type || Py_TYPE(this) == find_type->type)
+        return value_and_holder(this, find_type, 0, 0);
+
+    detail::values_and_holders vhs(this);
+    auto it = vhs.find(find_type);
+    if (it != vhs.end())
+        return *it;
+
+    if (!throw_if_missing)
+        return value_and_holder();
+
+#if defined(NDEBUG)
+    pybind11_fail("pybind11::detail::instance::get_value_and_holder: "
+            "type is not a pybind11 base of the given instance "
+            "(compile in debug mode for type details)");
+#else
+    pybind11_fail("pybind11::detail::instance::get_value_and_holder: `" +
+            std::string(find_type->type->tp_name) + "' is not a pybind11 base of the given `" +
+            std::string(Py_TYPE(this)->tp_name) + "' instance");
+#endif
+}
+
+PYBIND11_NOINLINE inline void instance::allocate_layout() {
+    auto &tinfo = all_type_info(Py_TYPE(this));
+
+    const size_t n_types = tinfo.size();
+
+    if (n_types == 0)
+        pybind11_fail("instance allocation failed: new instance has no pybind11-registered base types");
+
+    simple_layout =
+        n_types == 1 && tinfo.front()->holder_size_in_ptrs <= instance_simple_holder_in_ptrs();
+
+    // Simple path: no python-side multiple inheritance, and a small-enough holder
+    if (simple_layout) {
+        simple_value_holder[0] = nullptr;
+        simple_holder_constructed = false;
+        simple_instance_registered = false;
+    }
+    else { // multiple base types or a too-large holder
+        // Allocate space to hold: [v1*][h1][v2*][h2]...[bb...] where [vN*] is a value pointer,
+        // [hN] is the (uninitialized) holder instance for value N, and [bb...] is a set of bool
+        // values that tracks whether each associated holder has been initialized.  Each [block] is
+        // padded, if necessary, to an integer multiple of sizeof(void *).
+        size_t space = 0;
+        for (auto t : tinfo) {
+            space += 1; // value pointer
+            space += t->holder_size_in_ptrs; // holder instance
+        }
+        size_t flags_at = space;
+        space += size_in_ptrs(n_types); // status bytes (holder_constructed and instance_registered)
+
+        // Allocate space for flags, values, and holders, and initialize it to 0 (flags and values,
+        // in particular, need to be 0).  Use Python's memory allocation functions: in Python 3.6
+        // they default to using pymalloc, which is designed to be efficient for small allocations
+        // like the one we're doing here; in earlier versions (and for larger allocations) they are
+        // just wrappers around malloc.
+#if PY_VERSION_HEX >= 0x03050000
+        nonsimple.values_and_holders = (void **) PyMem_Calloc(space, sizeof(void *));
+        if (!nonsimple.values_and_holders) throw std::bad_alloc();
+#else
+        nonsimple.values_and_holders = (void **) PyMem_New(void *, space);
+        if (!nonsimple.values_and_holders) throw std::bad_alloc();
+        std::memset(nonsimple.values_and_holders, 0, space * sizeof(void *));
+#endif
+        nonsimple.status = reinterpret_cast<uint8_t *>(&nonsimple.values_and_holders[flags_at]);
+    }
+    owned = true;
+}
+
+PYBIND11_NOINLINE inline void instance::deallocate_layout() {
+    if (!simple_layout)
+        PyMem_Free(nonsimple.values_and_holders);
+}
+
+PYBIND11_NOINLINE inline bool isinstance_generic(handle obj, const std::type_info &tp) {
+    handle type = detail::get_type_handle(tp, false);
+    if (!type)
+        return false;
+    return isinstance(obj, type);
+}
+
+PYBIND11_NOINLINE inline std::string error_string() {
+    if (!PyErr_Occurred()) {
+        PyErr_SetString(PyExc_RuntimeError, "Unknown internal error occurred");
+        return "Unknown internal error occurred";
+    }
+
+    error_scope scope; // Preserve error state
+
+    std::string errorString;
+    if (scope.type) {
+        errorString += handle(scope.type).attr("__name__").cast<std::string>();
+        errorString += ": ";
+    }
+    if (scope.value)
+        errorString += (std::string) str(scope.value);
+
+    PyErr_NormalizeException(&scope.type, &scope.value, &scope.trace);
+
+#if PY_MAJOR_VERSION >= 3
+    if (scope.trace != nullptr)
+        PyException_SetTraceback(scope.value, scope.trace);
+#endif
+
+#if !defined(PYPY_VERSION)
+    if (scope.trace) {
+        PyTracebackObject *trace = (PyTracebackObject *) scope.trace;
+
+        /* Get the deepest trace possible */
+        while (trace->tb_next)
+            trace = trace->tb_next;
+
+        PyFrameObject *frame = trace->tb_frame;
+        errorString += "\n\nAt:\n";
+        while (frame) {
+            int lineno = PyFrame_GetLineNumber(frame);
+            errorString +=
+                "  " + handle(frame->f_code->co_filename).cast<std::string>() +
+                "(" + std::to_string(lineno) + "): " +
+                handle(frame->f_code->co_name).cast<std::string>() + "\n";
+            frame = frame->f_back;
+        }
+    }
+#endif
+
+    return errorString;
+}
+
+PYBIND11_NOINLINE inline handle get_object_handle(const void *ptr, const detail::type_info *type ) {
+    auto &instances = get_internals().registered_instances;
+    auto range = instances.equal_range(ptr);
+    for (auto it = range.first; it != range.second; ++it) {
+        for (auto vh : values_and_holders(it->second)) {
+            if (vh.type == type)
+                return handle((PyObject *) it->second);
+        }
+    }
+    return handle();
+}
+
+inline PyThreadState *get_thread_state_unchecked() {
+#if defined(PYPY_VERSION)
+    return PyThreadState_GET();
+#elif PY_VERSION_HEX < 0x03000000
+    return _PyThreadState_Current;
+#elif PY_VERSION_HEX < 0x03050000
+    return (PyThreadState*) _Py_atomic_load_relaxed(&_PyThreadState_Current);
+#elif PY_VERSION_HEX < 0x03050200
+    return (PyThreadState*) _PyThreadState_Current.value;
+#else
+    return _PyThreadState_UncheckedGet();
+#endif
+}
+
+// Forward declarations
+inline void keep_alive_impl(handle nurse, handle patient);
+inline PyObject *make_new_instance(PyTypeObject *type);
+
+class type_caster_generic {
+public:
+    PYBIND11_NOINLINE type_caster_generic(const std::type_info &type_info)
+        : typeinfo(get_type_info(type_info)), cpptype(&type_info) { }
+
+    type_caster_generic(const type_info *typeinfo)
+        : typeinfo(typeinfo), cpptype(typeinfo ? typeinfo->cpptype : nullptr) { }
+
+    bool load(handle src, bool convert) {
+        return load_impl<type_caster_generic>(src, convert);
+    }
+
+    PYBIND11_NOINLINE static handle cast(const void *_src, return_value_policy policy, handle parent,
+                                         const detail::type_info *tinfo,
+                                         void *(*copy_constructor)(const void *),
+                                         void *(*move_constructor)(const void *),
+                                         const void *existing_holder = nullptr) {
+        if (!tinfo) // no type info: error will be set already
+            return handle();
+
+        void *src = const_cast<void *>(_src);
+        if (src == nullptr)
+            return none().release();
+
+        auto it_instances = get_internals().registered_instances.equal_range(src);
+        for (auto it_i = it_instances.first; it_i != it_instances.second; ++it_i) {
+            for (auto instance_type : detail::all_type_info(Py_TYPE(it_i->second))) {
+                if (instance_type && same_type(*instance_type->cpptype, *tinfo->cpptype))
+                    return handle((PyObject *) it_i->second).inc_ref();
+            }
+        }
+
+        auto inst = reinterpret_steal<object>(make_new_instance(tinfo->type));
+        auto wrapper = reinterpret_cast<instance *>(inst.ptr());
+        wrapper->owned = false;
+        void *&valueptr = values_and_holders(wrapper).begin()->value_ptr();
+
+        switch (policy) {
+            case return_value_policy::automatic:
+            case return_value_policy::take_ownership:
+                valueptr = src;
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::automatic_reference:
+            case return_value_policy::reference:
+                valueptr = src;
+                wrapper->owned = false;
+                break;
+
+            case return_value_policy::copy:
+                if (copy_constructor)
+                    valueptr = copy_constructor(src);
+                else
+                    throw cast_error("return_value_policy = copy, but the "
+                                     "object is non-copyable!");
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::move:
+                if (move_constructor)
+                    valueptr = move_constructor(src);
+                else if (copy_constructor)
+                    valueptr = copy_constructor(src);
+                else
+                    throw cast_error("return_value_policy = move, but the "
+                                     "object is neither movable nor copyable!");
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::reference_internal:
+                valueptr = src;
+                wrapper->owned = false;
+                keep_alive_impl(inst, parent);
+                break;
+
+            default:
+                throw cast_error("unhandled return_value_policy: should not happen!");
+        }
+
+        tinfo->init_instance(wrapper, existing_holder);
+
+        return inst.release();
+    }
+
+    // Base methods for generic caster; there are overridden in copyable_holder_caster
+    void load_value(value_and_holder &&v_h) {
+        auto *&vptr = v_h.value_ptr();
+        // Lazy allocation for unallocated values:
+        if (vptr == nullptr) {
+            auto *type = v_h.type ? v_h.type : typeinfo;
+            vptr = type->operator_new(type->type_size);
+        }
+        value = vptr;
+    }
+    bool try_implicit_casts(handle src, bool convert) {
+        for (auto &cast : typeinfo->implicit_casts) {
+            type_caster_generic sub_caster(*cast.first);
+            if (sub_caster.load(src, convert)) {
+                value = cast.second(sub_caster.value);
+                return true;
+            }
+        }
+        return false;
+    }
+    bool try_direct_conversions(handle src) {
+        for (auto &converter : *typeinfo->direct_conversions) {
+            if (converter(src.ptr(), value))
+                return true;
+        }
+        return false;
+    }
+    void check_holder_compat() {}
+
+    PYBIND11_NOINLINE static void *local_load(PyObject *src, const type_info *ti) {
+        auto caster = type_caster_generic(ti);
+        if (caster.load(src, false))
+            return caster.value;
+        return nullptr;
+    }
+
+    /// Try to load with foreign typeinfo, if available. Used when there is no
+    /// native typeinfo, or when the native one wasn't able to produce a value.
+    PYBIND11_NOINLINE bool try_load_foreign_module_local(handle src) {
+        constexpr auto *local_key = PYBIND11_MODULE_LOCAL_ID;
+        const auto pytype = src.get_type();
+        if (!hasattr(pytype, local_key))
+            return false;
+
+        type_info *foreign_typeinfo = reinterpret_borrow<capsule>(getattr(pytype, local_key));
+        // Only consider this foreign loader if actually foreign and is a loader of the correct cpp type
+        if (foreign_typeinfo->module_local_load == &local_load
+            || (cpptype && !same_type(*cpptype, *foreign_typeinfo->cpptype)))
+            return false;
+
+        if (auto result = foreign_typeinfo->module_local_load(src.ptr(), foreign_typeinfo)) {
+            value = result;
+            return true;
+        }
+        return false;
+    }
+
+    // Implementation of `load`; this takes the type of `this` so that it can dispatch the relevant
+    // bits of code between here and copyable_holder_caster where the two classes need different
+    // logic (without having to resort to virtual inheritance).
+    template <typename ThisT>
+    PYBIND11_NOINLINE bool load_impl(handle src, bool convert) {
+        if (!src) return false;
+        if (!typeinfo) return try_load_foreign_module_local(src);
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) return false;
+            value = nullptr;
+            return true;
+        }
+
+        auto &this_ = static_cast<ThisT &>(*this);
+        this_.check_holder_compat();
+
+        PyTypeObject *srctype = Py_TYPE(src.ptr());
+
+        // Case 1: If src is an exact type match for the target type then we can reinterpret_cast
+        // the instance's value pointer to the target type:
+        if (srctype == typeinfo->type) {
+            this_.load_value(reinterpret_cast<instance *>(src.ptr())->get_value_and_holder());
+            return true;
+        }
+        // Case 2: We have a derived class
+        else if (PyType_IsSubtype(srctype, typeinfo->type)) {
+            auto &bases = all_type_info(srctype);
+            bool no_cpp_mi = typeinfo->simple_type;
+
+            // Case 2a: the python type is a Python-inherited derived class that inherits from just
+            // one simple (no MI) pybind11 class, or is an exact match, so the C++ instance is of
+            // the right type and we can use reinterpret_cast.
+            // (This is essentially the same as case 2b, but because not using multiple inheritance
+            // is extremely common, we handle it specially to avoid the loop iterator and type
+            // pointer lookup overhead)
+            if (bases.size() == 1 && (no_cpp_mi || bases.front()->type == typeinfo->type)) {
+                this_.load_value(reinterpret_cast<instance *>(src.ptr())->get_value_and_holder());
+                return true;
+            }
+            // Case 2b: the python type inherits from multiple C++ bases.  Check the bases to see if
+            // we can find an exact match (or, for a simple C++ type, an inherited match); if so, we
+            // can safely reinterpret_cast to the relevant pointer.
+            else if (bases.size() > 1) {
+                for (auto base : bases) {
+                    if (no_cpp_mi ? PyType_IsSubtype(base->type, typeinfo->type) : base->type == typeinfo->type) {
+                        this_.load_value(reinterpret_cast<instance *>(src.ptr())->get_value_and_holder(base));
+                        return true;
+                    }
+                }
+            }
+
+            // Case 2c: C++ multiple inheritance is involved and we couldn't find an exact type match
+            // in the registered bases, above, so try implicit casting (needed for proper C++ casting
+            // when MI is involved).
+            if (this_.try_implicit_casts(src, convert))
+                return true;
+        }
+
+        // Perform an implicit conversion
+        if (convert) {
+            for (auto &converter : typeinfo->implicit_conversions) {
+                auto temp = reinterpret_steal<object>(converter(src.ptr(), typeinfo->type));
+                if (load_impl<ThisT>(temp, false)) {
+                    loader_life_support::add_patient(temp);
+                    return true;
+                }
+            }
+            if (this_.try_direct_conversions(src))
+                return true;
+        }
+
+        // Failed to match local typeinfo. Try again with global.
+        if (typeinfo->module_local) {
+            if (auto gtype = get_global_type_info(*typeinfo->cpptype)) {
+                typeinfo = gtype;
+                return load(src, false);
+            }
+        }
+
+        // Global typeinfo has precedence over foreign module_local
+        return try_load_foreign_module_local(src);
+    }
+
+
+    // Called to do type lookup and wrap the pointer and type in a pair when a dynamic_cast
+    // isn't needed or can't be used.  If the type is unknown, sets the error and returns a pair
+    // with .second = nullptr.  (p.first = nullptr is not an error: it becomes None).
+    PYBIND11_NOINLINE static std::pair<const void *, const type_info *> src_and_type(
+            const void *src, const std::type_info &cast_type, const std::type_info *rtti_type = nullptr) {
+        if (auto *tpi = get_type_info(cast_type))
+            return {src, const_cast<const type_info *>(tpi)};
+
+        // Not found, set error:
+        std::string tname = rtti_type ? rtti_type->name() : cast_type.name();
+        detail::clean_type_id(tname);
+        std::string msg = "Unregistered type : " + tname;
+        PyErr_SetString(PyExc_TypeError, msg.c_str());
+        return {nullptr, nullptr};
+    }
+
+    const type_info *typeinfo = nullptr;
+    const std::type_info *cpptype = nullptr;
+    void *value = nullptr;
+};
+
+/**
+ * Determine suitable casting operator for pointer-or-lvalue-casting type casters.  The type caster
+ * needs to provide `operator T*()` and `operator T&()` operators.
+ *
+ * If the type supports moving the value away via an `operator T&&() &&` method, it should use
+ * `movable_cast_op_type` instead.
+ */
+template <typename T>
+using cast_op_type =
+    conditional_t<std::is_pointer<remove_reference_t<T>>::value,
+        typename std::add_pointer<intrinsic_t<T>>::type,
+        typename std::add_lvalue_reference<intrinsic_t<T>>::type>;
+
+/**
+ * Determine suitable casting operator for a type caster with a movable value.  Such a type caster
+ * needs to provide `operator T*()`, `operator T&()`, and `operator T&&() &&`.  The latter will be
+ * called in appropriate contexts where the value can be moved rather than copied.
+ *
+ * These operator are automatically provided when using the PYBIND11_TYPE_CASTER macro.
+ */
+template <typename T>
+using movable_cast_op_type =
+    conditional_t<std::is_pointer<typename std::remove_reference<T>::type>::value,
+        typename std::add_pointer<intrinsic_t<T>>::type,
+    conditional_t<std::is_rvalue_reference<T>::value,
+        typename std::add_rvalue_reference<intrinsic_t<T>>::type,
+        typename std::add_lvalue_reference<intrinsic_t<T>>::type>>;
+
+// std::is_copy_constructible isn't quite enough: it lets std::vector<T> (and similar) through when
+// T is non-copyable, but code containing such a copy constructor fails to actually compile.
+template <typename T, typename SFINAE = void> struct is_copy_constructible : std::is_copy_constructible<T> {};
+
+// Specialization for types that appear to be copy constructible but also look like stl containers
+// (we specifically check for: has `value_type` and `reference` with `reference = value_type&`): if
+// so, copy constructability depends on whether the value_type is copy constructible.
+template <typename Container> struct is_copy_constructible<Container, enable_if_t<all_of<
+        std::is_copy_constructible<Container>,
+        std::is_same<typename Container::value_type &, typename Container::reference>
+    >::value>> : is_copy_constructible<typename Container::value_type> {};
+
+#if !defined(PYBIND11_CPP17)
+// Likewise for std::pair before C++17 (which mandates that the copy constructor not exist when the
+// two types aren't themselves copy constructible).
+template <typename T1, typename T2> struct is_copy_constructible<std::pair<T1, T2>>
+    : all_of<is_copy_constructible<T1>, is_copy_constructible<T2>> {};
+#endif
+
+/// Generic type caster for objects stored on the heap
+template <typename type> class type_caster_base : public type_caster_generic {
+    using itype = intrinsic_t<type>;
+public:
+    static PYBIND11_DESCR name() { return type_descr(_<type>()); }
+
+    type_caster_base() : type_caster_base(typeid(type)) { }
+    explicit type_caster_base(const std::type_info &info) : type_caster_generic(info) { }
+
+    static handle cast(const itype &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic || policy == return_value_policy::automatic_reference)
+            policy = return_value_policy::copy;
+        return cast(&src, policy, parent);
+    }
+
+    static handle cast(itype &&src, return_value_policy, handle parent) {
+        return cast(&src, return_value_policy::move, parent);
+    }
+
+    // Returns a (pointer, type_info) pair taking care of necessary RTTI type lookup for a
+    // polymorphic type.  If the instance isn't derived, returns the non-RTTI base version.
+    template <typename T = itype, enable_if_t<std::is_polymorphic<T>::value, int> = 0>
+    static std::pair<const void *, const type_info *> src_and_type(const itype *src) {
+        const void *vsrc = src;
+        auto &cast_type = typeid(itype);
+        const std::type_info *instance_type = nullptr;
+        if (vsrc) {
+            instance_type = &typeid(*src);
+            if (!same_type(cast_type, *instance_type)) {
+                // This is a base pointer to a derived type; if it is a pybind11-registered type, we
+                // can get the correct derived pointer (which may be != base pointer) by a
+                // dynamic_cast to most derived type:
+                if (auto *tpi = get_type_info(*instance_type))
+                    return {dynamic_cast<const void *>(src), const_cast<const type_info *>(tpi)};
+            }
+        }
+        // Otherwise we have either a nullptr, an `itype` pointer, or an unknown derived pointer, so
+        // don't do a cast
+        return type_caster_generic::src_and_type(vsrc, cast_type, instance_type);
+    }
+
+    // Non-polymorphic type, so no dynamic casting; just call the generic version directly
+    template <typename T = itype, enable_if_t<!std::is_polymorphic<T>::value, int> = 0>
+    static std::pair<const void *, const type_info *> src_and_type(const itype *src) {
+        return type_caster_generic::src_and_type(src, typeid(itype));
+    }
+
+    static handle cast(const itype *src, return_value_policy policy, handle parent) {
+        auto st = src_and_type(src);
+        return type_caster_generic::cast(
+            st.first, policy, parent, st.second,
+            make_copy_constructor(src), make_move_constructor(src));
+    }
+
+    static handle cast_holder(const itype *src, const void *holder) {
+        auto st = src_and_type(src);
+        return type_caster_generic::cast(
+            st.first, return_value_policy::take_ownership, {}, st.second,
+            nullptr, nullptr, holder);
+    }
+
+    template <typename T> using cast_op_type = cast_op_type<T>;
+
+    operator itype*() { return (type *) value; }
+    operator itype&() { if (!value) throw reference_cast_error(); return *((itype *) value); }
+
+protected:
+    using Constructor = void *(*)(const void *);
+
+    /* Only enabled when the types are {copy,move}-constructible *and* when the type
+       does not have a private operator new implementation. */
+    template <typename T, typename = enable_if_t<is_copy_constructible<T>::value>>
+    static auto make_copy_constructor(const T *x) -> decltype(new T(*x), Constructor{}) {
+        return [](const void *arg) -> void * {
+            return new T(*reinterpret_cast<const T *>(arg));
+        };
+    }
+
+    template <typename T, typename = enable_if_t<std::is_move_constructible<T>::value>>
+    static auto make_move_constructor(const T *x) -> decltype(new T(std::move(*const_cast<T *>(x))), Constructor{}) {
+        return [](const void *arg) -> void * {
+            return new T(std::move(*const_cast<T *>(reinterpret_cast<const T *>(arg))));
+        };
+    }
+
+    static Constructor make_copy_constructor(...) { return nullptr; }
+    static Constructor make_move_constructor(...) { return nullptr; }
+};
+
+template <typename type, typename SFINAE = void> class type_caster : public type_caster_base<type> { };
+template <typename type> using make_caster = type_caster<intrinsic_t<type>>;
+
+// Shortcut for calling a caster's `cast_op_type` cast operator for casting a type_caster to a T
+template <typename T> typename make_caster<T>::template cast_op_type<T> cast_op(make_caster<T> &caster) {
+    return caster.operator typename make_caster<T>::template cast_op_type<T>();
+}
+template <typename T> typename make_caster<T>::template cast_op_type<typename std::add_rvalue_reference<T>::type>
+cast_op(make_caster<T> &&caster) {
+    return std::move(caster).operator
+        typename make_caster<T>::template cast_op_type<typename std::add_rvalue_reference<T>::type>();
+}
+
+template <typename type> class type_caster<std::reference_wrapper<type>> {
+private:
+    using caster_t = make_caster<type>;
+    caster_t subcaster;
+    using subcaster_cast_op_type = typename caster_t::template cast_op_type<type>;
+    static_assert(std::is_same<typename std::remove_const<type>::type &, subcaster_cast_op_type>::value,
+            "std::reference_wrapper<T> caster requires T to have a caster with an `T &` operator");
+public:
+    bool load(handle src, bool convert) { return subcaster.load(src, convert); }
+    static PYBIND11_DESCR name() { return caster_t::name(); }
+    static handle cast(const std::reference_wrapper<type> &src, return_value_policy policy, handle parent) {
+        // It is definitely wrong to take ownership of this pointer, so mask that rvp
+        if (policy == return_value_policy::take_ownership || policy == return_value_policy::automatic)
+            policy = return_value_policy::automatic_reference;
+        return caster_t::cast(&src.get(), policy, parent);
+    }
+    template <typename T> using cast_op_type = std::reference_wrapper<type>;
+    operator std::reference_wrapper<type>() { return subcaster.operator subcaster_cast_op_type&(); }
+};
+
+#define PYBIND11_TYPE_CASTER(type, py_name) \
+    protected: \
+        type value; \
+    public: \
+        static PYBIND11_DESCR name() { return type_descr(py_name); } \
+        template <typename T_, enable_if_t<std::is_same<type, remove_cv_t<T_>>::value, int> = 0> \
+        static handle cast(T_ *src, return_value_policy policy, handle parent) { \
+            if (!src) return none().release(); \
+            if (policy == return_value_policy::take_ownership) { \
+                auto h = cast(std::move(*src), policy, parent); delete src; return h; \
+            } else { \
+                return cast(*src, policy, parent); \
+            } \
+        } \
+        operator type*() { return &value; } \
+        operator type&() { return value; } \
+        operator type&&() && { return std::move(value); } \
+        template <typename T_> using cast_op_type = pybind11::detail::movable_cast_op_type<T_>
+
+
+template <typename CharT> using is_std_char_type = any_of<
+    std::is_same<CharT, char>, /* std::string */
+    std::is_same<CharT, char16_t>, /* std::u16string */
+    std::is_same<CharT, char32_t>, /* std::u32string */
+    std::is_same<CharT, wchar_t> /* std::wstring */
+>;
+
+template <typename T>
+struct type_caster<T, enable_if_t<std::is_arithmetic<T>::value && !is_std_char_type<T>::value>> {
+    using _py_type_0 = conditional_t<sizeof(T) <= sizeof(long), long, long long>;
+    using _py_type_1 = conditional_t<std::is_signed<T>::value, _py_type_0, typename std::make_unsigned<_py_type_0>::type>;
+    using py_type = conditional_t<std::is_floating_point<T>::value, double, _py_type_1>;
+public:
+
+    bool load(handle src, bool convert) {
+        py_type py_value;
+
+        if (!src)
+            return false;
+
+        if (std::is_floating_point<T>::value) {
+            if (convert || PyFloat_Check(src.ptr()))
+                py_value = (py_type) PyFloat_AsDouble(src.ptr());
+            else
+                return false;
+        } else if (PyFloat_Check(src.ptr())) {
+            return false;
+        } else if (std::is_unsigned<py_type>::value) {
+            py_value = as_unsigned<py_type>(src.ptr());
+        } else { // signed integer:
+            py_value = sizeof(T) <= sizeof(long)
+                ? (py_type) PyLong_AsLong(src.ptr())
+                : (py_type) PYBIND11_LONG_AS_LONGLONG(src.ptr());
+        }
+
+        bool py_err = py_value == (py_type) -1 && PyErr_Occurred();
+        if (py_err || (std::is_integral<T>::value && sizeof(py_type) != sizeof(T) &&
+                       (py_value < (py_type) std::numeric_limits<T>::min() ||
+                        py_value > (py_type) std::numeric_limits<T>::max()))) {
+            bool type_error = py_err && PyErr_ExceptionMatches(
+#if PY_VERSION_HEX < 0x03000000 && !defined(PYPY_VERSION)
+                PyExc_SystemError
+#else
+                PyExc_TypeError
+#endif
+            );
+            PyErr_Clear();
+            if (type_error && convert && PyNumber_Check(src.ptr())) {
+                auto tmp = reinterpret_steal<object>(std::is_floating_point<T>::value
+                                                     ? PyNumber_Float(src.ptr())
+                                                     : PyNumber_Long(src.ptr()));
+                PyErr_Clear();
+                return load(tmp, false);
+            }
+            return false;
+        }
+
+        value = (T) py_value;
+        return true;
+    }
+
+    static handle cast(T src, return_value_policy /* policy */, handle /* parent */) {
+        if (std::is_floating_point<T>::value) {
+            return PyFloat_FromDouble((double) src);
+        } else if (sizeof(T) <= sizeof(long)) {
+            if (std::is_signed<T>::value)
+                return PyLong_FromLong((long) src);
+            else
+                return PyLong_FromUnsignedLong((unsigned long) src);
+        } else {
+            if (std::is_signed<T>::value)
+                return PyLong_FromLongLong((long long) src);
+            else
+                return PyLong_FromUnsignedLongLong((unsigned long long) src);
+        }
+    }
+
+    PYBIND11_TYPE_CASTER(T, _<std::is_integral<T>::value>("int", "float"));
+};
+
+template<typename T> struct void_caster {
+public:
+    bool load(handle src, bool) {
+        if (src && src.is_none())
+            return true;
+        return false;
+    }
+    static handle cast(T, return_value_policy /* policy */, handle /* parent */) {
+        return none().inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(T, _("None"));
+};
+
+template <> class type_caster<void_type> : public void_caster<void_type> {};
+
+template <> class type_caster<void> : public type_caster<void_type> {
+public:
+    using type_caster<void_type>::cast;
+
+    bool load(handle h, bool) {
+        if (!h) {
+            return false;
+        } else if (h.is_none()) {
+            value = nullptr;
+            return true;
+        }
+
+        /* Check if this is a capsule */
+        if (isinstance<capsule>(h)) {
+            value = reinterpret_borrow<capsule>(h);
+            return true;
+        }
+
+        /* Check if this is a C++ type */
+        auto &bases = all_type_info((PyTypeObject *) h.get_type().ptr());
+        if (bases.size() == 1) { // Only allowing loading from a single-value type
+            value = values_and_holders(reinterpret_cast<instance *>(h.ptr())).begin()->value_ptr();
+            return true;
+        }
+
+        /* Fail */
+        return false;
+    }
+
+    static handle cast(const void *ptr, return_value_policy /* policy */, handle /* parent */) {
+        if (ptr)
+            return capsule(ptr).release();
+        else
+            return none().inc_ref();
+    }
+
+    template <typename T> using cast_op_type = void*&;
+    operator void *&() { return value; }
+    static PYBIND11_DESCR name() { return type_descr(_("capsule")); }
+private:
+    void *value = nullptr;
+};
+
+template <> class type_caster<std::nullptr_t> : public void_caster<std::nullptr_t> { };
+
+template <> class type_caster<bool> {
+public:
+    bool load(handle src, bool convert) {
+        if (!src) return false;
+        else if (src.ptr() == Py_True) { value = true; return true; }
+        else if (src.ptr() == Py_False) { value = false; return true; }
+        else if (convert || !strcmp("numpy.bool_", Py_TYPE(src.ptr())->tp_name)) {
+            // (allow non-implicit conversion for numpy booleans)
+
+            Py_ssize_t res = -1;
+            if (src.is_none()) {
+                res = 0;  // None is implicitly converted to False
+            }
+            #if defined(PYPY_VERSION)
+            // On PyPy, check that "__bool__" (or "__nonzero__" on Python 2.7) attr exists
+            else if (hasattr(src, PYBIND11_BOOL_ATTR)) {
+                res = PyObject_IsTrue(src.ptr());
+            }
+            #else
+            // Alternate approach for CPython: this does the same as the above, but optimized
+            // using the CPython API so as to avoid an unneeded attribute lookup.
+            else if (auto tp_as_number = src.ptr()->ob_type->tp_as_number) {
+                if (PYBIND11_NB_BOOL(tp_as_number)) {
+                    res = (*PYBIND11_NB_BOOL(tp_as_number))(src.ptr());
+                }
+            }
+            #endif
+            if (res == 0 || res == 1) {
+                value = (bool) res;
+                return true;
+            }
+        }
+        return false;
+    }
+    static handle cast(bool src, return_value_policy /* policy */, handle /* parent */) {
+        return handle(src ? Py_True : Py_False).inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(bool, _("bool"));
+};
+
+// Helper class for UTF-{8,16,32} C++ stl strings:
+template <typename StringType, bool IsView = false> struct string_caster {
+    using CharT = typename StringType::value_type;
+
+    // Simplify life by being able to assume standard char sizes (the standard only guarantees
+    // minimums, but Python requires exact sizes)
+    static_assert(!std::is_same<CharT, char>::value || sizeof(CharT) == 1, "Unsupported char size != 1");
+    static_assert(!std::is_same<CharT, char16_t>::value || sizeof(CharT) == 2, "Unsupported char16_t size != 2");
+    static_assert(!std::is_same<CharT, char32_t>::value || sizeof(CharT) == 4, "Unsupported char32_t size != 4");
+    // wchar_t can be either 16 bits (Windows) or 32 (everywhere else)
+    static_assert(!std::is_same<CharT, wchar_t>::value || sizeof(CharT) == 2 || sizeof(CharT) == 4,
+            "Unsupported wchar_t size != 2/4");
+    static constexpr size_t UTF_N = 8 * sizeof(CharT);
+
+    bool load(handle src, bool) {
+#if PY_MAJOR_VERSION < 3
+        object temp;
+#endif
+        handle load_src = src;
+        if (!src) {
+            return false;
+        } else if (!PyUnicode_Check(load_src.ptr())) {
+#if PY_MAJOR_VERSION >= 3
+            return load_bytes(load_src);
+#else
+            if (sizeof(CharT) == 1) {
+                return load_bytes(load_src);
+            }
+
+            // The below is a guaranteed failure in Python 3 when PyUnicode_Check returns false
+            if (!PYBIND11_BYTES_CHECK(load_src.ptr()))
+                return false;
+
+            temp = reinterpret_steal<object>(PyUnicode_FromObject(load_src.ptr()));
+            if (!temp) { PyErr_Clear(); return false; }
+            load_src = temp;
+#endif
+        }
+
+        object utfNbytes = reinterpret_steal<object>(PyUnicode_AsEncodedString(
+            load_src.ptr(), UTF_N == 8 ? "utf-8" : UTF_N == 16 ? "utf-16" : "utf-32", nullptr));
+        if (!utfNbytes) { PyErr_Clear(); return false; }
+
+        const CharT *buffer = reinterpret_cast<const CharT *>(PYBIND11_BYTES_AS_STRING(utfNbytes.ptr()));
+        size_t length = (size_t) PYBIND11_BYTES_SIZE(utfNbytes.ptr()) / sizeof(CharT);
+        if (UTF_N > 8) { buffer++; length--; } // Skip BOM for UTF-16/32
+        value = StringType(buffer, length);
+
+        // If we're loading a string_view we need to keep the encoded Python object alive:
+        if (IsView)
+            loader_life_support::add_patient(utfNbytes);
+
+        return true;
+    }
+
+    static handle cast(const StringType &src, return_value_policy /* policy */, handle /* parent */) {
+        const char *buffer = reinterpret_cast<const char *>(src.data());
+        ssize_t nbytes = ssize_t(src.size() * sizeof(CharT));
+        handle s = decode_utfN(buffer, nbytes);
+        if (!s) throw error_already_set();
+        return s;
+    }
+
+    PYBIND11_TYPE_CASTER(StringType, _(PYBIND11_STRING_NAME));
+
+private:
+    static handle decode_utfN(const char *buffer, ssize_t nbytes) {
+#if !defined(PYPY_VERSION)
+        return
+            UTF_N == 8  ? PyUnicode_DecodeUTF8(buffer, nbytes, nullptr) :
+            UTF_N == 16 ? PyUnicode_DecodeUTF16(buffer, nbytes, nullptr, nullptr) :
+                          PyUnicode_DecodeUTF32(buffer, nbytes, nullptr, nullptr);
+#else
+        // PyPy seems to have multiple problems related to PyUnicode_UTF*: the UTF8 version
+        // sometimes segfaults for unknown reasons, while the UTF16 and 32 versions require a
+        // non-const char * arguments, which is also a nuissance, so bypass the whole thing by just
+        // passing the encoding as a string value, which works properly:
+        return PyUnicode_Decode(buffer, nbytes, UTF_N == 8 ? "utf-8" : UTF_N == 16 ? "utf-16" : "utf-32", nullptr);
+#endif
+    }
+
+    // When loading into a std::string or char*, accept a bytes object as-is (i.e.
+    // without any encoding/decoding attempt).  For other C++ char sizes this is a no-op.
+    // which supports loading a unicode from a str, doesn't take this path.
+    template <typename C = CharT>
+    bool load_bytes(enable_if_t<sizeof(C) == 1, handle> src) {
+        if (PYBIND11_BYTES_CHECK(src.ptr())) {
+            // We were passed a Python 3 raw bytes; accept it into a std::string or char*
+            // without any encoding attempt.
+            const char *bytes = PYBIND11_BYTES_AS_STRING(src.ptr());
+            if (bytes) {
+                value = StringType(bytes, (size_t) PYBIND11_BYTES_SIZE(src.ptr()));
+                return true;
+            }
+        }
+
+        return false;
+    }
+
+    template <typename C = CharT>
+    bool load_bytes(enable_if_t<sizeof(C) != 1, handle>) { return false; }
+};
+
+template <typename CharT, class Traits, class Allocator>
+struct type_caster<std::basic_string<CharT, Traits, Allocator>, enable_if_t<is_std_char_type<CharT>::value>>
+    : string_caster<std::basic_string<CharT, Traits, Allocator>> {};
+
+#ifdef PYBIND11_HAS_STRING_VIEW
+template <typename CharT, class Traits>
+struct type_caster<std::basic_string_view<CharT, Traits>, enable_if_t<is_std_char_type<CharT>::value>>
+    : string_caster<std::basic_string_view<CharT, Traits>, true> {};
+#endif
+
+// Type caster for C-style strings.  We basically use a std::string type caster, but also add the
+// ability to use None as a nullptr char* (which the string caster doesn't allow).
+template <typename CharT> struct type_caster<CharT, enable_if_t<is_std_char_type<CharT>::value>> {
+    using StringType = std::basic_string<CharT>;
+    using StringCaster = type_caster<StringType>;
+    StringCaster str_caster;
+    bool none = false;
+    CharT one_char = 0;
+public:
+    bool load(handle src, bool convert) {
+        if (!src) return false;
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) return false;
+            none = true;
+            return true;
+        }
+        return str_caster.load(src, convert);
+    }
+
+    static handle cast(const CharT *src, return_value_policy policy, handle parent) {
+        if (src == nullptr) return pybind11::none().inc_ref();
+        return StringCaster::cast(StringType(src), policy, parent);
+    }
+
+    static handle cast(CharT src, return_value_policy policy, handle parent) {
+        if (std::is_same<char, CharT>::value) {
+            handle s = PyUnicode_DecodeLatin1((const char *) &src, 1, nullptr);
+            if (!s) throw error_already_set();
+            return s;
+        }
+        return StringCaster::cast(StringType(1, src), policy, parent);
+    }
+
+    operator CharT*() { return none ? nullptr : const_cast<CharT *>(static_cast<StringType &>(str_caster).c_str()); }
+    operator CharT&() {
+        if (none)
+            throw value_error("Cannot convert None to a character");
+
+        auto &value = static_cast<StringType &>(str_caster);
+        size_t str_len = value.size();
+        if (str_len == 0)
+            throw value_error("Cannot convert empty string to a character");
+
+        // If we're in UTF-8 mode, we have two possible failures: one for a unicode character that
+        // is too high, and one for multiple unicode characters (caught later), so we need to figure
+        // out how long the first encoded character is in bytes to distinguish between these two
+        // errors.  We also allow want to allow unicode characters U+0080 through U+00FF, as those
+        // can fit into a single char value.
+        if (StringCaster::UTF_N == 8 && str_len > 1 && str_len <= 4) {
+            unsigned char v0 = static_cast<unsigned char>(value[0]);
+            size_t char0_bytes = !(v0 & 0x80) ? 1 : // low bits only: 0-127
+                (v0 & 0xE0) == 0xC0 ? 2 : // 0b110xxxxx - start of 2-byte sequence
+                (v0 & 0xF0) == 0xE0 ? 3 : // 0b1110xxxx - start of 3-byte sequence
+                4; // 0b11110xxx - start of 4-byte sequence
+
+            if (char0_bytes == str_len) {
+                // If we have a 128-255 value, we can decode it into a single char:
+                if (char0_bytes == 2 && (v0 & 0xFC) == 0xC0) { // 0x110000xx 0x10xxxxxx
+                    one_char = static_cast<CharT>(((v0 & 3) << 6) + (static_cast<unsigned char>(value[1]) & 0x3F));
+                    return one_char;
+                }
+                // Otherwise we have a single character, but it's > U+00FF
+                throw value_error("Character code point not in range(0x100)");
+            }
+        }
+
+        // UTF-16 is much easier: we can only have a surrogate pair for values above U+FFFF, thus a
+        // surrogate pair with total length 2 instantly indicates a range error (but not a "your
+        // string was too long" error).
+        else if (StringCaster::UTF_N == 16 && str_len == 2) {
+            one_char = static_cast<CharT>(value[0]);
+            if (one_char >= 0xD800 && one_char < 0xE000)
+                throw value_error("Character code point not in range(0x10000)");
+        }
+
+        if (str_len != 1)
+            throw value_error("Expected a character, but multi-character string found");
+
+        one_char = value[0];
+        return one_char;
+    }
+
+    static PYBIND11_DESCR name() { return type_descr(_(PYBIND11_STRING_NAME)); }
+    template <typename _T> using cast_op_type = pybind11::detail::cast_op_type<_T>;
+};
+
+// Base implementation for std::tuple and std::pair
+template <template<typename...> class Tuple, typename... Ts> class tuple_caster {
+    using type = Tuple<Ts...>;
+    static constexpr auto size = sizeof...(Ts);
+    using indices = make_index_sequence<size>;
+public:
+
+    bool load(handle src, bool convert) {
+        if (!isinstance<sequence>(src))
+            return false;
+        const auto seq = reinterpret_borrow<sequence>(src);
+        if (seq.size() != size)
+            return false;
+        return load_impl(seq, convert, indices{});
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        return cast_impl(std::forward<T>(src), policy, parent, indices{});
+    }
+
+    static PYBIND11_DESCR name() {
+        return type_descr(_("Tuple[") + detail::concat(make_caster<Ts>::name()...) + _("]"));
+    }
+
+    template <typename T> using cast_op_type = type;
+
+    operator type() & { return implicit_cast(indices{}); }
+    operator type() && { return std::move(*this).implicit_cast(indices{}); }
+
+protected:
+    template <size_t... Is>
+    type implicit_cast(index_sequence<Is...>) & { return type(cast_op<Ts>(std::get<Is>(subcasters))...); }
+    template <size_t... Is>
+    type implicit_cast(index_sequence<Is...>) && { return type(cast_op<Ts>(std::move(std::get<Is>(subcasters)))...); }
+
+    static constexpr bool load_impl(const sequence &, bool, index_sequence<>) { return true; }
+
+    template <size_t... Is>
+    bool load_impl(const sequence &seq, bool convert, index_sequence<Is...>) {
+        for (bool r : {std::get<Is>(subcasters).load(seq[Is], convert)...})
+            if (!r)
+                return false;
+        return true;
+    }
+
+    /* Implementation: Convert a C++ tuple into a Python tuple */
+    template <typename T, size_t... Is>
+    static handle cast_impl(T &&src, return_value_policy policy, handle parent, index_sequence<Is...>) {
+        std::array<object, size> entries{{
+            reinterpret_steal<object>(make_caster<Ts>::cast(std::get<Is>(std::forward<T>(src)), policy, parent))...
+        }};
+        for (const auto &entry: entries)
+            if (!entry)
+                return handle();
+        tuple result(size);
+        int counter = 0;
+        for (auto & entry: entries)
+            PyTuple_SET_ITEM(result.ptr(), counter++, entry.release().ptr());
+        return result.release();
+    }
+
+    Tuple<make_caster<Ts>...> subcasters;
+};
+
+template <typename T1, typename T2> class type_caster<std::pair<T1, T2>>
+    : public tuple_caster<std::pair, T1, T2> {};
+
+template <typename... Ts> class type_caster<std::tuple<Ts...>>
+    : public tuple_caster<std::tuple, Ts...> {};
+
+/// Helper class which abstracts away certain actions. Users can provide specializations for
+/// custom holders, but it's only necessary if the type has a non-standard interface.
+template <typename T>
+struct holder_helper {
+    static auto get(const T &p) -> decltype(p.get()) { return p.get(); }
+};
+
+/// Type caster for holder types like std::shared_ptr, etc.
+template <typename type, typename holder_type>
+struct copyable_holder_caster : public type_caster_base<type> {
+public:
+    using base = type_caster_base<type>;
+    static_assert(std::is_base_of<base, type_caster<type>>::value,
+            "Holder classes are only supported for custom types");
+    using base::base;
+    using base::cast;
+    using base::typeinfo;
+    using base::value;
+
+    bool load(handle src, bool convert) {
+        return base::template load_impl<copyable_holder_caster<type, holder_type>>(src, convert);
+    }
+
+    explicit operator type*() { return this->value; }
+    explicit operator type&() { return *(this->value); }
+    explicit operator holder_type*() { return &holder; }
+
+    // Workaround for Intel compiler bug
+    // see pybind11 issue 94
+    #if defined(__ICC) || defined(__INTEL_COMPILER)
+    operator holder_type&() { return holder; }
+    #else
+    explicit operator holder_type&() { return holder; }
+    #endif
+
+    static handle cast(const holder_type &src, return_value_policy, handle) {
+        const auto *ptr = holder_helper<holder_type>::get(src);
+        return type_caster_base<type>::cast_holder(ptr, &src);
+    }
+
+protected:
+    friend class type_caster_generic;
+    void check_holder_compat() {
+        if (typeinfo->default_holder)
+            throw cast_error("Unable to load a custom holder type from a default-holder instance");
+    }
+
+    bool load_value(value_and_holder &&v_h) {
+        if (v_h.holder_constructed()) {
+            value = v_h.value_ptr();
+            holder = v_h.template holder<holder_type>();
+            return true;
+        } else {
+            throw cast_error("Unable to cast from non-held to held instance (T& to Holder<T>) "
+#if defined(NDEBUG)
+                             "(compile in debug mode for type information)");
+#else
+                             "of type '" + type_id<holder_type>() + "''");
+#endif
+        }
+    }
+
+    template <typename T = holder_type, detail::enable_if_t<!std::is_constructible<T, const T &, type*>::value, int> = 0>
+    bool try_implicit_casts(handle, bool) { return false; }
+
+    template <typename T = holder_type, detail::enable_if_t<std::is_constructible<T, const T &, type*>::value, int> = 0>
+    bool try_implicit_casts(handle src, bool convert) {
+        for (auto &cast : typeinfo->implicit_casts) {
+            copyable_holder_caster sub_caster(*cast.first);
+            if (sub_caster.load(src, convert)) {
+                value = cast.second(sub_caster.value);
+                holder = holder_type(sub_caster.holder, (type *) value);
+                return true;
+            }
+        }
+        return false;
+    }
+
+    static bool try_direct_conversions(handle) { return false; }
+
+
+    holder_type holder;
+};
+
+/// Specialize for the common std::shared_ptr, so users don't need to
+template <typename T>
+class type_caster<std::shared_ptr<T>> : public copyable_holder_caster<T, std::shared_ptr<T>> { };
+
+template <typename type, typename holder_type>
+struct move_only_holder_caster {
+    static_assert(std::is_base_of<type_caster_base<type>, type_caster<type>>::value,
+            "Holder classes are only supported for custom types");
+
+    static handle cast(holder_type &&src, return_value_policy, handle) {
+        auto *ptr = holder_helper<holder_type>::get(src);
+        return type_caster_base<type>::cast_holder(ptr, &src);
+    }
+    static PYBIND11_DESCR name() { return type_caster_base<type>::name(); }
+};
+
+template <typename type, typename deleter>
+class type_caster<std::unique_ptr<type, deleter>>
+    : public move_only_holder_caster<type, std::unique_ptr<type, deleter>> { };
+
+template <typename type, typename holder_type>
+using type_caster_holder = conditional_t<is_copy_constructible<holder_type>::value,
+                                         copyable_holder_caster<type, holder_type>,
+                                         move_only_holder_caster<type, holder_type>>;
+
+template <typename T, bool Value = false> struct always_construct_holder { static constexpr bool value = Value; };
+
+/// Create a specialization for custom holder types (silently ignores std::shared_ptr)
+#define PYBIND11_DECLARE_HOLDER_TYPE(type, holder_type, ...) \
+    namespace pybind11 { namespace detail { \
+    template <typename type> \
+    struct always_construct_holder<holder_type> : always_construct_holder<void, ##__VA_ARGS__>  { }; \
+    template <typename type> \
+    class type_caster<holder_type, enable_if_t<!is_shared_ptr<holder_type>::value>> \
+        : public type_caster_holder<type, holder_type> { }; \
+    }}
+
+// PYBIND11_DECLARE_HOLDER_TYPE holder types:
+template <typename base, typename holder> struct is_holder_type :
+    std::is_base_of<detail::type_caster_holder<base, holder>, detail::type_caster<holder>> {};
+// Specialization for always-supported unique_ptr holders:
+template <typename base, typename deleter> struct is_holder_type<base, std::unique_ptr<base, deleter>> :
+    std::true_type {};
+
+template <typename T> struct handle_type_name { static PYBIND11_DESCR name() { return _<T>(); } };
+template <> struct handle_type_name<bytes> { static PYBIND11_DESCR name() { return _(PYBIND11_BYTES_NAME); } };
+template <> struct handle_type_name<args> { static PYBIND11_DESCR name() { return _("*args"); } };
+template <> struct handle_type_name<kwargs> { static PYBIND11_DESCR name() { return _("**kwargs"); } };
+
+template <typename type>
+struct pyobject_caster {
+    template <typename T = type, enable_if_t<std::is_same<T, handle>::value, int> = 0>
+    bool load(handle src, bool /* convert */) { value = src; return static_cast<bool>(value); }
+
+    template <typename T = type, enable_if_t<std::is_base_of<object, T>::value, int> = 0>
+    bool load(handle src, bool /* convert */) {
+        if (!isinstance<type>(src))
+            return false;
+        value = reinterpret_borrow<type>(src);
+        return true;
+    }
+
+    static handle cast(const handle &src, return_value_policy /* policy */, handle /* parent */) {
+        return src.inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(type, handle_type_name<type>::name());
+};
+
+template <typename T>
+class type_caster<T, enable_if_t<is_pyobject<T>::value>> : public pyobject_caster<T> { };
+
+// Our conditions for enabling moving are quite restrictive:
+// At compile time:
+// - T needs to be a non-const, non-pointer, non-reference type
+// - type_caster<T>::operator T&() must exist
+// - the type must be move constructible (obviously)
+// At run-time:
+// - if the type is non-copy-constructible, the object must be the sole owner of the type (i.e. it
+//   must have ref_count() == 1)h
+// If any of the above are not satisfied, we fall back to copying.
+template <typename T> using move_is_plain_type = satisfies_none_of<T,
+    std::is_void, std::is_pointer, std::is_reference, std::is_const
+>;
+template <typename T, typename SFINAE = void> struct move_always : std::false_type {};
+template <typename T> struct move_always<T, enable_if_t<all_of<
+    move_is_plain_type<T>,
+    negation<is_copy_constructible<T>>,
+    std::is_move_constructible<T>,
+    std::is_same<decltype(std::declval<make_caster<T>>().operator T&()), T&>
+>::value>> : std::true_type {};
+template <typename T, typename SFINAE = void> struct move_if_unreferenced : std::false_type {};
+template <typename T> struct move_if_unreferenced<T, enable_if_t<all_of<
+    move_is_plain_type<T>,
+    negation<move_always<T>>,
+    std::is_move_constructible<T>,
+    std::is_same<decltype(std::declval<make_caster<T>>().operator T&()), T&>
+>::value>> : std::true_type {};
+template <typename T> using move_never = none_of<move_always<T>, move_if_unreferenced<T>>;
+
+// Detect whether returning a `type` from a cast on type's type_caster is going to result in a
+// reference or pointer to a local variable of the type_caster.  Basically, only
+// non-reference/pointer `type`s and reference/pointers from a type_caster_generic are safe;
+// everything else returns a reference/pointer to a local variable.
+template <typename type> using cast_is_temporary_value_reference = bool_constant<
+    (std::is_reference<type>::value || std::is_pointer<type>::value) &&
+    !std::is_base_of<type_caster_generic, make_caster<type>>::value
+>;
+
+// When a value returned from a C++ function is being cast back to Python, we almost always want to
+// force `policy = move`, regardless of the return value policy the function/method was declared
+// with.  Some classes (most notably Eigen::Ref and related) need to avoid this, and so can do so by
+// specializing this struct.
+template <typename Return, typename SFINAE = void> struct return_value_policy_override {
+    static return_value_policy policy(return_value_policy p) {
+        return !std::is_lvalue_reference<Return>::value && !std::is_pointer<Return>::value
+            ? return_value_policy::move : p;
+    }
+};
+
+// Basic python -> C++ casting; throws if casting fails
+template <typename T, typename SFINAE> type_caster<T, SFINAE> &load_type(type_caster<T, SFINAE> &conv, const handle &handle) {
+    if (!conv.load(handle, true)) {
+#if defined(NDEBUG)
+        throw cast_error("Unable to cast Python instance to C++ type (compile in debug mode for details)");
+#else
+        throw cast_error("Unable to cast Python instance of type " +
+            (std::string) str(handle.get_type()) + " to C++ type '" + type_id<T>() + "'");
+#endif
+    }
+    return conv;
+}
+// Wrapper around the above that also constructs and returns a type_caster
+template <typename T> make_caster<T> load_type(const handle &handle) {
+    make_caster<T> conv;
+    load_type(conv, handle);
+    return conv;
+}
+
+NAMESPACE_END(detail)
+
+// pytype -> C++ type
+template <typename T, detail::enable_if_t<!detail::is_pyobject<T>::value, int> = 0>
+T cast(const handle &handle) {
+    using namespace detail;
+    static_assert(!cast_is_temporary_value_reference<T>::value,
+            "Unable to cast type to reference: value is local to type caster");
+    return cast_op<T>(load_type<T>(handle));
+}
+
+// pytype -> pytype (calls converting constructor)
+template <typename T, detail::enable_if_t<detail::is_pyobject<T>::value, int> = 0>
+T cast(const handle &handle) { return T(reinterpret_borrow<object>(handle)); }
+
+// C++ type -> py::object
+template <typename T, detail::enable_if_t<!detail::is_pyobject<T>::value, int> = 0>
+object cast(const T &value, return_value_policy policy = return_value_policy::automatic_reference,
+            handle parent = handle()) {
+    if (policy == return_value_policy::automatic)
+        policy = std::is_pointer<T>::value ? return_value_policy::take_ownership : return_value_policy::copy;
+    else if (policy == return_value_policy::automatic_reference)
+        policy = std::is_pointer<T>::value ? return_value_policy::reference : return_value_policy::copy;
+    return reinterpret_steal<object>(detail::make_caster<T>::cast(value, policy, parent));
+}
+
+template <typename T> T handle::cast() const { return pybind11::cast<T>(*this); }
+template <> inline void handle::cast() const { return; }
+
+template <typename T>
+detail::enable_if_t<!detail::move_never<T>::value, T> move(object &&obj) {
+    if (obj.ref_count() > 1)
+#if defined(NDEBUG)
+        throw cast_error("Unable to cast Python instance to C++ rvalue: instance has multiple references"
+            " (compile in debug mode for details)");
+#else
+        throw cast_error("Unable to move from Python " + (std::string) str(obj.get_type()) +
+                " instance to C++ " + type_id<T>() + " instance: instance has multiple references");
+#endif
+
+    // Move into a temporary and return that, because the reference may be a local value of `conv`
+    T ret = std::move(detail::load_type<T>(obj).operator T&());
+    return ret;
+}
+
+// Calling cast() on an rvalue calls pybind::cast with the object rvalue, which does:
+// - If we have to move (because T has no copy constructor), do it.  This will fail if the moved
+//   object has multiple references, but trying to copy will fail to compile.
+// - If both movable and copyable, check ref count: if 1, move; otherwise copy
+// - Otherwise (not movable), copy.
+template <typename T> detail::enable_if_t<detail::move_always<T>::value, T> cast(object &&object) {
+    return move<T>(std::move(object));
+}
+template <typename T> detail::enable_if_t<detail::move_if_unreferenced<T>::value, T> cast(object &&object) {
+    if (object.ref_count() > 1)
+        return cast<T>(object);
+    else
+        return move<T>(std::move(object));
+}
+template <typename T> detail::enable_if_t<detail::move_never<T>::value, T> cast(object &&object) {
+    return cast<T>(object);
+}
+
+template <typename T> T object::cast() const & { return pybind11::cast<T>(*this); }
+template <typename T> T object::cast() && { return pybind11::cast<T>(std::move(*this)); }
+template <> inline void object::cast() const & { return; }
+template <> inline void object::cast() && { return; }
+
+NAMESPACE_BEGIN(detail)
+
+// Declared in pytypes.h:
+template <typename T, enable_if_t<!is_pyobject<T>::value, int>>
+object object_or_cast(T &&o) { return pybind11::cast(std::forward<T>(o)); }
+
+struct overload_unused {}; // Placeholder type for the unneeded (and dead code) static variable in the OVERLOAD_INT macro
+template <typename ret_type> using overload_caster_t = conditional_t<
+    cast_is_temporary_value_reference<ret_type>::value, make_caster<ret_type>, overload_unused>;
+
+// Trampoline use: for reference/pointer types to value-converted values, we do a value cast, then
+// store the result in the given variable.  For other types, this is a no-op.
+template <typename T> enable_if_t<cast_is_temporary_value_reference<T>::value, T> cast_ref(object &&o, make_caster<T> &caster) {
+    return cast_op<T>(load_type(caster, o));
+}
+template <typename T> enable_if_t<!cast_is_temporary_value_reference<T>::value, T> cast_ref(object &&, overload_unused &) {
+    pybind11_fail("Internal error: cast_ref fallback invoked"); }
+
+// Trampoline use: Having a pybind11::cast with an invalid reference type is going to static_assert, even
+// though if it's in dead code, so we provide a "trampoline" to pybind11::cast that only does anything in
+// cases where pybind11::cast is valid.
+template <typename T> enable_if_t<!cast_is_temporary_value_reference<T>::value, T> cast_safe(object &&o) {
+    return pybind11::cast<T>(std::move(o)); }
+template <typename T> enable_if_t<cast_is_temporary_value_reference<T>::value, T> cast_safe(object &&) {
+    pybind11_fail("Internal error: cast_safe fallback invoked"); }
+template <> inline void cast_safe<void>(object &&) {}
+
+NAMESPACE_END(detail)
+
+template <return_value_policy policy = return_value_policy::automatic_reference>
+tuple make_tuple() { return tuple(0); }
+
+template <return_value_policy policy = return_value_policy::automatic_reference,
+          typename... Args> tuple make_tuple(Args&&... args_) {
+    constexpr size_t size = sizeof...(Args);
+    std::array<object, size> args {
+        { reinterpret_steal<object>(detail::make_caster<Args>::cast(
+            std::forward<Args>(args_), policy, nullptr))... }
+    };
+    for (size_t i = 0; i < args.size(); i++) {
+        if (!args[i]) {
+#if defined(NDEBUG)
+            throw cast_error("make_tuple(): unable to convert arguments to Python object (compile in debug mode for details)");
+#else
+            std::array<std::string, size> argtypes { {type_id<Args>()...} };
+            throw cast_error("make_tuple(): unable to convert argument of type '" +
+                argtypes[i] + "' to Python object");
+#endif
+        }
+    }
+    tuple result(size);
+    int counter = 0;
+    for (auto &arg_value : args)
+        PyTuple_SET_ITEM(result.ptr(), counter++, arg_value.release().ptr());
+    return result;
+}
+
+/// \ingroup annotations
+/// Annotation for arguments
+struct arg {
+    /// Constructs an argument with the name of the argument; if null or omitted, this is a positional argument.
+    constexpr explicit arg(const char *name = nullptr) : name(name), flag_noconvert(false), flag_none(true) { }
+    /// Assign a value to this argument
+    template <typename T> arg_v operator=(T &&value) const;
+    /// Indicate that the type should not be converted in the type caster
+    arg &noconvert(bool flag = true) { flag_noconvert = flag; return *this; }
+    /// Indicates that the argument should/shouldn't allow None (e.g. for nullable pointer args)
+    arg &none(bool flag = true) { flag_none = flag; return *this; }
+
+    const char *name; ///< If non-null, this is a named kwargs argument
+    bool flag_noconvert : 1; ///< If set, do not allow conversion (requires a supporting type caster!)
+    bool flag_none : 1; ///< If set (the default), allow None to be passed to this argument
+};
+
+/// \ingroup annotations
+/// Annotation for arguments with values
+struct arg_v : arg {
+private:
+    template <typename T>
+    arg_v(arg &&base, T &&x, const char *descr = nullptr)
+        : arg(base),
+          value(reinterpret_steal<object>(
+              detail::make_caster<T>::cast(x, return_value_policy::automatic, {})
+          )),
+          descr(descr)
+#if !defined(NDEBUG)
+        , type(type_id<T>())
+#endif
+    { }
+
+public:
+    /// Direct construction with name, default, and description
+    template <typename T>
+    arg_v(const char *name, T &&x, const char *descr = nullptr)
+        : arg_v(arg(name), std::forward<T>(x), descr) { }
+
+    /// Called internally when invoking `py::arg("a") = value`
+    template <typename T>
+    arg_v(const arg &base, T &&x, const char *descr = nullptr)
+        : arg_v(arg(base), std::forward<T>(x), descr) { }
+
+    /// Same as `arg::noconvert()`, but returns *this as arg_v&, not arg&
+    arg_v &noconvert(bool flag = true) { arg::noconvert(flag); return *this; }
+
+    /// Same as `arg::nonone()`, but returns *this as arg_v&, not arg&
+    arg_v &none(bool flag = true) { arg::none(flag); return *this; }
+
+    /// The default value
+    object value;
+    /// The (optional) description of the default value
+    const char *descr;
+#if !defined(NDEBUG)
+    /// The C++ type name of the default value (only available when compiled in debug mode)
+    std::string type;
+#endif
+};
+
+template <typename T>
+arg_v arg::operator=(T &&value) const { return {std::move(*this), std::forward<T>(value)}; }
+
+/// Alias for backward compatibility -- to be removed in version 2.0
+template <typename /*unused*/> using arg_t = arg_v;
+
+inline namespace literals {
+/** \rst
+    String literal version of `arg`
+ \endrst */
+constexpr arg operator"" _a(const char *name, size_t) { return arg(name); }
+}
+
+NAMESPACE_BEGIN(detail)
+
+// forward declaration (definition in attr.h)
+struct function_record;
+
+/// Internal data associated with a single function call
+struct function_call {
+    function_call(function_record &f, handle p); // Implementation in attr.h
+
+    /// The function data:
+    const function_record &func;
+
+    /// Arguments passed to the function:
+    std::vector<handle> args;
+
+    /// The `convert` value the arguments should be loaded with
+    std::vector<bool> args_convert;
+
+    /// Extra references for the optional `py::args` and/or `py::kwargs` arguments (which, if
+    /// present, are also in `args` but without a reference).
+    object args_ref, kwargs_ref;
+
+    /// The parent, if any
+    handle parent;
+
+    /// If this is a call to an initializer, this argument contains `self`
+    handle init_self;
+};
+
+
+/// Helper class which loads arguments for C++ functions called from Python
+template <typename... Args>
+class argument_loader {
+    using indices = make_index_sequence<sizeof...(Args)>;
+
+    template <typename Arg> using argument_is_args   = std::is_same<intrinsic_t<Arg>, args>;
+    template <typename Arg> using argument_is_kwargs = std::is_same<intrinsic_t<Arg>, kwargs>;
+    // Get args/kwargs argument positions relative to the end of the argument list:
+    static constexpr auto args_pos = constexpr_first<argument_is_args, Args...>() - (int) sizeof...(Args),
+                        kwargs_pos = constexpr_first<argument_is_kwargs, Args...>() - (int) sizeof...(Args);
+
+    static constexpr bool args_kwargs_are_last = kwargs_pos >= - 1 && args_pos >= kwargs_pos - 1;
+
+    static_assert(args_kwargs_are_last, "py::args/py::kwargs are only permitted as the last argument(s) of a function");
+
+public:
+    static constexpr bool has_kwargs = kwargs_pos < 0;
+    static constexpr bool has_args = args_pos < 0;
+
+    static PYBIND11_DESCR arg_names() { return detail::concat(make_caster<Args>::name()...); }
+
+    bool load_args(function_call &call) {
+        return load_impl_sequence(call, indices{});
+    }
+
+    template <typename Return, typename Guard, typename Func>
+    enable_if_t<!std::is_void<Return>::value, Return> call(Func &&f) && {
+        return std::move(*this).template call_impl<Return>(std::forward<Func>(f), indices{}, Guard{});
+    }
+
+    template <typename Return, typename Guard, typename Func>
+    enable_if_t<std::is_void<Return>::value, void_type> call(Func &&f) && {
+        std::move(*this).template call_impl<Return>(std::forward<Func>(f), indices{}, Guard{});
+        return void_type();
+    }
+
+private:
+
+    static bool load_impl_sequence(function_call &, index_sequence<>) { return true; }
+
+    template <size_t... Is>
+    bool load_impl_sequence(function_call &call, index_sequence<Is...>) {
+        for (bool r : {std::get<Is>(argcasters).load(call.args[Is], call.args_convert[Is])...})
+            if (!r)
+                return false;
+        return true;
+    }
+
+    template <typename Return, typename Func, size_t... Is, typename Guard>
+    Return call_impl(Func &&f, index_sequence<Is...>, Guard &&) {
+        return std::forward<Func>(f)(cast_op<Args>(std::move(std::get<Is>(argcasters)))...);
+    }
+
+    std::tuple<make_caster<Args>...> argcasters;
+};
+
+/// Helper class which collects only positional arguments for a Python function call.
+/// A fancier version below can collect any argument, but this one is optimal for simple calls.
+template <return_value_policy policy>
+class simple_collector {
+public:
+    template <typename... Ts>
+    explicit simple_collector(Ts &&...values)
+        : m_args(pybind11::make_tuple<policy>(std::forward<Ts>(values)...)) { }
+
+    const tuple &args() const & { return m_args; }
+    dict kwargs() const { return {}; }
+
+    tuple args() && { return std::move(m_args); }
+
+    /// Call a Python function and pass the collected arguments
+    object call(PyObject *ptr) const {
+        PyObject *result = PyObject_CallObject(ptr, m_args.ptr());
+        if (!result)
+            throw error_already_set();
+        return reinterpret_steal<object>(result);
+    }
+
+private:
+    tuple m_args;
+};
+
+/// Helper class which collects positional, keyword, * and ** arguments for a Python function call
+template <return_value_policy policy>
+class unpacking_collector {
+public:
+    template <typename... Ts>
+    explicit unpacking_collector(Ts &&...values) {
+        // Tuples aren't (easily) resizable so a list is needed for collection,
+        // but the actual function call strictly requires a tuple.
+        auto args_list = list();
+        int _[] = { 0, (process(args_list, std::forward<Ts>(values)), 0)... };
+        ignore_unused(_);
+
+        m_args = std::move(args_list);
+    }
+
+    const tuple &args() const & { return m_args; }
+    const dict &kwargs() const & { return m_kwargs; }
+
+    tuple args() && { return std::move(m_args); }
+    dict kwargs() && { return std::move(m_kwargs); }
+
+    /// Call a Python function and pass the collected arguments
+    object call(PyObject *ptr) const {
+        PyObject *result = PyObject_Call(ptr, m_args.ptr(), m_kwargs.ptr());
+        if (!result)
+            throw error_already_set();
+        return reinterpret_steal<object>(result);
+    }
+
+private:
+    template <typename T>
+    void process(list &args_list, T &&x) {
+        auto o = reinterpret_steal<object>(detail::make_caster<T>::cast(std::forward<T>(x), policy, {}));
+        if (!o) {
+#if defined(NDEBUG)
+            argument_cast_error();
+#else
+            argument_cast_error(std::to_string(args_list.size()), type_id<T>());
+#endif
+        }
+        args_list.append(o);
+    }
+
+    void process(list &args_list, detail::args_proxy ap) {
+        for (const auto &a : ap)
+            args_list.append(a);
+    }
+
+    void process(list &/*args_list*/, arg_v a) {
+        if (!a.name)
+#if defined(NDEBUG)
+            nameless_argument_error();
+#else
+            nameless_argument_error(a.type);
+#endif
+
+        if (m_kwargs.contains(a.name)) {
+#if defined(NDEBUG)
+            multiple_values_error();
+#else
+            multiple_values_error(a.name);
+#endif
+        }
+        if (!a.value) {
+#if defined(NDEBUG)
+            argument_cast_error();
+#else
+            argument_cast_error(a.name, a.type);
+#endif
+        }
+        m_kwargs[a.name] = a.value;
+    }
+
+    void process(list &/*args_list*/, detail::kwargs_proxy kp) {
+        if (!kp)
+            return;
+        for (const auto &k : reinterpret_borrow<dict>(kp)) {
+            if (m_kwargs.contains(k.first)) {
+#if defined(NDEBUG)
+                multiple_values_error();
+#else
+                multiple_values_error(str(k.first));
+#endif
+            }
+            m_kwargs[k.first] = k.second;
+        }
+    }
+
+    [[noreturn]] static void nameless_argument_error() {
+        throw type_error("Got kwargs without a name; only named arguments "
+                         "may be passed via py::arg() to a python function call. "
+                         "(compile in debug mode for details)");
+    }
+    [[noreturn]] static void nameless_argument_error(std::string type) {
+        throw type_error("Got kwargs without a name of type '" + type + "'; only named "
+                         "arguments may be passed via py::arg() to a python function call. ");
+    }
+    [[noreturn]] static void multiple_values_error() {
+        throw type_error("Got multiple values for keyword argument "
+                         "(compile in debug mode for details)");
+    }
+
+    [[noreturn]] static void multiple_values_error(std::string name) {
+        throw type_error("Got multiple values for keyword argument '" + name + "'");
+    }
+
+    [[noreturn]] static void argument_cast_error() {
+        throw cast_error("Unable to convert call argument to Python object "
+                         "(compile in debug mode for details)");
+    }
+
+    [[noreturn]] static void argument_cast_error(std::string name, std::string type) {
+        throw cast_error("Unable to convert call argument '" + name
+                         + "' of type '" + type + "' to Python object");
+    }
+
+private:
+    tuple m_args;
+    dict m_kwargs;
+};
+
+/// Collect only positional arguments for a Python function call
+template <return_value_policy policy, typename... Args,
+          typename = enable_if_t<all_of<is_positional<Args>...>::value>>
+simple_collector<policy> collect_arguments(Args &&...args) {
+    return simple_collector<policy>(std::forward<Args>(args)...);
+}
+
+/// Collect all arguments, including keywords and unpacking (only instantiated when needed)
+template <return_value_policy policy, typename... Args,
+          typename = enable_if_t<!all_of<is_positional<Args>...>::value>>
+unpacking_collector<policy> collect_arguments(Args &&...args) {
+    // Following argument order rules for generalized unpacking according to PEP 448
+    static_assert(
+        constexpr_last<is_positional, Args...>() < constexpr_first<is_keyword_or_ds, Args...>()
+        && constexpr_last<is_s_unpacking, Args...>() < constexpr_first<is_ds_unpacking, Args...>(),
+        "Invalid function call: positional args must precede keywords and ** unpacking; "
+        "* unpacking must precede ** unpacking"
+    );
+    return unpacking_collector<policy>(std::forward<Args>(args)...);
+}
+
+template <typename Derived>
+template <return_value_policy policy, typename... Args>
+object object_api<Derived>::operator()(Args &&...args) const {
+    return detail::collect_arguments<policy>(std::forward<Args>(args)...).call(derived().ptr());
+}
+
+template <typename Derived>
+template <return_value_policy policy, typename... Args>
+object object_api<Derived>::call(Args &&...args) const {
+    return operator()<policy>(std::forward<Args>(args)...);
+}
+
+NAMESPACE_END(detail)
+
+#define PYBIND11_MAKE_OPAQUE(Type) \
+    namespace pybind11 { namespace detail { \
+        template<> class type_caster<Type> : public type_caster_base<Type> { }; \
+    }}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/chrono.h b/ml/dlib/dlib/external/pybind11/include/pybind11/chrono.h
new file mode 100644
index 000000000..95ada76e0
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/chrono.h
@@ -0,0 +1,162 @@
+/*
+    pybind11/chrono.h: Transparent conversion between std::chrono and python's datetime
+
+    Copyright (c) 2016 Trent Houliston <trent@houliston.me> and
+                       Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include <cmath>
+#include <ctime>
+#include <chrono>
+#include <datetime.h>
+
+// Backport the PyDateTime_DELTA functions from Python3.3 if required
+#ifndef PyDateTime_DELTA_GET_DAYS
+#define PyDateTime_DELTA_GET_DAYS(o)         (((PyDateTime_Delta*)o)->days)
+#endif
+#ifndef PyDateTime_DELTA_GET_SECONDS
+#define PyDateTime_DELTA_GET_SECONDS(o)      (((PyDateTime_Delta*)o)->seconds)
+#endif
+#ifndef PyDateTime_DELTA_GET_MICROSECONDS
+#define PyDateTime_DELTA_GET_MICROSECONDS(o) (((PyDateTime_Delta*)o)->microseconds)
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+template <typename type> class duration_caster {
+public:
+    typedef typename type::rep rep;
+    typedef typename type::period period;
+
+    typedef std::chrono::duration<uint_fast32_t, std::ratio<86400>> days;
+
+    bool load(handle src, bool) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) { PyDateTime_IMPORT; }
+
+        if (!src) return false;
+        // If invoked with datetime.delta object
+        if (PyDelta_Check(src.ptr())) {
+            value = type(duration_cast<duration<rep, period>>(
+                  days(PyDateTime_DELTA_GET_DAYS(src.ptr()))
+                + seconds(PyDateTime_DELTA_GET_SECONDS(src.ptr()))
+                + microseconds(PyDateTime_DELTA_GET_MICROSECONDS(src.ptr()))));
+            return true;
+        }
+        // If invoked with a float we assume it is seconds and convert
+        else if (PyFloat_Check(src.ptr())) {
+            value = type(duration_cast<duration<rep, period>>(duration<double>(PyFloat_AsDouble(src.ptr()))));
+            return true;
+        }
+        else return false;
+    }
+
+    // If this is a duration just return it back
+    static const std::chrono::duration<rep, period>& get_duration(const std::chrono::duration<rep, period> &src) {
+        return src;
+    }
+
+    // If this is a time_point get the time_since_epoch
+    template <typename Clock> static std::chrono::duration<rep, period> get_duration(const std::chrono::time_point<Clock, std::chrono::duration<rep, period>> &src) {
+        return src.time_since_epoch();
+    }
+
+    static handle cast(const type &src, return_value_policy /* policy */, handle /* parent */) {
+        using namespace std::chrono;
+
+        // Use overloaded function to get our duration from our source
+        // Works out if it is a duration or time_point and get the duration
+        auto d = get_duration(src);
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) { PyDateTime_IMPORT; }
+
+        // Declare these special duration types so the conversions happen with the correct primitive types (int)
+        using dd_t = duration<int, std::ratio<86400>>;
+        using ss_t = duration<int, std::ratio<1>>;
+        using us_t = duration<int, std::micro>;
+
+        auto dd = duration_cast<dd_t>(d);
+        auto subd = d - dd;
+        auto ss = duration_cast<ss_t>(subd);
+        auto us = duration_cast<us_t>(subd - ss);
+        return PyDelta_FromDSU(dd.count(), ss.count(), us.count());
+    }
+
+    PYBIND11_TYPE_CASTER(type, _("datetime.timedelta"));
+};
+
+// This is for casting times on the system clock into datetime.datetime instances
+template <typename Duration> class type_caster<std::chrono::time_point<std::chrono::system_clock, Duration>> {
+public:
+    typedef std::chrono::time_point<std::chrono::system_clock, Duration> type;
+    bool load(handle src, bool) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) { PyDateTime_IMPORT; }
+
+        if (!src) return false;
+        if (PyDateTime_Check(src.ptr())) {
+            std::tm cal;
+            cal.tm_sec   = PyDateTime_DATE_GET_SECOND(src.ptr());
+            cal.tm_min   = PyDateTime_DATE_GET_MINUTE(src.ptr());
+            cal.tm_hour  = PyDateTime_DATE_GET_HOUR(src.ptr());
+            cal.tm_mday  = PyDateTime_GET_DAY(src.ptr());
+            cal.tm_mon   = PyDateTime_GET_MONTH(src.ptr()) - 1;
+            cal.tm_year  = PyDateTime_GET_YEAR(src.ptr()) - 1900;
+            cal.tm_isdst = -1;
+
+            value = system_clock::from_time_t(std::mktime(&cal)) + microseconds(PyDateTime_DATE_GET_MICROSECOND(src.ptr()));
+            return true;
+        }
+        else return false;
+    }
+
+    static handle cast(const std::chrono::time_point<std::chrono::system_clock, Duration> &src, return_value_policy /* policy */, handle /* parent */) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) { PyDateTime_IMPORT; }
+
+        std::time_t tt = system_clock::to_time_t(src);
+        // this function uses static memory so it's best to copy it out asap just in case
+        // otherwise other code that is using localtime may break this (not just python code)
+        std::tm localtime = *std::localtime(&tt);
+
+        // Declare these special duration types so the conversions happen with the correct primitive types (int)
+        using us_t = duration<int, std::micro>;
+
+        return PyDateTime_FromDateAndTime(localtime.tm_year + 1900,
+                                          localtime.tm_mon + 1,
+                                          localtime.tm_mday,
+                                          localtime.tm_hour,
+                                          localtime.tm_min,
+                                          localtime.tm_sec,
+                                          (duration_cast<us_t>(src.time_since_epoch() % seconds(1))).count());
+    }
+    PYBIND11_TYPE_CASTER(type, _("datetime.datetime"));
+};
+
+// Other clocks that are not the system clock are not measured as datetime.datetime objects
+// since they are not measured on calendar time. So instead we just make them timedeltas
+// Or if they have passed us a time as a float we convert that
+template <typename Clock, typename Duration> class type_caster<std::chrono::time_point<Clock, Duration>>
+: public duration_caster<std::chrono::time_point<Clock, Duration>> {
+};
+
+template <typename Rep, typename Period> class type_caster<std::chrono::duration<Rep, Period>>
+: public duration_caster<std::chrono::duration<Rep, Period>> {
+};
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/common.h b/ml/dlib/dlib/external/pybind11/include/pybind11/common.h
new file mode 100644
index 000000000..6c8a4f1e8
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/common.h
@@ -0,0 +1,2 @@
+#include "detail/common.h"
+#warning "Including 'common.h' is deprecated. It will be removed in v3.0. Use 'pybind11.h'."
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/complex.h b/ml/dlib/dlib/external/pybind11/include/pybind11/complex.h
new file mode 100644
index 000000000..5dac27cc4
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/complex.h
@@ -0,0 +1,61 @@
+/*
+    pybind11/complex.h: Complex number support
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include <complex>
+
+/// glibc defines I as a macro which breaks things, e.g., boost template names
+#ifdef I
+#  undef I
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+template <typename T> struct format_descriptor<std::complex<T>, detail::enable_if_t<std::is_floating_point<T>::value>> {
+    static constexpr const char c = format_descriptor<T>::c;
+    static constexpr const char value[3] = { 'Z', c, '\0' };
+    static std::string format() { return std::string(value); }
+};
+
+template <typename T> constexpr const char format_descriptor<
+    std::complex<T>, detail::enable_if_t<std::is_floating_point<T>::value>>::value[3];
+
+NAMESPACE_BEGIN(detail)
+
+template <typename T> struct is_fmt_numeric<std::complex<T>, detail::enable_if_t<std::is_floating_point<T>::value>> {
+    static constexpr bool value = true;
+    static constexpr int index = is_fmt_numeric<T>::index + 3;
+};
+
+template <typename T> class type_caster<std::complex<T>> {
+public:
+    bool load(handle src, bool convert) {
+        if (!src)
+            return false;
+        if (!convert && !PyComplex_Check(src.ptr()))
+            return false;
+        Py_complex result = PyComplex_AsCComplex(src.ptr());
+        if (result.real == -1.0 && PyErr_Occurred()) {
+            PyErr_Clear();
+            return false;
+        }
+        value = std::complex<T>((T) result.real, (T) result.imag);
+        return true;
+    }
+
+    static handle cast(const std::complex<T> &src, return_value_policy /* policy */, handle /* parent */) {
+        return PyComplex_FromDoubles((double) src.real(), (double) src.imag());
+    }
+
+    PYBIND11_TYPE_CASTER(std::complex<T>, _("complex"));
+};
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/class.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/class.h
new file mode 100644
index 000000000..ff06370fa
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/class.h
@@ -0,0 +1,626 @@
+/*
+    pybind11/detail/class.h: Python C API implementation details for py::class_
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "../attr.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+#if PY_VERSION_HEX >= 0x03030000
+#  define PYBIND11_BUILTIN_QUALNAME
+#  define PYBIND11_SET_OLDPY_QUALNAME(obj, nameobj)
+#else
+// In pre-3.3 Python, we still set __qualname__ so that we can produce reliable function type
+// signatures; in 3.3+ this macro expands to nothing:
+#  define PYBIND11_SET_OLDPY_QUALNAME(obj, nameobj) setattr((PyObject *) obj, "__qualname__", nameobj)
+#endif
+
+inline PyTypeObject *type_incref(PyTypeObject *type) {
+    Py_INCREF(type);
+    return type;
+}
+
+#if !defined(PYPY_VERSION)
+
+/// `pybind11_static_property.__get__()`: Always pass the class instead of the instance.
+extern "C" inline PyObject *pybind11_static_get(PyObject *self, PyObject * /*ob*/, PyObject *cls) {
+    return PyProperty_Type.tp_descr_get(self, cls, cls);
+}
+
+/// `pybind11_static_property.__set__()`: Just like the above `__get__()`.
+extern "C" inline int pybind11_static_set(PyObject *self, PyObject *obj, PyObject *value) {
+    PyObject *cls = PyType_Check(obj) ? obj : (PyObject *) Py_TYPE(obj);
+    return PyProperty_Type.tp_descr_set(self, cls, value);
+}
+
+/** A `static_property` is the same as a `property` but the `__get__()` and `__set__()`
+    methods are modified to always use the object type instead of a concrete instance.
+    Return value: New reference. */
+inline PyTypeObject *make_static_property_type() {
+    constexpr auto *name = "pybind11_static_property";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto heap_type = (PyHeapTypeObject *) PyType_Type.tp_alloc(&PyType_Type, 0);
+    if (!heap_type)
+        pybind11_fail("make_static_property_type(): error allocating type!");
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#endif
+
+    auto type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyProperty_Type);
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+    type->tp_descr_get = pybind11_static_get;
+    type->tp_descr_set = pybind11_static_set;
+
+    if (PyType_Ready(type) < 0)
+        pybind11_fail("make_static_property_type(): failure in PyType_Ready()!");
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    return type;
+}
+
+#else // PYPY
+
+/** PyPy has some issues with the above C API, so we evaluate Python code instead.
+    This function will only be called once so performance isn't really a concern.
+    Return value: New reference. */
+inline PyTypeObject *make_static_property_type() {
+    auto d = dict();
+    PyObject *result = PyRun_String(R"(\
+        class pybind11_static_property(property):
+            def __get__(self, obj, cls):
+                return property.__get__(self, cls, cls)
+
+            def __set__(self, obj, value):
+                cls = obj if isinstance(obj, type) else type(obj)
+                property.__set__(self, cls, value)
+        )", Py_file_input, d.ptr(), d.ptr()
+    );
+    if (result == nullptr)
+        throw error_already_set();
+    Py_DECREF(result);
+    return (PyTypeObject *) d["pybind11_static_property"].cast<object>().release().ptr();
+}
+
+#endif // PYPY
+
+/** Types with static properties need to handle `Type.static_prop = x` in a specific way.
+    By default, Python replaces the `static_property` itself, but for wrapped C++ types
+    we need to call `static_property.__set__()` in order to propagate the new value to
+    the underlying C++ data structure. */
+extern "C" inline int pybind11_meta_setattro(PyObject* obj, PyObject* name, PyObject* value) {
+    // Use `_PyType_Lookup()` instead of `PyObject_GetAttr()` in order to get the raw
+    // descriptor (`property`) instead of calling `tp_descr_get` (`property.__get__()`).
+    PyObject *descr = _PyType_Lookup((PyTypeObject *) obj, name);
+
+    // The following assignment combinations are possible:
+    //   1. `Type.static_prop = value`             --> descr_set: `Type.static_prop.__set__(value)`
+    //   2. `Type.static_prop = other_static_prop` --> setattro:  replace existing `static_prop`
+    //   3. `Type.regular_attribute = value`       --> setattro:  regular attribute assignment
+    const auto static_prop = (PyObject *) get_internals().static_property_type;
+    const auto call_descr_set = descr && PyObject_IsInstance(descr, static_prop)
+                                && !PyObject_IsInstance(value, static_prop);
+    if (call_descr_set) {
+        // Call `static_property.__set__()` instead of replacing the `static_property`.
+#if !defined(PYPY_VERSION)
+        return Py_TYPE(descr)->tp_descr_set(descr, obj, value);
+#else
+        if (PyObject *result = PyObject_CallMethod(descr, "__set__", "OO", obj, value)) {
+            Py_DECREF(result);
+            return 0;
+        } else {
+            return -1;
+        }
+#endif
+    } else {
+        // Replace existing attribute.
+        return PyType_Type.tp_setattro(obj, name, value);
+    }
+}
+
+#if PY_MAJOR_VERSION >= 3
+/**
+ * Python 3's PyInstanceMethod_Type hides itself via its tp_descr_get, which prevents aliasing
+ * methods via cls.attr("m2") = cls.attr("m1"): instead the tp_descr_get returns a plain function,
+ * when called on a class, or a PyMethod, when called on an instance.  Override that behaviour here
+ * to do a special case bypass for PyInstanceMethod_Types.
+ */
+extern "C" inline PyObject *pybind11_meta_getattro(PyObject *obj, PyObject *name) {
+    PyObject *descr = _PyType_Lookup((PyTypeObject *) obj, name);
+    if (descr && PyInstanceMethod_Check(descr)) {
+        Py_INCREF(descr);
+        return descr;
+    }
+    else {
+        return PyType_Type.tp_getattro(obj, name);
+    }
+}
+#endif
+
+/** This metaclass is assigned by default to all pybind11 types and is required in order
+    for static properties to function correctly. Users may override this using `py::metaclass`.
+    Return value: New reference. */
+inline PyTypeObject* make_default_metaclass() {
+    constexpr auto *name = "pybind11_type";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto heap_type = (PyHeapTypeObject *) PyType_Type.tp_alloc(&PyType_Type, 0);
+    if (!heap_type)
+        pybind11_fail("make_default_metaclass(): error allocating metaclass!");
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#endif
+
+    auto type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyType_Type);
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+
+    type->tp_setattro = pybind11_meta_setattro;
+#if PY_MAJOR_VERSION >= 3
+    type->tp_getattro = pybind11_meta_getattro;
+#endif
+
+    if (PyType_Ready(type) < 0)
+        pybind11_fail("make_default_metaclass(): failure in PyType_Ready()!");
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    return type;
+}
+
+/// For multiple inheritance types we need to recursively register/deregister base pointers for any
+/// base classes with pointers that are difference from the instance value pointer so that we can
+/// correctly recognize an offset base class pointer. This calls a function with any offset base ptrs.
+inline void traverse_offset_bases(void *valueptr, const detail::type_info *tinfo, instance *self,
+        bool (*f)(void * /*parentptr*/, instance * /*self*/)) {
+    for (handle h : reinterpret_borrow<tuple>(tinfo->type->tp_bases)) {
+        if (auto parent_tinfo = get_type_info((PyTypeObject *) h.ptr())) {
+            for (auto &c : parent_tinfo->implicit_casts) {
+                if (c.first == tinfo->cpptype) {
+                    auto *parentptr = c.second(valueptr);
+                    if (parentptr != valueptr)
+                        f(parentptr, self);
+                    traverse_offset_bases(parentptr, parent_tinfo, self, f);
+                    break;
+                }
+            }
+        }
+    }
+}
+
+inline bool register_instance_impl(void *ptr, instance *self) {
+    get_internals().registered_instances.emplace(ptr, self);
+    return true; // unused, but gives the same signature as the deregister func
+}
+inline bool deregister_instance_impl(void *ptr, instance *self) {
+    auto &registered_instances = get_internals().registered_instances;
+    auto range = registered_instances.equal_range(ptr);
+    for (auto it = range.first; it != range.second; ++it) {
+        if (Py_TYPE(self) == Py_TYPE(it->second)) {
+            registered_instances.erase(it);
+            return true;
+        }
+    }
+    return false;
+}
+
+inline void register_instance(instance *self, void *valptr, const type_info *tinfo) {
+    register_instance_impl(valptr, self);
+    if (!tinfo->simple_ancestors)
+        traverse_offset_bases(valptr, tinfo, self, register_instance_impl);
+}
+
+inline bool deregister_instance(instance *self, void *valptr, const type_info *tinfo) {
+    bool ret = deregister_instance_impl(valptr, self);
+    if (!tinfo->simple_ancestors)
+        traverse_offset_bases(valptr, tinfo, self, deregister_instance_impl);
+    return ret;
+}
+
+/// Instance creation function for all pybind11 types. It allocates the internal instance layout for
+/// holding C++ objects and holders.  Allocation is done lazily (the first time the instance is cast
+/// to a reference or pointer), and initialization is done by an `__init__` function.
+inline PyObject *make_new_instance(PyTypeObject *type) {
+#if defined(PYPY_VERSION)
+    // PyPy gets tp_basicsize wrong (issue 2482) under multiple inheritance when the first inherited
+    // object is a a plain Python type (i.e. not derived from an extension type).  Fix it.
+    ssize_t instance_size = static_cast<ssize_t>(sizeof(instance));
+    if (type->tp_basicsize < instance_size) {
+        type->tp_basicsize = instance_size;
+    }
+#endif
+    PyObject *self = type->tp_alloc(type, 0);
+    auto inst = reinterpret_cast<instance *>(self);
+    // Allocate the value/holder internals:
+    inst->allocate_layout();
+
+    inst->owned = true;
+
+    return self;
+}
+
+/// Instance creation function for all pybind11 types. It only allocates space for the
+/// C++ object, but doesn't call the constructor -- an `__init__` function must do that.
+extern "C" inline PyObject *pybind11_object_new(PyTypeObject *type, PyObject *, PyObject *) {
+    return make_new_instance(type);
+}
+
+/// An `__init__` function constructs the C++ object. Users should provide at least one
+/// of these using `py::init` or directly with `.def(__init__, ...)`. Otherwise, the
+/// following default function will be used which simply throws an exception.
+extern "C" inline int pybind11_object_init(PyObject *self, PyObject *, PyObject *) {
+    PyTypeObject *type = Py_TYPE(self);
+    std::string msg;
+#if defined(PYPY_VERSION)
+    msg += handle((PyObject *) type).attr("__module__").cast<std::string>() + ".";
+#endif
+    msg += type->tp_name;
+    msg += ": No constructor defined!";
+    PyErr_SetString(PyExc_TypeError, msg.c_str());
+    return -1;
+}
+
+inline void add_patient(PyObject *nurse, PyObject *patient) {
+    auto &internals = get_internals();
+    auto instance = reinterpret_cast<detail::instance *>(nurse);
+    auto &current_patients = internals.patients[nurse];
+    instance->has_patients = true;
+    for (auto &p : current_patients)
+        if (p == patient)
+            return;
+    Py_INCREF(patient);
+    current_patients.push_back(patient);
+}
+
+inline void clear_patients(PyObject *self) {
+    auto instance = reinterpret_cast<detail::instance *>(self);
+    auto &internals = get_internals();
+    auto pos = internals.patients.find(self);
+    assert(pos != internals.patients.end());
+    // Clearing the patients can cause more Python code to run, which
+    // can invalidate the iterator. Extract the vector of patients
+    // from the unordered_map first.
+    auto patients = std::move(pos->second);
+    internals.patients.erase(pos);
+    instance->has_patients = false;
+    for (PyObject *&patient : patients)
+        Py_CLEAR(patient);
+}
+
+/// Clears all internal data from the instance and removes it from registered instances in
+/// preparation for deallocation.
+inline void clear_instance(PyObject *self) {
+    auto instance = reinterpret_cast<detail::instance *>(self);
+
+    // Deallocate any values/holders, if present:
+    for (auto &v_h : values_and_holders(instance)) {
+        if (v_h) {
+
+            // We have to deregister before we call dealloc because, for virtual MI types, we still
+            // need to be able to get the parent pointers.
+            if (v_h.instance_registered() && !deregister_instance(instance, v_h.value_ptr(), v_h.type))
+                pybind11_fail("pybind11_object_dealloc(): Tried to deallocate unregistered instance!");
+
+            if (instance->owned || v_h.holder_constructed())
+                v_h.type->dealloc(v_h);
+        }
+    }
+    // Deallocate the value/holder layout internals:
+    instance->deallocate_layout();
+
+    if (instance->weakrefs)
+        PyObject_ClearWeakRefs(self);
+
+    PyObject **dict_ptr = _PyObject_GetDictPtr(self);
+    if (dict_ptr)
+        Py_CLEAR(*dict_ptr);
+
+    if (instance->has_patients)
+        clear_patients(self);
+}
+
+/// Instance destructor function for all pybind11 types. It calls `type_info.dealloc`
+/// to destroy the C++ object itself, while the rest is Python bookkeeping.
+extern "C" inline void pybind11_object_dealloc(PyObject *self) {
+    clear_instance(self);
+
+    auto type = Py_TYPE(self);
+    type->tp_free(self);
+
+    // `type->tp_dealloc != pybind11_object_dealloc` means that we're being called
+    // as part of a derived type's dealloc, in which case we're not allowed to decref
+    // the type here. For cross-module compatibility, we shouldn't compare directly
+    // with `pybind11_object_dealloc`, but with the common one stashed in internals.
+    auto pybind11_object_type = (PyTypeObject *) get_internals().instance_base;
+    if (type->tp_dealloc == pybind11_object_type->tp_dealloc)
+        Py_DECREF(type);
+}
+
+/** Create the type which can be used as a common base for all classes.  This is
+    needed in order to satisfy Python's requirements for multiple inheritance.
+    Return value: New reference. */
+inline PyObject *make_object_base_type(PyTypeObject *metaclass) {
+    constexpr auto *name = "pybind11_object";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto heap_type = (PyHeapTypeObject *) metaclass->tp_alloc(metaclass, 0);
+    if (!heap_type)
+        pybind11_fail("make_object_base_type(): error allocating type!");
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#endif
+
+    auto type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyBaseObject_Type);
+    type->tp_basicsize = static_cast<ssize_t>(sizeof(instance));
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+
+    type->tp_new = pybind11_object_new;
+    type->tp_init = pybind11_object_init;
+    type->tp_dealloc = pybind11_object_dealloc;
+
+    /* Support weak references (needed for the keep_alive feature) */
+    type->tp_weaklistoffset = offsetof(instance, weakrefs);
+
+    if (PyType_Ready(type) < 0)
+        pybind11_fail("PyType_Ready failed in make_object_base_type():" + error_string());
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    assert(!PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC));
+    return (PyObject *) heap_type;
+}
+
+/// dynamic_attr: Support for `d = instance.__dict__`.
+extern "C" inline PyObject *pybind11_get_dict(PyObject *self, void *) {
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    if (!dict)
+        dict = PyDict_New();
+    Py_XINCREF(dict);
+    return dict;
+}
+
+/// dynamic_attr: Support for `instance.__dict__ = dict()`.
+extern "C" inline int pybind11_set_dict(PyObject *self, PyObject *new_dict, void *) {
+    if (!PyDict_Check(new_dict)) {
+        PyErr_Format(PyExc_TypeError, "__dict__ must be set to a dictionary, not a '%.200s'",
+                     Py_TYPE(new_dict)->tp_name);
+        return -1;
+    }
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    Py_INCREF(new_dict);
+    Py_CLEAR(dict);
+    dict = new_dict;
+    return 0;
+}
+
+/// dynamic_attr: Allow the garbage collector to traverse the internal instance `__dict__`.
+extern "C" inline int pybind11_traverse(PyObject *self, visitproc visit, void *arg) {
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    Py_VISIT(dict);
+    return 0;
+}
+
+/// dynamic_attr: Allow the GC to clear the dictionary.
+extern "C" inline int pybind11_clear(PyObject *self) {
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    Py_CLEAR(dict);
+    return 0;
+}
+
+/// Give instances of this type a `__dict__` and opt into garbage collection.
+inline void enable_dynamic_attributes(PyHeapTypeObject *heap_type) {
+    auto type = &heap_type->ht_type;
+#if defined(PYPY_VERSION)
+    pybind11_fail(std::string(type->tp_name) + ": dynamic attributes are "
+                                               "currently not supported in "
+                                               "conjunction with PyPy!");
+#endif
+    type->tp_flags |= Py_TPFLAGS_HAVE_GC;
+    type->tp_dictoffset = type->tp_basicsize; // place dict at the end
+    type->tp_basicsize += (ssize_t)sizeof(PyObject *); // and allocate enough space for it
+    type->tp_traverse = pybind11_traverse;
+    type->tp_clear = pybind11_clear;
+
+    static PyGetSetDef getset[] = {
+        {const_cast<char*>("__dict__"), pybind11_get_dict, pybind11_set_dict, nullptr, nullptr},
+        {nullptr, nullptr, nullptr, nullptr, nullptr}
+    };
+    type->tp_getset = getset;
+}
+
+/// buffer_protocol: Fill in the view as specified by flags.
+extern "C" inline int pybind11_getbuffer(PyObject *obj, Py_buffer *view, int flags) {
+    // Look for a `get_buffer` implementation in this type's info or any bases (following MRO).
+    type_info *tinfo = nullptr;
+    for (auto type : reinterpret_borrow<tuple>(Py_TYPE(obj)->tp_mro)) {
+        tinfo = get_type_info((PyTypeObject *) type.ptr());
+        if (tinfo && tinfo->get_buffer)
+            break;
+    }
+    if (view == nullptr || obj == nullptr || !tinfo || !tinfo->get_buffer) {
+        if (view)
+            view->obj = nullptr;
+        PyErr_SetString(PyExc_BufferError, "pybind11_getbuffer(): Internal error");
+        return -1;
+    }
+    std::memset(view, 0, sizeof(Py_buffer));
+    buffer_info *info = tinfo->get_buffer(obj, tinfo->get_buffer_data);
+    view->obj = obj;
+    view->ndim = 1;
+    view->internal = info;
+    view->buf = info->ptr;
+    view->itemsize = info->itemsize;
+    view->len = view->itemsize;
+    for (auto s : info->shape)
+        view->len *= s;
+    if ((flags & PyBUF_FORMAT) == PyBUF_FORMAT)
+        view->format = const_cast<char *>(info->format.c_str());
+    if ((flags & PyBUF_STRIDES) == PyBUF_STRIDES) {
+        view->ndim = (int) info->ndim;
+        view->strides = &info->strides[0];
+        view->shape = &info->shape[0];
+    }
+    Py_INCREF(view->obj);
+    return 0;
+}
+
+/// buffer_protocol: Release the resources of the buffer.
+extern "C" inline void pybind11_releasebuffer(PyObject *, Py_buffer *view) {
+    delete (buffer_info *) view->internal;
+}
+
+/// Give this type a buffer interface.
+inline void enable_buffer_protocol(PyHeapTypeObject *heap_type) {
+    heap_type->ht_type.tp_as_buffer = &heap_type->as_buffer;
+#if PY_MAJOR_VERSION < 3
+    heap_type->ht_type.tp_flags |= Py_TPFLAGS_HAVE_NEWBUFFER;
+#endif
+
+    heap_type->as_buffer.bf_getbuffer = pybind11_getbuffer;
+    heap_type->as_buffer.bf_releasebuffer = pybind11_releasebuffer;
+}
+
+/** Create a brand new Python type according to the `type_record` specification.
+    Return value: New reference. */
+inline PyObject* make_new_python_type(const type_record &rec) {
+    auto name = reinterpret_steal<object>(PYBIND11_FROM_STRING(rec.name));
+
+    auto qualname = name;
+    if (rec.scope && !PyModule_Check(rec.scope.ptr()) && hasattr(rec.scope, "__qualname__")) {
+#if PY_MAJOR_VERSION >= 3
+        qualname = reinterpret_steal<object>(
+            PyUnicode_FromFormat("%U.%U", rec.scope.attr("__qualname__").ptr(), name.ptr()));
+#else
+        qualname = str(rec.scope.attr("__qualname__").cast<std::string>() + "." + rec.name);
+#endif
+    }
+
+    object module;
+    if (rec.scope) {
+        if (hasattr(rec.scope, "__module__"))
+            module = rec.scope.attr("__module__");
+        else if (hasattr(rec.scope, "__name__"))
+            module = rec.scope.attr("__name__");
+    }
+
+    auto full_name = c_str(
+#if !defined(PYPY_VERSION)
+        module ? str(module).cast<std::string>() + "." + rec.name :
+#endif
+        rec.name);
+
+    char *tp_doc = nullptr;
+    if (rec.doc && options::show_user_defined_docstrings()) {
+        /* Allocate memory for docstring (using PyObject_MALLOC, since
+           Python will free this later on) */
+        size_t size = strlen(rec.doc) + 1;
+        tp_doc = (char *) PyObject_MALLOC(size);
+        memcpy((void *) tp_doc, rec.doc, size);
+    }
+
+    auto &internals = get_internals();
+    auto bases = tuple(rec.bases);
+    auto base = (bases.size() == 0) ? internals.instance_base
+                                    : bases[0].ptr();
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto metaclass = rec.metaclass.ptr() ? (PyTypeObject *) rec.metaclass.ptr()
+                                         : internals.default_metaclass;
+
+    auto heap_type = (PyHeapTypeObject *) metaclass->tp_alloc(metaclass, 0);
+    if (!heap_type)
+        pybind11_fail(std::string(rec.name) + ": Unable to create type object!");
+
+    heap_type->ht_name = name.release().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = qualname.inc_ref().ptr();
+#endif
+
+    auto type = &heap_type->ht_type;
+    type->tp_name = full_name;
+    type->tp_doc = tp_doc;
+    type->tp_base = type_incref((PyTypeObject *)base);
+    type->tp_basicsize = static_cast<ssize_t>(sizeof(instance));
+    if (bases.size() > 0)
+        type->tp_bases = bases.release().ptr();
+
+    /* Don't inherit base __init__ */
+    type->tp_init = pybind11_object_init;
+
+    /* Supported protocols */
+    type->tp_as_number = &heap_type->as_number;
+    type->tp_as_sequence = &heap_type->as_sequence;
+    type->tp_as_mapping = &heap_type->as_mapping;
+
+    /* Flags */
+    type->tp_flags |= Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+#if PY_MAJOR_VERSION < 3
+    type->tp_flags |= Py_TPFLAGS_CHECKTYPES;
+#endif
+
+    if (rec.dynamic_attr)
+        enable_dynamic_attributes(heap_type);
+
+    if (rec.buffer_protocol)
+        enable_buffer_protocol(heap_type);
+
+    if (PyType_Ready(type) < 0)
+        pybind11_fail(std::string(rec.name) + ": PyType_Ready failed (" + error_string() + ")!");
+
+    assert(rec.dynamic_attr ? PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC)
+                            : !PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC));
+
+    /* Register type with the parent scope */
+    if (rec.scope)
+        setattr(rec.scope, rec.name, (PyObject *) type);
+    else
+        Py_INCREF(type); // Keep it alive forever (reference leak)
+
+    if (module) // Needed by pydoc
+        setattr((PyObject *) type, "__module__", module);
+
+    PYBIND11_SET_OLDPY_QUALNAME(type, qualname);
+
+    return (PyObject *) type;
+}
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/common.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/common.h
new file mode 100644
index 000000000..7d41cd63b
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/common.h
@@ -0,0 +1,802 @@
+/*
+    pybind11/detail/common.h -- Basic macros
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#if !defined(NAMESPACE_BEGIN)
+#  define NAMESPACE_BEGIN(name) namespace name {
+#endif
+#if !defined(NAMESPACE_END)
+#  define NAMESPACE_END(name) }
+#endif
+
+// Robust support for some features and loading modules compiled against different pybind versions
+// requires forcing hidden visibility on pybind code, so we enforce this by setting the attribute on
+// the main `pybind11` namespace.
+#if !defined(PYBIND11_NAMESPACE)
+#  ifdef __GNUG__
+#    define PYBIND11_NAMESPACE pybind11 __attribute__((visibility("hidden")))
+#  else
+#    define PYBIND11_NAMESPACE pybind11
+#  endif
+#endif
+
+#if !defined(_MSC_VER) && !defined(__INTEL_COMPILER)
+#  if __cplusplus >= 201402L
+#    define PYBIND11_CPP14
+#    if __cplusplus > 201402L /* Temporary: should be updated to >= the final C++17 value once known */
+#      define PYBIND11_CPP17
+#    endif
+#  endif
+#elif defined(_MSC_VER)
+// MSVC sets _MSVC_LANG rather than __cplusplus (supposedly until the standard is fully implemented)
+#  if _MSVC_LANG >= 201402L
+#    define PYBIND11_CPP14
+#    if _MSVC_LANG > 201402L && _MSC_VER >= 1910
+#      define PYBIND11_CPP17
+#    endif
+#  endif
+#endif
+
+// Compiler version assertions
+#if defined(__INTEL_COMPILER)
+#  if __INTEL_COMPILER < 1500
+#    error pybind11 requires Intel C++ compiler v15 or newer
+#  endif
+#elif defined(__clang__) && !defined(__apple_build_version__)
+#  if __clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ < 3)
+#    error pybind11 requires clang 3.3 or newer
+#  endif
+#elif defined(__clang__)
+// Apple changes clang version macros to its Xcode version; the first Xcode release based on
+// (upstream) clang 3.3 was Xcode 5:
+#  if __clang_major__ < 5
+#    error pybind11 requires Xcode/clang 5.0 or newer
+#  endif
+#elif defined(__GNUG__)
+#  if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 8)
+#    error pybind11 requires gcc 4.8 or newer
+#  endif
+#elif defined(_MSC_VER)
+// Pybind hits various compiler bugs in 2015u2 and earlier, and also makes use of some stl features
+// (e.g. std::negation) added in 2015u3:
+#  if _MSC_FULL_VER < 190024210
+#    error pybind11 requires MSVC 2015 update 3 or newer
+#  endif
+#endif
+
+#if !defined(PYBIND11_EXPORT)
+#  if defined(WIN32) || defined(_WIN32)
+#    define PYBIND11_EXPORT __declspec(dllexport)
+#  else
+#    define PYBIND11_EXPORT __attribute__ ((visibility("default")))
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  define PYBIND11_NOINLINE __declspec(noinline)
+#else
+#  define PYBIND11_NOINLINE __attribute__ ((noinline))
+#endif
+
+#if defined(PYBIND11_CPP14)
+#  define PYBIND11_DEPRECATED(reason) [[deprecated(reason)]]
+#else
+#  define PYBIND11_DEPRECATED(reason) __attribute__((deprecated(reason)))
+#endif
+
+#define PYBIND11_VERSION_MAJOR 2
+#define PYBIND11_VERSION_MINOR 2
+#define PYBIND11_VERSION_PATCH 2
+
+/// Include Python header, disable linking to pythonX_d.lib on Windows in debug mode
+#if defined(_MSC_VER)
+#  if (PY_MAJOR_VERSION == 3 && PY_MINOR_VERSION < 4)
+#    define HAVE_ROUND 1
+#  endif
+#  pragma warning(push)
+#  pragma warning(disable: 4510 4610 4512 4005)
+#  if defined(_DEBUG)
+#    define PYBIND11_DEBUG_MARKER
+#    undef _DEBUG
+#  endif
+#endif
+
+#include <Python.h>
+#include <frameobject.h>
+#include <pythread.h>
+
+#if defined(_WIN32) && (defined(min) || defined(max))
+#  error Macro clash with min and max -- define NOMINMAX when compiling your program on Windows
+#endif
+
+#if defined(isalnum)
+#  undef isalnum
+#  undef isalpha
+#  undef islower
+#  undef isspace
+#  undef isupper
+#  undef tolower
+#  undef toupper
+#endif
+
+#if defined(_MSC_VER)
+#  if defined(PYBIND11_DEBUG_MARKER)
+#    define _DEBUG
+#    undef PYBIND11_DEBUG_MARKER
+#  endif
+#  pragma warning(pop)
+#endif
+
+#include <cstddef>
+#include <cstring>
+#include <forward_list>
+#include <vector>
+#include <string>
+#include <stdexcept>
+#include <unordered_set>
+#include <unordered_map>
+#include <memory>
+#include <typeindex>
+#include <type_traits>
+
+#if PY_MAJOR_VERSION >= 3 /// Compatibility macros for various Python versions
+#define PYBIND11_INSTANCE_METHOD_NEW(ptr, class_) PyInstanceMethod_New(ptr)
+#define PYBIND11_INSTANCE_METHOD_CHECK PyInstanceMethod_Check
+#define PYBIND11_INSTANCE_METHOD_GET_FUNCTION PyInstanceMethod_GET_FUNCTION
+#define PYBIND11_BYTES_CHECK PyBytes_Check
+#define PYBIND11_BYTES_FROM_STRING PyBytes_FromString
+#define PYBIND11_BYTES_FROM_STRING_AND_SIZE PyBytes_FromStringAndSize
+#define PYBIND11_BYTES_AS_STRING_AND_SIZE PyBytes_AsStringAndSize
+#define PYBIND11_BYTES_AS_STRING PyBytes_AsString
+#define PYBIND11_BYTES_SIZE PyBytes_Size
+#define PYBIND11_LONG_CHECK(o) PyLong_Check(o)
+#define PYBIND11_LONG_AS_LONGLONG(o) PyLong_AsLongLong(o)
+#define PYBIND11_BYTES_NAME "bytes"
+#define PYBIND11_STRING_NAME "str"
+#define PYBIND11_SLICE_OBJECT PyObject
+#define PYBIND11_FROM_STRING PyUnicode_FromString
+#define PYBIND11_STR_TYPE ::pybind11::str
+#define PYBIND11_BOOL_ATTR "__bool__"
+#define PYBIND11_NB_BOOL(ptr) ((ptr)->nb_bool)
+#define PYBIND11_PLUGIN_IMPL(name) \
+    extern "C" PYBIND11_EXPORT PyObject *PyInit_##name()
+
+#else
+#define PYBIND11_INSTANCE_METHOD_NEW(ptr, class_) PyMethod_New(ptr, nullptr, class_)
+#define PYBIND11_INSTANCE_METHOD_CHECK PyMethod_Check
+#define PYBIND11_INSTANCE_METHOD_GET_FUNCTION PyMethod_GET_FUNCTION
+#define PYBIND11_BYTES_CHECK PyString_Check
+#define PYBIND11_BYTES_FROM_STRING PyString_FromString
+#define PYBIND11_BYTES_FROM_STRING_AND_SIZE PyString_FromStringAndSize
+#define PYBIND11_BYTES_AS_STRING_AND_SIZE PyString_AsStringAndSize
+#define PYBIND11_BYTES_AS_STRING PyString_AsString
+#define PYBIND11_BYTES_SIZE PyString_Size
+#define PYBIND11_LONG_CHECK(o) (PyInt_Check(o) || PyLong_Check(o))
+#define PYBIND11_LONG_AS_LONGLONG(o) (PyInt_Check(o) ? (long long) PyLong_AsLong(o) : PyLong_AsLongLong(o))
+#define PYBIND11_BYTES_NAME "str"
+#define PYBIND11_STRING_NAME "unicode"
+#define PYBIND11_SLICE_OBJECT PySliceObject
+#define PYBIND11_FROM_STRING PyString_FromString
+#define PYBIND11_STR_TYPE ::pybind11::bytes
+#define PYBIND11_BOOL_ATTR "__nonzero__"
+#define PYBIND11_NB_BOOL(ptr) ((ptr)->nb_nonzero)
+#define PYBIND11_PLUGIN_IMPL(name) \
+    static PyObject *pybind11_init_wrapper();               \
+    extern "C" PYBIND11_EXPORT void init##name() {          \
+        (void)pybind11_init_wrapper();                      \
+    }                                                       \
+    PyObject *pybind11_init_wrapper()
+#endif
+
+#if PY_VERSION_HEX >= 0x03050000 && PY_VERSION_HEX < 0x03050200
+extern "C" {
+    struct _Py_atomic_address { void *value; };
+    PyAPI_DATA(_Py_atomic_address) _PyThreadState_Current;
+}
+#endif
+
+#define PYBIND11_TRY_NEXT_OVERLOAD ((PyObject *) 1) // special failure return code
+#define PYBIND11_STRINGIFY(x) #x
+#define PYBIND11_TOSTRING(x) PYBIND11_STRINGIFY(x)
+#define PYBIND11_CONCAT(first, second) first##second
+
+/** \rst
+    ***Deprecated in favor of PYBIND11_MODULE***
+
+    This macro creates the entry point that will be invoked when the Python interpreter
+    imports a plugin library. Please create a `module` in the function body and return
+    the pointer to its underlying Python object at the end.
+
+    .. code-block:: cpp
+
+        PYBIND11_PLUGIN(example) {
+            pybind11::module m("example", "pybind11 example plugin");
+            /// Set up bindings here
+            return m.ptr();
+        }
+\endrst */
+#define PYBIND11_PLUGIN(name)                                                  \
+    PYBIND11_DEPRECATED("PYBIND11_PLUGIN is deprecated, use PYBIND11_MODULE")  \
+    static PyObject *pybind11_init();                                          \
+    PYBIND11_PLUGIN_IMPL(name) {                                               \
+        int major, minor;                                                      \
+        if (sscanf(Py_GetVersion(), "%i.%i", &major, &minor) != 2) {           \
+            PyErr_SetString(PyExc_ImportError, "Can't parse Python version."); \
+            return nullptr;                                                    \
+        } else if (major != PY_MAJOR_VERSION || minor != PY_MINOR_VERSION) {   \
+            PyErr_Format(PyExc_ImportError,                                    \
+                         "Python version mismatch: module was compiled for "   \
+                         "version %i.%i, while the interpreter is running "    \
+                         "version %i.%i.", PY_MAJOR_VERSION, PY_MINOR_VERSION, \
+                         major, minor);                                        \
+            return nullptr;                                                    \
+        }                                                                      \
+        try {                                                                  \
+            return pybind11_init();                                            \
+        } catch (pybind11::error_already_set &e) {                             \
+            PyErr_SetString(PyExc_ImportError, e.what());                      \
+            return nullptr;                                                    \
+        } catch (const std::exception &e) {                                    \
+            PyErr_SetString(PyExc_ImportError, e.what());                      \
+            return nullptr;                                                    \
+        }                                                                      \
+    }                                                                          \
+    PyObject *pybind11_init()
+
+/** \rst
+    This macro creates the entry point that will be invoked when the Python interpreter
+    imports an extension module. The module name is given as the fist argument and it
+    should not be in quotes. The second macro argument defines a variable of type
+    `py::module` which can be used to initialize the module.
+
+    .. code-block:: cpp
+
+        PYBIND11_MODULE(example, m) {
+            m.doc() = "pybind11 example module";
+
+            // Add bindings here
+            m.def("foo", []() {
+                return "Hello, World!";
+            });
+        }
+\endrst */
+#define PYBIND11_MODULE(name, variable)                                        \
+    static void PYBIND11_CONCAT(pybind11_init_, name)(pybind11::module &);     \
+    PYBIND11_PLUGIN_IMPL(name) {                                               \
+        int major, minor;                                                      \
+        if (sscanf(Py_GetVersion(), "%i.%i", &major, &minor) != 2) {           \
+            PyErr_SetString(PyExc_ImportError, "Can't parse Python version."); \
+            return nullptr;                                                    \
+        } else if (major != PY_MAJOR_VERSION || minor != PY_MINOR_VERSION) {   \
+            PyErr_Format(PyExc_ImportError,                                    \
+                         "Python version mismatch: module was compiled for "   \
+                         "version %i.%i, while the interpreter is running "    \
+                         "version %i.%i.", PY_MAJOR_VERSION, PY_MINOR_VERSION, \
+                         major, minor);                                        \
+            return nullptr;                                                    \
+        }                                                                      \
+        auto m = pybind11::module(PYBIND11_TOSTRING(name));                    \
+        try {                                                                  \
+            PYBIND11_CONCAT(pybind11_init_, name)(m);                          \
+            return m.ptr();                                                    \
+        } catch (pybind11::error_already_set &e) {                             \
+            PyErr_SetString(PyExc_ImportError, e.what());                      \
+            return nullptr;                                                    \
+        } catch (const std::exception &e) {                                    \
+            PyErr_SetString(PyExc_ImportError, e.what());                      \
+            return nullptr;                                                    \
+        }                                                                      \
+    }                                                                          \
+    void PYBIND11_CONCAT(pybind11_init_, name)(pybind11::module &variable)
+
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+using ssize_t = Py_ssize_t;
+using size_t  = std::size_t;
+
+/// Approach used to cast a previously unknown C++ instance into a Python object
+enum class return_value_policy : uint8_t {
+    /** This is the default return value policy, which falls back to the policy
+        return_value_policy::take_ownership when the return value is a pointer.
+        Otherwise, it uses return_value::move or return_value::copy for rvalue
+        and lvalue references, respectively. See below for a description of what
+        all of these different policies do. */
+    automatic = 0,
+
+    /** As above, but use policy return_value_policy::reference when the return
+        value is a pointer. This is the default conversion policy for function
+        arguments when calling Python functions manually from C++ code (i.e. via
+        handle::operator()). You probably won't need to use this. */
+    automatic_reference,
+
+    /** Reference an existing object (i.e. do not create a new copy) and take
+        ownership. Python will call the destructor and delete operator when the
+        object’s reference count reaches zero. Undefined behavior ensues when
+        the C++ side does the same.. */
+    take_ownership,
+
+    /** Create a new copy of the returned object, which will be owned by
+        Python. This policy is comparably safe because the lifetimes of the two
+        instances are decoupled. */
+    copy,
+
+    /** Use std::move to move the return value contents into a new instance
+        that will be owned by Python. This policy is comparably safe because the
+        lifetimes of the two instances (move source and destination) are
+        decoupled. */
+    move,
+
+    /** Reference an existing object, but do not take ownership. The C++ side
+        is responsible for managing the object’s lifetime and deallocating it
+        when it is no longer used. Warning: undefined behavior will ensue when
+        the C++ side deletes an object that is still referenced and used by
+        Python. */
+    reference,
+
+    /** This policy only applies to methods and properties. It references the
+        object without taking ownership similar to the above
+        return_value_policy::reference policy. In contrast to that policy, the
+        function or property’s implicit this argument (called the parent) is
+        considered to be the the owner of the return value (the child).
+        pybind11 then couples the lifetime of the parent to the child via a
+        reference relationship that ensures that the parent cannot be garbage
+        collected while Python is still using the child. More advanced
+        variations of this scheme are also possible using combinations of
+        return_value_policy::reference and the keep_alive call policy */
+    reference_internal
+};
+
+NAMESPACE_BEGIN(detail)
+
+inline static constexpr int log2(size_t n, int k = 0) { return (n <= 1) ? k : log2(n >> 1, k + 1); }
+
+// Returns the size as a multiple of sizeof(void *), rounded up.
+inline static constexpr size_t size_in_ptrs(size_t s) { return 1 + ((s - 1) >> log2(sizeof(void *))); }
+
+/**
+ * The space to allocate for simple layout instance holders (see below) in multiple of the size of
+ * a pointer (e.g.  2 means 16 bytes on 64-bit architectures).  The default is the minimum required
+ * to holder either a std::unique_ptr or std::shared_ptr (which is almost always
+ * sizeof(std::shared_ptr<T>)).
+ */
+constexpr size_t instance_simple_holder_in_ptrs() {
+    static_assert(sizeof(std::shared_ptr<int>) >= sizeof(std::unique_ptr<int>),
+            "pybind assumes std::shared_ptrs are at least as big as std::unique_ptrs");
+    return size_in_ptrs(sizeof(std::shared_ptr<int>));
+}
+
+// Forward declarations
+struct type_info;
+struct value_and_holder;
+
+struct nonsimple_values_and_holders {
+    void **values_and_holders;
+    uint8_t *status;
+};
+
+/// The 'instance' type which needs to be standard layout (need to be able to use 'offsetof')
+struct instance {
+    PyObject_HEAD
+    /// Storage for pointers and holder; see simple_layout, below, for a description
+    union {
+        void *simple_value_holder[1 + instance_simple_holder_in_ptrs()];
+        nonsimple_values_and_holders nonsimple;
+    };
+    /// Weak references (needed for keep alive):
+    PyObject *weakrefs;
+    /// If true, the pointer is owned which means we're free to manage it with a holder.
+    bool owned : 1;
+    /**
+     * An instance has two possible value/holder layouts.
+     *
+     * Simple layout (when this flag is true), means the `simple_value_holder` is set with a pointer
+     * and the holder object governing that pointer, i.e. [val1*][holder].  This layout is applied
+     * whenever there is no python-side multiple inheritance of bound C++ types *and* the type's
+     * holder will fit in the default space (which is large enough to hold either a std::unique_ptr
+     * or std::shared_ptr).
+     *
+     * Non-simple layout applies when using custom holders that require more space than `shared_ptr`
+     * (which is typically the size of two pointers), or when multiple inheritance is used on the
+     * python side.  Non-simple layout allocates the required amount of memory to have multiple
+     * bound C++ classes as parents.  Under this layout, `nonsimple.values_and_holders` is set to a
+     * pointer to allocated space of the required space to hold a a sequence of value pointers and
+     * holders followed `status`, a set of bit flags (1 byte each), i.e.
+     * [val1*][holder1][val2*][holder2]...[bb...]  where each [block] is rounded up to a multiple of
+     * `sizeof(void *)`.  `nonsimple.holder_constructed` is, for convenience, a pointer to the
+     * beginning of the [bb...] block (but not independently allocated).
+     *
+     * Status bits indicate whether the associated holder is constructed (&
+     * status_holder_constructed) and whether the value pointer is registered (&
+     * status_instance_registered) in `registered_instances`.
+     */
+    bool simple_layout : 1;
+    /// For simple layout, tracks whether the holder has been constructed
+    bool simple_holder_constructed : 1;
+    /// For simple layout, tracks whether the instance is registered in `registered_instances`
+    bool simple_instance_registered : 1;
+    /// If true, get_internals().patients has an entry for this object
+    bool has_patients : 1;
+
+    /// Initializes all of the above type/values/holders data (but not the instance values themselves)
+    void allocate_layout();
+
+    /// Destroys/deallocates all of the above
+    void deallocate_layout();
+
+    /// Returns the value_and_holder wrapper for the given type (or the first, if `find_type`
+    /// omitted).  Returns a default-constructed (with `.inst = nullptr`) object on failure if
+    /// `throw_if_missing` is false.
+    value_and_holder get_value_and_holder(const type_info *find_type = nullptr, bool throw_if_missing = true);
+
+    /// Bit values for the non-simple status flags
+    static constexpr uint8_t status_holder_constructed  = 1;
+    static constexpr uint8_t status_instance_registered = 2;
+};
+
+static_assert(std::is_standard_layout<instance>::value, "Internal error: `pybind11::detail::instance` is not standard layout!");
+
+/// from __cpp_future__ import (convenient aliases from C++14/17)
+#if defined(PYBIND11_CPP14) && (!defined(_MSC_VER) || _MSC_VER >= 1910)
+using std::enable_if_t;
+using std::conditional_t;
+using std::remove_cv_t;
+using std::remove_reference_t;
+#else
+template <bool B, typename T = void> using enable_if_t = typename std::enable_if<B, T>::type;
+template <bool B, typename T, typename F> using conditional_t = typename std::conditional<B, T, F>::type;
+template <typename T> using remove_cv_t = typename std::remove_cv<T>::type;
+template <typename T> using remove_reference_t = typename std::remove_reference<T>::type;
+#endif
+
+/// Index sequences
+#if defined(PYBIND11_CPP14)
+using std::index_sequence;
+using std::make_index_sequence;
+#else
+template<size_t ...> struct index_sequence  { };
+template<size_t N, size_t ...S> struct make_index_sequence_impl : make_index_sequence_impl <N - 1, N - 1, S...> { };
+template<size_t ...S> struct make_index_sequence_impl <0, S...> { typedef index_sequence<S...> type; };
+template<size_t N> using make_index_sequence = typename make_index_sequence_impl<N>::type;
+#endif
+
+/// Make an index sequence of the indices of true arguments
+template <typename ISeq, size_t, bool...> struct select_indices_impl { using type = ISeq; };
+template <size_t... IPrev, size_t I, bool B, bool... Bs> struct select_indices_impl<index_sequence<IPrev...>, I, B, Bs...>
+    : select_indices_impl<conditional_t<B, index_sequence<IPrev..., I>, index_sequence<IPrev...>>, I + 1, Bs...> {};
+template <bool... Bs> using select_indices = typename select_indices_impl<index_sequence<>, 0, Bs...>::type;
+
+/// Backports of std::bool_constant and std::negation to accommodate older compilers
+template <bool B> using bool_constant = std::integral_constant<bool, B>;
+template <typename T> struct negation : bool_constant<!T::value> { };
+
+template <typename...> struct void_t_impl { using type = void; };
+template <typename... Ts> using void_t = typename void_t_impl<Ts...>::type;
+
+/// Compile-time all/any/none of that check the boolean value of all template types
+#ifdef __cpp_fold_expressions
+template <class... Ts> using all_of = bool_constant<(Ts::value && ...)>;
+template <class... Ts> using any_of = bool_constant<(Ts::value || ...)>;
+#elif !defined(_MSC_VER)
+template <bool...> struct bools {};
+template <class... Ts> using all_of = std::is_same<
+    bools<Ts::value..., true>,
+    bools<true, Ts::value...>>;
+template <class... Ts> using any_of = negation<all_of<negation<Ts>...>>;
+#else
+// MSVC has trouble with the above, but supports std::conjunction, which we can use instead (albeit
+// at a slight loss of compilation efficiency).
+template <class... Ts> using all_of = std::conjunction<Ts...>;
+template <class... Ts> using any_of = std::disjunction<Ts...>;
+#endif
+template <class... Ts> using none_of = negation<any_of<Ts...>>;
+
+template <class T, template<class> class... Predicates> using satisfies_all_of = all_of<Predicates<T>...>;
+template <class T, template<class> class... Predicates> using satisfies_any_of = any_of<Predicates<T>...>;
+template <class T, template<class> class... Predicates> using satisfies_none_of = none_of<Predicates<T>...>;
+
+/// Strip the class from a method type
+template <typename T> struct remove_class { };
+template <typename C, typename R, typename... A> struct remove_class<R (C::*)(A...)> { typedef R type(A...); };
+template <typename C, typename R, typename... A> struct remove_class<R (C::*)(A...) const> { typedef R type(A...); };
+
+/// Helper template to strip away type modifiers
+template <typename T> struct intrinsic_type                       { typedef T type; };
+template <typename T> struct intrinsic_type<const T>              { typedef typename intrinsic_type<T>::type type; };
+template <typename T> struct intrinsic_type<T*>                   { typedef typename intrinsic_type<T>::type type; };
+template <typename T> struct intrinsic_type<T&>                   { typedef typename intrinsic_type<T>::type type; };
+template <typename T> struct intrinsic_type<T&&>                  { typedef typename intrinsic_type<T>::type type; };
+template <typename T, size_t N> struct intrinsic_type<const T[N]> { typedef typename intrinsic_type<T>::type type; };
+template <typename T, size_t N> struct intrinsic_type<T[N]>       { typedef typename intrinsic_type<T>::type type; };
+template <typename T> using intrinsic_t = typename intrinsic_type<T>::type;
+
+/// Helper type to replace 'void' in some expressions
+struct void_type { };
+
+/// Helper template which holds a list of types
+template <typename...> struct type_list { };
+
+/// Compile-time integer sum
+#ifdef __cpp_fold_expressions
+template <typename... Ts> constexpr size_t constexpr_sum(Ts... ns) { return (0 + ... + size_t{ns}); }
+#else
+constexpr size_t constexpr_sum() { return 0; }
+template <typename T, typename... Ts>
+constexpr size_t constexpr_sum(T n, Ts... ns) { return size_t{n} + constexpr_sum(ns...); }
+#endif
+
+NAMESPACE_BEGIN(constexpr_impl)
+/// Implementation details for constexpr functions
+constexpr int first(int i) { return i; }
+template <typename T, typename... Ts>
+constexpr int first(int i, T v, Ts... vs) { return v ? i : first(i + 1, vs...); }
+
+constexpr int last(int /*i*/, int result) { return result; }
+template <typename T, typename... Ts>
+constexpr int last(int i, int result, T v, Ts... vs) { return last(i + 1, v ? i : result, vs...); }
+NAMESPACE_END(constexpr_impl)
+
+/// Return the index of the first type in Ts which satisfies Predicate<T>.  Returns sizeof...(Ts) if
+/// none match.
+template <template<typename> class Predicate, typename... Ts>
+constexpr int constexpr_first() { return constexpr_impl::first(0, Predicate<Ts>::value...); }
+
+/// Return the index of the last type in Ts which satisfies Predicate<T>, or -1 if none match.
+template <template<typename> class Predicate, typename... Ts>
+constexpr int constexpr_last() { return constexpr_impl::last(0, -1, Predicate<Ts>::value...); }
+
+/// Return the Nth element from the parameter pack
+template <size_t N, typename T, typename... Ts>
+struct pack_element { using type = typename pack_element<N - 1, Ts...>::type; };
+template <typename T, typename... Ts>
+struct pack_element<0, T, Ts...> { using type = T; };
+
+/// Return the one and only type which matches the predicate, or Default if none match.
+/// If more than one type matches the predicate, fail at compile-time.
+template <template<typename> class Predicate, typename Default, typename... Ts>
+struct exactly_one {
+    static constexpr auto found = constexpr_sum(Predicate<Ts>::value...);
+    static_assert(found <= 1, "Found more than one type matching the predicate");
+
+    static constexpr auto index = found ? constexpr_first<Predicate, Ts...>() : 0;
+    using type = conditional_t<found, typename pack_element<index, Ts...>::type, Default>;
+};
+template <template<typename> class P, typename Default>
+struct exactly_one<P, Default> { using type = Default; };
+
+template <template<typename> class Predicate, typename Default, typename... Ts>
+using exactly_one_t = typename exactly_one<Predicate, Default, Ts...>::type;
+
+/// Defer the evaluation of type T until types Us are instantiated
+template <typename T, typename... /*Us*/> struct deferred_type { using type = T; };
+template <typename T, typename... Us> using deferred_t = typename deferred_type<T, Us...>::type;
+
+/// Like is_base_of, but requires a strict base (i.e. `is_strict_base_of<T, T>::value == false`,
+/// unlike `std::is_base_of`)
+template <typename Base, typename Derived> using is_strict_base_of = bool_constant<
+    std::is_base_of<Base, Derived>::value && !std::is_same<Base, Derived>::value>;
+
+template <template<typename...> class Base>
+struct is_template_base_of_impl {
+    template <typename... Us> static std::true_type check(Base<Us...> *);
+    static std::false_type check(...);
+};
+
+/// Check if a template is the base of a type. For example:
+/// `is_template_base_of<Base, T>` is true if `struct T : Base<U> {}` where U can be anything
+template <template<typename...> class Base, typename T>
+#if !defined(_MSC_VER)
+using is_template_base_of = decltype(is_template_base_of_impl<Base>::check((intrinsic_t<T>*)nullptr));
+#else // MSVC2015 has trouble with decltype in template aliases
+struct is_template_base_of : decltype(is_template_base_of_impl<Base>::check((intrinsic_t<T>*)nullptr)) { };
+#endif
+
+/// Check if T is an instantiation of the template `Class`. For example:
+/// `is_instantiation<shared_ptr, T>` is true if `T == shared_ptr<U>` where U can be anything.
+template <template<typename...> class Class, typename T>
+struct is_instantiation : std::false_type { };
+template <template<typename...> class Class, typename... Us>
+struct is_instantiation<Class, Class<Us...>> : std::true_type { };
+
+/// Check if T is std::shared_ptr<U> where U can be anything
+template <typename T> using is_shared_ptr = is_instantiation<std::shared_ptr, T>;
+
+/// Check if T looks like an input iterator
+template <typename T, typename = void> struct is_input_iterator : std::false_type {};
+template <typename T>
+struct is_input_iterator<T, void_t<decltype(*std::declval<T &>()), decltype(++std::declval<T &>())>>
+    : std::true_type {};
+
+template <typename T> using is_function_pointer = bool_constant<
+    std::is_pointer<T>::value && std::is_function<typename std::remove_pointer<T>::type>::value>;
+
+template <typename F> struct strip_function_object {
+    using type = typename remove_class<decltype(&F::operator())>::type;
+};
+
+// Extracts the function signature from a function, function pointer or lambda.
+template <typename Function, typename F = remove_reference_t<Function>>
+using function_signature_t = conditional_t<
+    std::is_function<F>::value,
+    F,
+    typename conditional_t<
+        std::is_pointer<F>::value || std::is_member_pointer<F>::value,
+        std::remove_pointer<F>,
+        strip_function_object<F>
+    >::type
+>;
+
+/// Returns true if the type looks like a lambda: that is, isn't a function, pointer or member
+/// pointer.  Note that this can catch all sorts of other things, too; this is intended to be used
+/// in a place where passing a lambda makes sense.
+template <typename T> using is_lambda = satisfies_none_of<remove_reference_t<T>,
+        std::is_function, std::is_pointer, std::is_member_pointer>;
+
+/// Ignore that a variable is unused in compiler warnings
+inline void ignore_unused(const int *) { }
+
+/// Apply a function over each element of a parameter pack
+#ifdef __cpp_fold_expressions
+#define PYBIND11_EXPAND_SIDE_EFFECTS(PATTERN) (((PATTERN), void()), ...)
+#else
+using expand_side_effects = bool[];
+#define PYBIND11_EXPAND_SIDE_EFFECTS(PATTERN) pybind11::detail::expand_side_effects{ ((PATTERN), void(), false)..., false }
+#endif
+
+NAMESPACE_END(detail)
+
+/// C++ bindings of builtin Python exceptions
+class builtin_exception : public std::runtime_error {
+public:
+    using std::runtime_error::runtime_error;
+    /// Set the error using the Python C API
+    virtual void set_error() const = 0;
+};
+
+#define PYBIND11_RUNTIME_EXCEPTION(name, type) \
+    class name : public builtin_exception { public: \
+        using builtin_exception::builtin_exception; \
+        name() : name("") { } \
+        void set_error() const override { PyErr_SetString(type, what()); } \
+    };
+
+PYBIND11_RUNTIME_EXCEPTION(stop_iteration, PyExc_StopIteration)
+PYBIND11_RUNTIME_EXCEPTION(index_error, PyExc_IndexError)
+PYBIND11_RUNTIME_EXCEPTION(key_error, PyExc_KeyError)
+PYBIND11_RUNTIME_EXCEPTION(value_error, PyExc_ValueError)
+PYBIND11_RUNTIME_EXCEPTION(type_error, PyExc_TypeError)
+PYBIND11_RUNTIME_EXCEPTION(cast_error, PyExc_RuntimeError) /// Thrown when pybind11::cast or handle::call fail due to a type casting error
+PYBIND11_RUNTIME_EXCEPTION(reference_cast_error, PyExc_RuntimeError) /// Used internally
+
+[[noreturn]] PYBIND11_NOINLINE inline void pybind11_fail(const char *reason) { throw std::runtime_error(reason); }
+[[noreturn]] PYBIND11_NOINLINE inline void pybind11_fail(const std::string &reason) { throw std::runtime_error(reason); }
+
+template <typename T, typename SFINAE = void> struct format_descriptor { };
+
+NAMESPACE_BEGIN(detail)
+// Returns the index of the given type in the type char array below, and in the list in numpy.h
+// The order here is: bool; 8 ints ((signed,unsigned)x(8,16,32,64)bits); float,double,long double;
+// complex float,double,long double.  Note that the long double types only participate when long
+// double is actually longer than double (it isn't under MSVC).
+// NB: not only the string below but also complex.h and numpy.h rely on this order.
+template <typename T, typename SFINAE = void> struct is_fmt_numeric { static constexpr bool value = false; };
+template <typename T> struct is_fmt_numeric<T, enable_if_t<std::is_arithmetic<T>::value>> {
+    static constexpr bool value = true;
+    static constexpr int index = std::is_same<T, bool>::value ? 0 : 1 + (
+        std::is_integral<T>::value ? detail::log2(sizeof(T))*2 + std::is_unsigned<T>::value : 8 + (
+        std::is_same<T, double>::value ? 1 : std::is_same<T, long double>::value ? 2 : 0));
+};
+NAMESPACE_END(detail)
+
+template <typename T> struct format_descriptor<T, detail::enable_if_t<std::is_arithmetic<T>::value>> {
+    static constexpr const char c = "?bBhHiIqQfdg"[detail::is_fmt_numeric<T>::index];
+    static constexpr const char value[2] = { c, '\0' };
+    static std::string format() { return std::string(1, c); }
+};
+
+template <typename T> constexpr const char format_descriptor<
+    T, detail::enable_if_t<std::is_arithmetic<T>::value>>::value[2];
+
+/// RAII wrapper that temporarily clears any Python error state
+struct error_scope {
+    PyObject *type, *value, *trace;
+    error_scope() { PyErr_Fetch(&type, &value, &trace); }
+    ~error_scope() { PyErr_Restore(type, value, trace); }
+};
+
+/// Dummy destructor wrapper that can be used to expose classes with a private destructor
+struct nodelete { template <typename T> void operator()(T*) { } };
+
+// overload_cast requires variable templates: C++14
+#if defined(PYBIND11_CPP14)
+#define PYBIND11_OVERLOAD_CAST 1
+
+NAMESPACE_BEGIN(detail)
+template <typename... Args>
+struct overload_cast_impl {
+    constexpr overload_cast_impl() {} // MSVC 2015 needs this
+
+    template <typename Return>
+    constexpr auto operator()(Return (*pf)(Args...)) const noexcept
+                              -> decltype(pf) { return pf; }
+
+    template <typename Return, typename Class>
+    constexpr auto operator()(Return (Class::*pmf)(Args...), std::false_type = {}) const noexcept
+                              -> decltype(pmf) { return pmf; }
+
+    template <typename Return, typename Class>
+    constexpr auto operator()(Return (Class::*pmf)(Args...) const, std::true_type) const noexcept
+                              -> decltype(pmf) { return pmf; }
+};
+NAMESPACE_END(detail)
+
+/// Syntax sugar for resolving overloaded function pointers:
+///  - regular: static_cast<Return (Class::*)(Arg0, Arg1, Arg2)>(&Class::func)
+///  - sweet:   overload_cast<Arg0, Arg1, Arg2>(&Class::func)
+template <typename... Args>
+static constexpr detail::overload_cast_impl<Args...> overload_cast = {};
+// MSVC 2015 only accepts this particular initialization syntax for this variable template.
+
+/// Const member function selector for overload_cast
+///  - regular: static_cast<Return (Class::*)(Arg) const>(&Class::func)
+///  - sweet:   overload_cast<Arg>(&Class::func, const_)
+static constexpr auto const_ = std::true_type{};
+
+#else // no overload_cast: providing something that static_assert-fails:
+template <typename... Args> struct overload_cast {
+    static_assert(detail::deferred_t<std::false_type, Args...>::value,
+                  "pybind11::overload_cast<...> requires compiling in C++14 mode");
+};
+#endif // overload_cast
+
+NAMESPACE_BEGIN(detail)
+
+// Adaptor for converting arbitrary container arguments into a vector; implicitly convertible from
+// any standard container (or C-style array) supporting std::begin/std::end, any singleton
+// arithmetic type (if T is arithmetic), or explicitly constructible from an iterator pair.
+template <typename T>
+class any_container {
+    std::vector<T> v;
+public:
+    any_container() = default;
+
+    // Can construct from a pair of iterators
+    template <typename It, typename = enable_if_t<is_input_iterator<It>::value>>
+    any_container(It first, It last) : v(first, last) { }
+
+    // Implicit conversion constructor from any arbitrary container type with values convertible to T
+    template <typename Container, typename = enable_if_t<std::is_convertible<decltype(*std::begin(std::declval<const Container &>())), T>::value>>
+    any_container(const Container &c) : any_container(std::begin(c), std::end(c)) { }
+
+    // initializer_list's aren't deducible, so don't get matched by the above template; we need this
+    // to explicitly allow implicit conversion from one:
+    template <typename TIn, typename = enable_if_t<std::is_convertible<TIn, T>::value>>
+    any_container(const std::initializer_list<TIn> &c) : any_container(c.begin(), c.end()) { }
+
+    // Avoid copying if given an rvalue vector of the correct type.
+    any_container(std::vector<T> &&v) : v(std::move(v)) { }
+
+    // Moves the vector out of an rvalue any_container
+    operator std::vector<T> &&() && { return std::move(v); }
+
+    // Dereferencing obtains a reference to the underlying vector
+    std::vector<T> &operator*() { return v; }
+    const std::vector<T> &operator*() const { return v; }
+
+    // -> lets you call methods on the underlying vector
+    std::vector<T> *operator->() { return &v; }
+    const std::vector<T> *operator->() const { return &v; }
+};
+
+NAMESPACE_END(detail)
+
+
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/descr.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/descr.h
new file mode 100644
index 000000000..e3bf2ba97
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/descr.h
@@ -0,0 +1,185 @@
+/*
+    pybind11/detail/descr.h: Helper type for concatenating type signatures
+    either at runtime (C++11) or compile time (C++14)
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "common.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/* Concatenate type signatures at compile time using C++14 */
+#if defined(PYBIND11_CPP14) && !defined(_MSC_VER)
+#define PYBIND11_CONSTEXPR_DESCR
+
+template <size_t Size1, size_t Size2> class descr {
+    template <size_t Size1_, size_t Size2_> friend class descr;
+public:
+    constexpr descr(char const (&text) [Size1+1], const std::type_info * const (&types)[Size2+1])
+        : descr(text, types,
+                make_index_sequence<Size1>(),
+                make_index_sequence<Size2>()) { }
+
+    constexpr const char *text() const { return m_text; }
+    constexpr const std::type_info * const * types() const { return m_types; }
+
+    template <size_t OtherSize1, size_t OtherSize2>
+    constexpr descr<Size1 + OtherSize1, Size2 + OtherSize2> operator+(const descr<OtherSize1, OtherSize2> &other) const {
+        return concat(other,
+                      make_index_sequence<Size1>(),
+                      make_index_sequence<Size2>(),
+                      make_index_sequence<OtherSize1>(),
+                      make_index_sequence<OtherSize2>());
+    }
+
+protected:
+    template <size_t... Indices1, size_t... Indices2>
+    constexpr descr(
+        char const (&text) [Size1+1],
+        const std::type_info * const (&types) [Size2+1],
+        index_sequence<Indices1...>, index_sequence<Indices2...>)
+        : m_text{text[Indices1]..., '\0'},
+          m_types{types[Indices2]...,  nullptr } {}
+
+    template <size_t OtherSize1, size_t OtherSize2, size_t... Indices1,
+              size_t... Indices2, size_t... OtherIndices1, size_t... OtherIndices2>
+    constexpr descr<Size1 + OtherSize1, Size2 + OtherSize2>
+    concat(const descr<OtherSize1, OtherSize2> &other,
+           index_sequence<Indices1...>, index_sequence<Indices2...>,
+           index_sequence<OtherIndices1...>, index_sequence<OtherIndices2...>) const {
+        return descr<Size1 + OtherSize1, Size2 + OtherSize2>(
+            { m_text[Indices1]..., other.m_text[OtherIndices1]..., '\0' },
+            { m_types[Indices2]..., other.m_types[OtherIndices2]..., nullptr }
+        );
+    }
+
+protected:
+    char m_text[Size1 + 1];
+    const std::type_info * m_types[Size2 + 1];
+};
+
+template <size_t Size> constexpr descr<Size - 1, 0> _(char const(&text)[Size]) {
+    return descr<Size - 1, 0>(text, { nullptr });
+}
+
+template <size_t Rem, size_t... Digits> struct int_to_str : int_to_str<Rem/10, Rem%10, Digits...> { };
+template <size_t...Digits> struct int_to_str<0, Digits...> {
+    static constexpr auto digits = descr<sizeof...(Digits), 0>({ ('0' + Digits)..., '\0' }, { nullptr });
+};
+
+// Ternary description (like std::conditional)
+template <bool B, size_t Size1, size_t Size2>
+constexpr enable_if_t<B, descr<Size1 - 1, 0>> _(char const(&text1)[Size1], char const(&)[Size2]) {
+    return _(text1);
+}
+template <bool B, size_t Size1, size_t Size2>
+constexpr enable_if_t<!B, descr<Size2 - 1, 0>> _(char const(&)[Size1], char const(&text2)[Size2]) {
+    return _(text2);
+}
+template <bool B, size_t SizeA1, size_t SizeA2, size_t SizeB1, size_t SizeB2>
+constexpr enable_if_t<B, descr<SizeA1, SizeA2>> _(descr<SizeA1, SizeA2> d, descr<SizeB1, SizeB2>) { return d; }
+template <bool B, size_t SizeA1, size_t SizeA2, size_t SizeB1, size_t SizeB2>
+constexpr enable_if_t<!B, descr<SizeB1, SizeB2>> _(descr<SizeA1, SizeA2>, descr<SizeB1, SizeB2> d) { return d; }
+
+template <size_t Size> auto constexpr _() -> decltype(int_to_str<Size / 10, Size % 10>::digits) {
+    return int_to_str<Size / 10, Size % 10>::digits;
+}
+
+template <typename Type> constexpr descr<1, 1> _() {
+    return descr<1, 1>({ '%', '\0' }, { &typeid(Type), nullptr });
+}
+
+inline constexpr descr<0, 0> concat() { return _(""); }
+template <size_t Size1, size_t Size2, typename... Args> auto constexpr concat(descr<Size1, Size2> descr) { return descr; }
+template <size_t Size1, size_t Size2, typename... Args> auto constexpr concat(descr<Size1, Size2> descr, Args&&... args) { return descr + _(", ") + concat(args...); }
+template <size_t Size1, size_t Size2> auto constexpr type_descr(descr<Size1, Size2> descr) { return _("{") + descr + _("}"); }
+
+#define PYBIND11_DESCR constexpr auto
+
+#else /* Simpler C++11 implementation based on run-time memory allocation and copying */
+
+class descr {
+public:
+    PYBIND11_NOINLINE descr(const char *text, const std::type_info * const * types) {
+        size_t nChars = len(text), nTypes = len(types);
+        m_text  = new char[nChars];
+        m_types = new const std::type_info *[nTypes];
+        memcpy(m_text, text, nChars * sizeof(char));
+        memcpy(m_types, types, nTypes * sizeof(const std::type_info *));
+    }
+
+    PYBIND11_NOINLINE descr operator+(descr &&d2) && {
+        descr r;
+
+        size_t nChars1 = len(m_text),    nTypes1 = len(m_types);
+        size_t nChars2 = len(d2.m_text), nTypes2 = len(d2.m_types);
+
+        r.m_text  = new char[nChars1 + nChars2 - 1];
+        r.m_types = new const std::type_info *[nTypes1 + nTypes2 - 1];
+        memcpy(r.m_text, m_text, (nChars1-1) * sizeof(char));
+        memcpy(r.m_text + nChars1 - 1, d2.m_text, nChars2 * sizeof(char));
+        memcpy(r.m_types, m_types, (nTypes1-1) * sizeof(std::type_info *));
+        memcpy(r.m_types + nTypes1 - 1, d2.m_types, nTypes2 * sizeof(std::type_info *));
+
+        delete[] m_text;    delete[] m_types;
+        delete[] d2.m_text; delete[] d2.m_types;
+
+        return r;
+    }
+
+    char *text() { return m_text; }
+    const std::type_info * * types() { return m_types; }
+
+protected:
+    PYBIND11_NOINLINE descr() { }
+
+    template <typename T> static size_t len(const T *ptr) { // return length including null termination
+        const T *it = ptr;
+        while (*it++ != (T) 0)
+            ;
+        return static_cast<size_t>(it - ptr);
+    }
+
+    const std::type_info **m_types = nullptr;
+    char *m_text = nullptr;
+};
+
+/* The 'PYBIND11_NOINLINE inline' combinations below are intentional to get the desired linkage while producing as little object code as possible */
+
+PYBIND11_NOINLINE inline descr _(const char *text) {
+    const std::type_info *types[1] = { nullptr };
+    return descr(text, types);
+}
+
+template <bool B> PYBIND11_NOINLINE enable_if_t<B, descr> _(const char *text1, const char *) { return _(text1); }
+template <bool B> PYBIND11_NOINLINE enable_if_t<!B, descr> _(char const *, const char *text2) { return _(text2); }
+template <bool B> PYBIND11_NOINLINE enable_if_t<B, descr> _(descr d, descr) { return d; }
+template <bool B> PYBIND11_NOINLINE enable_if_t<!B, descr> _(descr, descr d) { return d; }
+
+template <typename Type> PYBIND11_NOINLINE descr _() {
+    const std::type_info *types[2] = { &typeid(Type), nullptr };
+    return descr("%", types);
+}
+
+template <size_t Size> PYBIND11_NOINLINE descr _() {
+    const std::type_info *types[1] = { nullptr };
+    return descr(std::to_string(Size).c_str(), types);
+}
+
+PYBIND11_NOINLINE inline descr concat() { return _(""); }
+PYBIND11_NOINLINE inline descr concat(descr &&d) { return d; }
+template <typename... Args> PYBIND11_NOINLINE descr concat(descr &&d, Args&&... args) { return std::move(d) + _(", ") + concat(std::forward<Args>(args)...); }
+PYBIND11_NOINLINE inline descr type_descr(descr&& d) { return _("{") + std::move(d) + _("}"); }
+
+#define PYBIND11_DESCR ::pybind11::detail::descr
+#endif
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/init.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/init.h
new file mode 100644
index 000000000..82f740760
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/init.h
@@ -0,0 +1,335 @@
+/*
+    pybind11/detail/init.h: init factory function implementation and support code.
+
+    Copyright (c) 2017 Jason Rhinelander <jason@imaginary.ca>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "class.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+template <>
+class type_caster<value_and_holder> {
+public:
+    bool load(handle h, bool) {
+        value = reinterpret_cast<value_and_holder *>(h.ptr());
+        return true;
+    }
+
+    template <typename> using cast_op_type = value_and_holder &;
+    operator value_and_holder &() { return *value; }
+    static PYBIND11_DESCR name() { return type_descr(_<value_and_holder>()); }
+
+private:
+    value_and_holder *value = nullptr;
+};
+
+NAMESPACE_BEGIN(initimpl)
+
+inline void no_nullptr(void *ptr) {
+    if (!ptr) throw type_error("pybind11::init(): factory function returned nullptr");
+}
+
+// Implementing functions for all forms of py::init<...> and py::init(...)
+template <typename Class> using Cpp = typename Class::type;
+template <typename Class> using Alias = typename Class::type_alias;
+template <typename Class> using Holder = typename Class::holder_type;
+
+template <typename Class> using is_alias_constructible = std::is_constructible<Alias<Class>, Cpp<Class> &&>;
+
+// Takes a Cpp pointer and returns true if it actually is a polymorphic Alias instance.
+template <typename Class, enable_if_t<Class::has_alias, int> = 0>
+bool is_alias(Cpp<Class> *ptr) {
+    return dynamic_cast<Alias<Class> *>(ptr) != nullptr;
+}
+// Failing fallback version of the above for a no-alias class (always returns false)
+template <typename /*Class*/>
+constexpr bool is_alias(void *) { return false; }
+
+// Constructs and returns a new object; if the given arguments don't map to a constructor, we fall
+// back to brace aggregate initiailization so that for aggregate initialization can be used with
+// py::init, e.g.  `py::init<int, int>` to initialize a `struct T { int a; int b; }`.  For
+// non-aggregate types, we need to use an ordinary T(...) constructor (invoking as `T{...}` usually
+// works, but will not do the expected thing when `T` has an `initializer_list<T>` constructor).
+template <typename Class, typename... Args, detail::enable_if_t<std::is_constructible<Class, Args...>::value, int> = 0>
+inline Class *construct_or_initialize(Args &&...args) { return new Class(std::forward<Args>(args)...); }
+template <typename Class, typename... Args, detail::enable_if_t<!std::is_constructible<Class, Args...>::value, int> = 0>
+inline Class *construct_or_initialize(Args &&...args) { return new Class{std::forward<Args>(args)...}; }
+
+// Attempts to constructs an alias using a `Alias(Cpp &&)` constructor.  This allows types with
+// an alias to provide only a single Cpp factory function as long as the Alias can be
+// constructed from an rvalue reference of the base Cpp type.  This means that Alias classes
+// can, when appropriate, simply define a `Alias(Cpp &&)` constructor rather than needing to
+// inherit all the base class constructors.
+template <typename Class>
+void construct_alias_from_cpp(std::true_type /*is_alias_constructible*/,
+                              value_and_holder &v_h, Cpp<Class> &&base) {
+    v_h.value_ptr() = new Alias<Class>(std::move(base));
+}
+template <typename Class>
+[[noreturn]] void construct_alias_from_cpp(std::false_type /*!is_alias_constructible*/,
+                                           value_and_holder &, Cpp<Class> &&) {
+    throw type_error("pybind11::init(): unable to convert returned instance to required "
+                     "alias class: no `Alias<Class>(Class &&)` constructor available");
+}
+
+// Error-generating fallback for factories that don't match one of the below construction
+// mechanisms.
+template <typename Class>
+void construct(...) {
+    static_assert(!std::is_same<Class, Class>::value /* always false */,
+            "pybind11::init(): init function must return a compatible pointer, "
+            "holder, or value");
+}
+
+// Pointer return v1: the factory function returns a class pointer for a registered class.
+// If we don't need an alias (because this class doesn't have one, or because the final type is
+// inherited on the Python side) we can simply take over ownership.  Otherwise we need to try to
+// construct an Alias from the returned base instance.
+template <typename Class>
+void construct(value_and_holder &v_h, Cpp<Class> *ptr, bool need_alias) {
+    no_nullptr(ptr);
+    if (Class::has_alias && need_alias && !is_alias<Class>(ptr)) {
+        // We're going to try to construct an alias by moving the cpp type.  Whether or not
+        // that succeeds, we still need to destroy the original cpp pointer (either the
+        // moved away leftover, if the alias construction works, or the value itself if we
+        // throw an error), but we can't just call `delete ptr`: it might have a special
+        // deleter, or might be shared_from_this.  So we construct a holder around it as if
+        // it was a normal instance, then steal the holder away into a local variable; thus
+        // the holder and destruction happens when we leave the C++ scope, and the holder
+        // class gets to handle the destruction however it likes.
+        v_h.value_ptr() = ptr;
+        v_h.set_instance_registered(true); // To prevent init_instance from registering it
+        v_h.type->init_instance(v_h.inst, nullptr); // Set up the holder
+        Holder<Class> temp_holder(std::move(v_h.holder<Holder<Class>>())); // Steal the holder
+        v_h.type->dealloc(v_h); // Destroys the moved-out holder remains, resets value ptr to null
+        v_h.set_instance_registered(false);
+
+        construct_alias_from_cpp<Class>(is_alias_constructible<Class>{}, v_h, std::move(*ptr));
+    } else {
+        // Otherwise the type isn't inherited, so we don't need an Alias
+        v_h.value_ptr() = ptr;
+    }
+}
+
+// Pointer return v2: a factory that always returns an alias instance ptr.  We simply take over
+// ownership of the pointer.
+template <typename Class, enable_if_t<Class::has_alias, int> = 0>
+void construct(value_and_holder &v_h, Alias<Class> *alias_ptr, bool) {
+    no_nullptr(alias_ptr);
+    v_h.value_ptr() = static_cast<Cpp<Class> *>(alias_ptr);
+}
+
+// Holder return: copy its pointer, and move or copy the returned holder into the new instance's
+// holder.  This also handles types like std::shared_ptr<T> and std::unique_ptr<T> where T is a
+// derived type (through those holder's implicit conversion from derived class holder constructors).
+template <typename Class>
+void construct(value_and_holder &v_h, Holder<Class> holder, bool need_alias) {
+    auto *ptr = holder_helper<Holder<Class>>::get(holder);
+    // If we need an alias, check that the held pointer is actually an alias instance
+    if (Class::has_alias && need_alias && !is_alias<Class>(ptr))
+        throw type_error("pybind11::init(): construction failed: returned holder-wrapped instance "
+                         "is not an alias instance");
+
+    v_h.value_ptr() = ptr;
+    v_h.type->init_instance(v_h.inst, &holder);
+}
+
+// return-by-value version 1: returning a cpp class by value.  If the class has an alias and an
+// alias is required the alias must have an `Alias(Cpp &&)` constructor so that we can construct
+// the alias from the base when needed (i.e. because of Python-side inheritance).  When we don't
+// need it, we simply move-construct the cpp value into a new instance.
+template <typename Class>
+void construct(value_and_holder &v_h, Cpp<Class> &&result, bool need_alias) {
+    static_assert(std::is_move_constructible<Cpp<Class>>::value,
+        "pybind11::init() return-by-value factory function requires a movable class");
+    if (Class::has_alias && need_alias)
+        construct_alias_from_cpp<Class>(is_alias_constructible<Class>{}, v_h, std::move(result));
+    else
+        v_h.value_ptr() = new Cpp<Class>(std::move(result));
+}
+
+// return-by-value version 2: returning a value of the alias type itself.  We move-construct an
+// Alias instance (even if no the python-side inheritance is involved).  The is intended for
+// cases where Alias initialization is always desired.
+template <typename Class>
+void construct(value_and_holder &v_h, Alias<Class> &&result, bool) {
+    static_assert(std::is_move_constructible<Alias<Class>>::value,
+        "pybind11::init() return-by-alias-value factory function requires a movable alias class");
+    v_h.value_ptr() = new Alias<Class>(std::move(result));
+}
+
+// Implementing class for py::init<...>()
+template <typename... Args>
+struct constructor {
+    template <typename Class, typename... Extra, enable_if_t<!Class::has_alias, int> = 0>
+    static void execute(Class &cl, const Extra&... extra) {
+        cl.def("__init__", [](value_and_holder &v_h, Args... args) {
+            v_h.value_ptr() = construct_or_initialize<Cpp<Class>>(std::forward<Args>(args)...);
+        }, is_new_style_constructor(), extra...);
+    }
+
+    template <typename Class, typename... Extra,
+              enable_if_t<Class::has_alias &&
+                          std::is_constructible<Cpp<Class>, Args...>::value, int> = 0>
+    static void execute(Class &cl, const Extra&... extra) {
+        cl.def("__init__", [](value_and_holder &v_h, Args... args) {
+            if (Py_TYPE(v_h.inst) == v_h.type->type)
+                v_h.value_ptr() = construct_or_initialize<Cpp<Class>>(std::forward<Args>(args)...);
+            else
+                v_h.value_ptr() = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+        }, is_new_style_constructor(), extra...);
+    }
+
+    template <typename Class, typename... Extra,
+              enable_if_t<Class::has_alias &&
+                          !std::is_constructible<Cpp<Class>, Args...>::value, int> = 0>
+    static void execute(Class &cl, const Extra&... extra) {
+        cl.def("__init__", [](value_and_holder &v_h, Args... args) {
+            v_h.value_ptr() = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+        }, is_new_style_constructor(), extra...);
+    }
+};
+
+// Implementing class for py::init_alias<...>()
+template <typename... Args> struct alias_constructor {
+    template <typename Class, typename... Extra,
+              enable_if_t<Class::has_alias && std::is_constructible<Alias<Class>, Args...>::value, int> = 0>
+    static void execute(Class &cl, const Extra&... extra) {
+        cl.def("__init__", [](value_and_holder &v_h, Args... args) {
+            v_h.value_ptr() = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+        }, is_new_style_constructor(), extra...);
+    }
+};
+
+// Implementation class for py::init(Func) and py::init(Func, AliasFunc)
+template <typename CFunc, typename AFunc = void_type (*)(),
+          typename = function_signature_t<CFunc>, typename = function_signature_t<AFunc>>
+struct factory;
+
+// Specialization for py::init(Func)
+template <typename Func, typename Return, typename... Args>
+struct factory<Func, void_type (*)(), Return(Args...)> {
+    remove_reference_t<Func> class_factory;
+
+    factory(Func &&f) : class_factory(std::forward<Func>(f)) { }
+
+    // The given class either has no alias or has no separate alias factory;
+    // this always constructs the class itself.  If the class is registered with an alias
+    // type and an alias instance is needed (i.e. because the final type is a Python class
+    // inheriting from the C++ type) the returned value needs to either already be an alias
+    // instance, or the alias needs to be constructible from a `Class &&` argument.
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) && {
+        #if defined(PYBIND11_CPP14)
+        cl.def("__init__", [func = std::move(class_factory)]
+        #else
+        auto &func = class_factory;
+        cl.def("__init__", [func]
+        #endif
+        (value_and_holder &v_h, Args... args) {
+            construct<Class>(v_h, func(std::forward<Args>(args)...),
+                             Py_TYPE(v_h.inst) != v_h.type->type);
+        }, is_new_style_constructor(), extra...);
+    }
+};
+
+// Specialization for py::init(Func, AliasFunc)
+template <typename CFunc, typename AFunc,
+          typename CReturn, typename... CArgs, typename AReturn, typename... AArgs>
+struct factory<CFunc, AFunc, CReturn(CArgs...), AReturn(AArgs...)> {
+    static_assert(sizeof...(CArgs) == sizeof...(AArgs),
+                  "pybind11::init(class_factory, alias_factory): class and alias factories "
+                  "must have identical argument signatures");
+    static_assert(all_of<std::is_same<CArgs, AArgs>...>::value,
+                  "pybind11::init(class_factory, alias_factory): class and alias factories "
+                  "must have identical argument signatures");
+
+    remove_reference_t<CFunc> class_factory;
+    remove_reference_t<AFunc> alias_factory;
+
+    factory(CFunc &&c, AFunc &&a)
+        : class_factory(std::forward<CFunc>(c)), alias_factory(std::forward<AFunc>(a)) { }
+
+    // The class factory is called when the `self` type passed to `__init__` is the direct
+    // class (i.e. not inherited), the alias factory when `self` is a Python-side subtype.
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra&... extra) && {
+        static_assert(Class::has_alias, "The two-argument version of `py::init()` can "
+                                        "only be used if the class has an alias");
+        #if defined(PYBIND11_CPP14)
+        cl.def("__init__", [class_func = std::move(class_factory), alias_func = std::move(alias_factory)]
+        #else
+        auto &class_func = class_factory;
+        auto &alias_func = alias_factory;
+        cl.def("__init__", [class_func, alias_func]
+        #endif
+        (value_and_holder &v_h, CArgs... args) {
+            if (Py_TYPE(v_h.inst) == v_h.type->type)
+                // If the instance type equals the registered type we don't have inheritance, so
+                // don't need the alias and can construct using the class function:
+                construct<Class>(v_h, class_func(std::forward<CArgs>(args)...), false);
+            else
+                construct<Class>(v_h, alias_func(std::forward<CArgs>(args)...), true);
+        }, is_new_style_constructor(), extra...);
+    }
+};
+
+/// Set just the C++ state. Same as `__init__`.
+template <typename Class, typename T>
+void setstate(value_and_holder &v_h, T &&result, bool need_alias) {
+    construct<Class>(v_h, std::forward<T>(result), need_alias);
+}
+
+/// Set both the C++ and Python states
+template <typename Class, typename T, typename O,
+          enable_if_t<std::is_convertible<O, handle>::value, int> = 0>
+void setstate(value_and_holder &v_h, std::pair<T, O> &&result, bool need_alias) {
+    construct<Class>(v_h, std::move(result.first), need_alias);
+    setattr((PyObject *) v_h.inst, "__dict__", result.second);
+}
+
+/// Implementation for py::pickle(GetState, SetState)
+template <typename Get, typename Set,
+          typename = function_signature_t<Get>, typename = function_signature_t<Set>>
+struct pickle_factory;
+
+template <typename Get, typename Set,
+          typename RetState, typename Self, typename NewInstance, typename ArgState>
+struct pickle_factory<Get, Set, RetState(Self), NewInstance(ArgState)> {
+    static_assert(std::is_same<intrinsic_t<RetState>, intrinsic_t<ArgState>>::value,
+                  "The type returned by `__getstate__` must be the same "
+                  "as the argument accepted by `__setstate__`");
+
+    remove_reference_t<Get> get;
+    remove_reference_t<Set> set;
+
+    pickle_factory(Get get, Set set)
+        : get(std::forward<Get>(get)), set(std::forward<Set>(set)) { }
+
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) && {
+        cl.def("__getstate__", std::move(get));
+
+#if defined(PYBIND11_CPP14)
+        cl.def("__setstate__", [func = std::move(set)]
+#else
+        auto &func = set;
+        cl.def("__setstate__", [func]
+#endif
+        (value_and_holder &v_h, ArgState state) {
+            setstate<Class>(v_h, func(std::forward<ArgState>(state)),
+                            Py_TYPE(v_h.inst) != v_h.type->type);
+        }, is_new_style_constructor(), extra...);
+    }
+};
+
+NAMESPACE_END(initimpl)
+NAMESPACE_END(detail)
+NAMESPACE_END(pybind11)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/internals.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/internals.h
new file mode 100644
index 000000000..e39f38695
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/internals.h
@@ -0,0 +1,249 @@
+/*
+    pybind11/detail/internals.h: Internal data structure and related functions
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "../pytypes.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+// Forward declarations
+inline PyTypeObject *make_static_property_type();
+inline PyTypeObject *make_default_metaclass();
+inline PyObject *make_object_base_type(PyTypeObject *metaclass);
+
+// Python loads modules by default with dlopen with the RTLD_LOCAL flag; under libc++ and possibly
+// other STLs, this means `typeid(A)` from one module won't equal `typeid(A)` from another module
+// even when `A` is the same, non-hidden-visibility type (e.g. from a common include).  Under
+// libstdc++, this doesn't happen: equality and the type_index hash are based on the type name,
+// which works.  If not under a known-good stl, provide our own name-based hash and equality
+// functions that use the type name.
+#if defined(__GLIBCXX__)
+inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) { return lhs == rhs; }
+using type_hash = std::hash<std::type_index>;
+using type_equal_to = std::equal_to<std::type_index>;
+#else
+inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) {
+    return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
+}
+
+struct type_hash {
+    size_t operator()(const std::type_index &t) const {
+        size_t hash = 5381;
+        const char *ptr = t.name();
+        while (auto c = static_cast<unsigned char>(*ptr++))
+            hash = (hash * 33) ^ c;
+        return hash;
+    }
+};
+
+struct type_equal_to {
+    bool operator()(const std::type_index &lhs, const std::type_index &rhs) const {
+        return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
+    }
+};
+#endif
+
+template <typename value_type>
+using type_map = std::unordered_map<std::type_index, value_type, type_hash, type_equal_to>;
+
+struct overload_hash {
+    inline size_t operator()(const std::pair<const PyObject *, const char *>& v) const {
+        size_t value = std::hash<const void *>()(v.first);
+        value ^= std::hash<const void *>()(v.second)  + 0x9e3779b9 + (value<<6) + (value>>2);
+        return value;
+    }
+};
+
+/// Internal data structure used to track registered instances and types.
+/// Whenever binary incompatible changes are made to this structure,
+/// `PYBIND11_INTERNALS_VERSION` must be incremented.
+struct internals {
+    type_map<type_info *> registered_types_cpp; // std::type_index -> pybind11's type information
+    std::unordered_map<PyTypeObject *, std::vector<type_info *>> registered_types_py; // PyTypeObject* -> base type_info(s)
+    std::unordered_multimap<const void *, instance*> registered_instances; // void * -> instance*
+    std::unordered_set<std::pair<const PyObject *, const char *>, overload_hash> inactive_overload_cache;
+    type_map<std::vector<bool (*)(PyObject *, void *&)>> direct_conversions;
+    std::unordered_map<const PyObject *, std::vector<PyObject *>> patients;
+    std::forward_list<void (*) (std::exception_ptr)> registered_exception_translators;
+    std::unordered_map<std::string, void *> shared_data; // Custom data to be shared across extensions
+    std::vector<PyObject *> loader_patient_stack; // Used by `loader_life_support`
+    std::forward_list<std::string> static_strings; // Stores the std::strings backing detail::c_str()
+    PyTypeObject *static_property_type;
+    PyTypeObject *default_metaclass;
+    PyObject *instance_base;
+#if defined(WITH_THREAD)
+    decltype(PyThread_create_key()) tstate = 0; // Usually an int but a long on Cygwin64 with Python 3.x
+    PyInterpreterState *istate = nullptr;
+#endif
+};
+
+/// Additional type information which does not fit into the PyTypeObject.
+/// Changes to this struct also require bumping `PYBIND11_INTERNALS_VERSION`.
+struct type_info {
+    PyTypeObject *type;
+    const std::type_info *cpptype;
+    size_t type_size, holder_size_in_ptrs;
+    void *(*operator_new)(size_t);
+    void (*init_instance)(instance *, const void *);
+    void (*dealloc)(value_and_holder &v_h);
+    std::vector<PyObject *(*)(PyObject *, PyTypeObject *)> implicit_conversions;
+    std::vector<std::pair<const std::type_info *, void *(*)(void *)>> implicit_casts;
+    std::vector<bool (*)(PyObject *, void *&)> *direct_conversions;
+    buffer_info *(*get_buffer)(PyObject *, void *) = nullptr;
+    void *get_buffer_data = nullptr;
+    void *(*module_local_load)(PyObject *, const type_info *) = nullptr;
+    /* A simple type never occurs as a (direct or indirect) parent
+     * of a class that makes use of multiple inheritance */
+    bool simple_type : 1;
+    /* True if there is no multiple inheritance in this type's inheritance tree */
+    bool simple_ancestors : 1;
+    /* for base vs derived holder_type checks */
+    bool default_holder : 1;
+    /* true if this is a type registered with py::module_local */
+    bool module_local : 1;
+};
+
+/// Tracks the `internals` and `type_info` ABI version independent of the main library version
+#define PYBIND11_INTERNALS_VERSION 1
+
+#if defined(WITH_THREAD)
+#  define PYBIND11_INTERNALS_KIND ""
+#else
+#  define PYBIND11_INTERNALS_KIND "_without_thread"
+#endif
+
+#define PYBIND11_INTERNALS_ID "__pybind11_internals_v" \
+    PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION) PYBIND11_INTERNALS_KIND "__"
+
+#define PYBIND11_MODULE_LOCAL_ID "__pybind11_module_local_v" \
+    PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION) PYBIND11_INTERNALS_KIND "__"
+
+/// Each module locally stores a pointer to the `internals` data. The data
+/// itself is shared among modules with the same `PYBIND11_INTERNALS_ID`.
+inline internals **&get_internals_pp() {
+    static internals **internals_pp = nullptr;
+    return internals_pp;
+}
+
+/// Return a reference to the current `internals` data
+PYBIND11_NOINLINE inline internals &get_internals() {
+    auto **&internals_pp = get_internals_pp();
+    if (internals_pp && *internals_pp)
+        return **internals_pp;
+
+    constexpr auto *id = PYBIND11_INTERNALS_ID;
+    auto builtins = handle(PyEval_GetBuiltins());
+    if (builtins.contains(id) && isinstance<capsule>(builtins[id])) {
+        internals_pp = static_cast<internals **>(capsule(builtins[id]));
+
+        // We loaded builtins through python's builtins, which means that our `error_already_set`
+        // and `builtin_exception` may be different local classes than the ones set up in the
+        // initial exception translator, below, so add another for our local exception classes.
+        //
+        // libstdc++ doesn't require this (types there are identified only by name)
+#if !defined(__GLIBCXX__)
+        (*internals_pp)->registered_exception_translators.push_front(
+            [](std::exception_ptr p) -> void {
+                try {
+                    if (p) std::rethrow_exception(p);
+                } catch (error_already_set &e)       { e.restore();   return;
+                } catch (const builtin_exception &e) { e.set_error(); return;
+                }
+            }
+        );
+#endif
+    } else {
+        if (!internals_pp) internals_pp = new internals*();
+        auto *&internals_ptr = *internals_pp;
+        internals_ptr = new internals();
+#if defined(WITH_THREAD)
+        PyEval_InitThreads();
+        PyThreadState *tstate = PyThreadState_Get();
+        internals_ptr->tstate = PyThread_create_key();
+        PyThread_set_key_value(internals_ptr->tstate, tstate);
+        internals_ptr->istate = tstate->interp;
+#endif
+        builtins[id] = capsule(internals_pp);
+        internals_ptr->registered_exception_translators.push_front(
+            [](std::exception_ptr p) -> void {
+                try {
+                    if (p) std::rethrow_exception(p);
+                } catch (error_already_set &e)           { e.restore();                                    return;
+                } catch (const builtin_exception &e)     { e.set_error();                                  return;
+                } catch (const std::bad_alloc &e)        { PyErr_SetString(PyExc_MemoryError,   e.what()); return;
+                } catch (const std::domain_error &e)     { PyErr_SetString(PyExc_ValueError,    e.what()); return;
+                } catch (const std::invalid_argument &e) { PyErr_SetString(PyExc_ValueError,    e.what()); return;
+                } catch (const std::length_error &e)     { PyErr_SetString(PyExc_ValueError,    e.what()); return;
+                } catch (const std::out_of_range &e)     { PyErr_SetString(PyExc_IndexError,    e.what()); return;
+                } catch (const std::range_error &e)      { PyErr_SetString(PyExc_ValueError,    e.what()); return;
+                } catch (const std::exception &e)        { PyErr_SetString(PyExc_RuntimeError,  e.what()); return;
+                } catch (...) {
+                    PyErr_SetString(PyExc_RuntimeError, "Caught an unknown exception!");
+                    return;
+                }
+            }
+        );
+        internals_ptr->static_property_type = make_static_property_type();
+        internals_ptr->default_metaclass = make_default_metaclass();
+        internals_ptr->instance_base = make_object_base_type(internals_ptr->default_metaclass);
+    }
+    return **internals_pp;
+}
+
+/// Works like `internals.registered_types_cpp`, but for module-local registered types:
+inline type_map<type_info *> &registered_local_types_cpp() {
+    static type_map<type_info *> locals{};
+    return locals;
+}
+
+/// Constructs a std::string with the given arguments, stores it in `internals`, and returns its
+/// `c_str()`.  Such strings objects have a long storage duration -- the internal strings are only
+/// cleared when the program exits or after interpreter shutdown (when embedding), and so are
+/// suitable for c-style strings needed by Python internals (such as PyTypeObject's tp_name).
+template <typename... Args>
+const char *c_str(Args &&...args) {
+    auto &strings = get_internals().static_strings;
+    strings.emplace_front(std::forward<Args>(args)...);
+    return strings.front().c_str();
+}
+
+NAMESPACE_END(detail)
+
+/// Returns a named pointer that is shared among all extension modules (using the same
+/// pybind11 version) running in the current interpreter. Names starting with underscores
+/// are reserved for internal usage. Returns `nullptr` if no matching entry was found.
+inline PYBIND11_NOINLINE void *get_shared_data(const std::string &name) {
+    auto &internals = detail::get_internals();
+    auto it = internals.shared_data.find(name);
+    return it != internals.shared_data.end() ? it->second : nullptr;
+}
+
+/// Set the shared data that can be later recovered by `get_shared_data()`.
+inline PYBIND11_NOINLINE void *set_shared_data(const std::string &name, void *data) {
+    detail::get_internals().shared_data[name] = data;
+    return data;
+}
+
+/// Returns a typed reference to a shared data entry (by using `get_shared_data()`) if
+/// such entry exists. Otherwise, a new object of default-constructible type `T` is
+/// added to the shared data under the given name and a reference to it is returned.
+template<typename T>
+T &get_or_create_shared_data(const std::string &name) {
+    auto &internals = detail::get_internals();
+    auto it = internals.shared_data.find(name);
+    T *ptr = (T *) (it != internals.shared_data.end() ? it->second : nullptr);
+    if (!ptr) {
+        ptr = new T();
+        internals.shared_data[name] = ptr;
+    }
+    return *ptr;
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/detail/typeid.h b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/typeid.h
new file mode 100644
index 000000000..6f36aab75
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/detail/typeid.h
@@ -0,0 +1,53 @@
+/*
+    pybind11/detail/typeid.h: Compiler-independent access to type identifiers
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+
+#if defined(__GNUG__)
+#include <cxxabi.h>
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+/// Erase all occurrences of a substring
+inline void erase_all(std::string &string, const std::string &search) {
+    for (size_t pos = 0;;) {
+        pos = string.find(search, pos);
+        if (pos == std::string::npos) break;
+        string.erase(pos, search.length());
+    }
+}
+
+PYBIND11_NOINLINE inline void clean_type_id(std::string &name) {
+#if defined(__GNUG__)
+    int status = 0;
+    std::unique_ptr<char, void (*)(void *)> res {
+        abi::__cxa_demangle(name.c_str(), nullptr, nullptr, &status), std::free };
+    if (status == 0)
+        name = res.get();
+#else
+    detail::erase_all(name, "class ");
+    detail::erase_all(name, "struct ");
+    detail::erase_all(name, "enum ");
+#endif
+    detail::erase_all(name, "pybind11::");
+}
+NAMESPACE_END(detail)
+
+/// Return a string representation of a C++ type
+template <typename T> static std::string type_id() {
+    std::string name(typeid(T).name());
+    detail::clean_type_id(name);
+    return name;
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/eigen.h b/ml/dlib/dlib/external/pybind11/include/pybind11/eigen.h
new file mode 100644
index 000000000..693a484dc
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/eigen.h
@@ -0,0 +1,612 @@
+/*
+    pybind11/eigen.h: Transparent conversion for dense and sparse Eigen matrices
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "numpy.h"
+
+#if defined(__INTEL_COMPILER)
+#  pragma warning(disable: 1682) // implicit conversion of a 64-bit integral type to a smaller integral type (potential portability problem)
+#elif defined(__GNUG__) || defined(__clang__)
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wconversion"
+#  pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+#  if __GNUC__ >= 7
+#    pragma GCC diagnostic ignored "-Wint-in-bool-context"
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  pragma warning(push)
+#  pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#  pragma warning(disable: 4996) // warning C4996: std::unary_negate is deprecated in C++17
+#endif
+
+#include <Eigen/Core>
+#include <Eigen/SparseCore>
+
+// Eigen prior to 3.2.7 doesn't have proper move constructors--but worse, some classes get implicit
+// move constructors that break things.  We could detect this an explicitly copy, but an extra copy
+// of matrices seems highly undesirable.
+static_assert(EIGEN_VERSION_AT_LEAST(3,2,7), "Eigen support in pybind11 requires Eigen >= 3.2.7");
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+// Provide a convenience alias for easier pass-by-ref usage with fully dynamic strides:
+using EigenDStride = Eigen::Stride<Eigen::Dynamic, Eigen::Dynamic>;
+template <typename MatrixType> using EigenDRef = Eigen::Ref<MatrixType, 0, EigenDStride>;
+template <typename MatrixType> using EigenDMap = Eigen::Map<MatrixType, 0, EigenDStride>;
+
+NAMESPACE_BEGIN(detail)
+
+#if EIGEN_VERSION_AT_LEAST(3,3,0)
+using EigenIndex = Eigen::Index;
+#else
+using EigenIndex = EIGEN_DEFAULT_DENSE_INDEX_TYPE;
+#endif
+
+// Matches Eigen::Map, Eigen::Ref, blocks, etc:
+template <typename T> using is_eigen_dense_map = all_of<is_template_base_of<Eigen::DenseBase, T>, std::is_base_of<Eigen::MapBase<T, Eigen::ReadOnlyAccessors>, T>>;
+template <typename T> using is_eigen_mutable_map = std::is_base_of<Eigen::MapBase<T, Eigen::WriteAccessors>, T>;
+template <typename T> using is_eigen_dense_plain = all_of<negation<is_eigen_dense_map<T>>, is_template_base_of<Eigen::PlainObjectBase, T>>;
+template <typename T> using is_eigen_sparse = is_template_base_of<Eigen::SparseMatrixBase, T>;
+// Test for objects inheriting from EigenBase<Derived> that aren't captured by the above.  This
+// basically covers anything that can be assigned to a dense matrix but that don't have a typical
+// matrix data layout that can be copied from their .data().  For example, DiagonalMatrix and
+// SelfAdjointView fall into this category.
+template <typename T> using is_eigen_other = all_of<
+    is_template_base_of<Eigen::EigenBase, T>,
+    negation<any_of<is_eigen_dense_map<T>, is_eigen_dense_plain<T>, is_eigen_sparse<T>>>
+>;
+
+// Captures numpy/eigen conformability status (returned by EigenProps::conformable()):
+template <bool EigenRowMajor> struct EigenConformable {
+    bool conformable = false;
+    EigenIndex rows = 0, cols = 0;
+    EigenDStride stride{0, 0};      // Only valid if negativestrides is false!
+    bool negativestrides = false;   // If true, do not use stride!
+
+    EigenConformable(bool fits = false) : conformable{fits} {}
+    // Matrix type:
+    EigenConformable(EigenIndex r, EigenIndex c,
+            EigenIndex rstride, EigenIndex cstride) :
+        conformable{true}, rows{r}, cols{c} {
+        // TODO: when Eigen bug #747 is fixed, remove the tests for non-negativity. http://eigen.tuxfamily.org/bz/show_bug.cgi?id=747
+        if (rstride < 0 || cstride < 0) {
+            negativestrides = true;
+        } else {
+            stride = {EigenRowMajor ? rstride : cstride /* outer stride */,
+                      EigenRowMajor ? cstride : rstride /* inner stride */ };
+        }
+    }
+    // Vector type:
+    EigenConformable(EigenIndex r, EigenIndex c, EigenIndex stride)
+        : EigenConformable(r, c, r == 1 ? c*stride : stride, c == 1 ? r : r*stride) {}
+
+    template <typename props> bool stride_compatible() const {
+        // To have compatible strides, we need (on both dimensions) one of fully dynamic strides,
+        // matching strides, or a dimension size of 1 (in which case the stride value is irrelevant)
+        return
+            !negativestrides &&
+            (props::inner_stride == Eigen::Dynamic || props::inner_stride == stride.inner() ||
+                (EigenRowMajor ? cols : rows) == 1) &&
+            (props::outer_stride == Eigen::Dynamic || props::outer_stride == stride.outer() ||
+                (EigenRowMajor ? rows : cols) == 1);
+    }
+    operator bool() const { return conformable; }
+};
+
+template <typename Type> struct eigen_extract_stride { using type = Type; };
+template <typename PlainObjectType, int MapOptions, typename StrideType>
+struct eigen_extract_stride<Eigen::Map<PlainObjectType, MapOptions, StrideType>> { using type = StrideType; };
+template <typename PlainObjectType, int Options, typename StrideType>
+struct eigen_extract_stride<Eigen::Ref<PlainObjectType, Options, StrideType>> { using type = StrideType; };
+
+// Helper struct for extracting information from an Eigen type
+template <typename Type_> struct EigenProps {
+    using Type = Type_;
+    using Scalar = typename Type::Scalar;
+    using StrideType = typename eigen_extract_stride<Type>::type;
+    static constexpr EigenIndex
+        rows = Type::RowsAtCompileTime,
+        cols = Type::ColsAtCompileTime,
+        size = Type::SizeAtCompileTime;
+    static constexpr bool
+        row_major = Type::IsRowMajor,
+        vector = Type::IsVectorAtCompileTime, // At least one dimension has fixed size 1
+        fixed_rows = rows != Eigen::Dynamic,
+        fixed_cols = cols != Eigen::Dynamic,
+        fixed = size != Eigen::Dynamic, // Fully-fixed size
+        dynamic = !fixed_rows && !fixed_cols; // Fully-dynamic size
+
+    template <EigenIndex i, EigenIndex ifzero> using if_zero = std::integral_constant<EigenIndex, i == 0 ? ifzero : i>;
+    static constexpr EigenIndex inner_stride = if_zero<StrideType::InnerStrideAtCompileTime, 1>::value,
+                                outer_stride = if_zero<StrideType::OuterStrideAtCompileTime,
+                                                       vector ? size : row_major ? cols : rows>::value;
+    static constexpr bool dynamic_stride = inner_stride == Eigen::Dynamic && outer_stride == Eigen::Dynamic;
+    static constexpr bool requires_row_major = !dynamic_stride && !vector && (row_major ? inner_stride : outer_stride) == 1;
+    static constexpr bool requires_col_major = !dynamic_stride && !vector && (row_major ? outer_stride : inner_stride) == 1;
+
+    // Takes an input array and determines whether we can make it fit into the Eigen type.  If
+    // the array is a vector, we attempt to fit it into either an Eigen 1xN or Nx1 vector
+    // (preferring the latter if it will fit in either, i.e. for a fully dynamic matrix type).
+    static EigenConformable<row_major> conformable(const array &a) {
+        const auto dims = a.ndim();
+        if (dims < 1 || dims > 2)
+            return false;
+
+        if (dims == 2) { // Matrix type: require exact match (or dynamic)
+
+            EigenIndex
+                np_rows = a.shape(0),
+                np_cols = a.shape(1),
+                np_rstride = a.strides(0) / static_cast<ssize_t>(sizeof(Scalar)),
+                np_cstride = a.strides(1) / static_cast<ssize_t>(sizeof(Scalar));
+            if ((fixed_rows && np_rows != rows) || (fixed_cols && np_cols != cols))
+                return false;
+
+            return {np_rows, np_cols, np_rstride, np_cstride};
+        }
+
+        // Otherwise we're storing an n-vector.  Only one of the strides will be used, but whichever
+        // is used, we want the (single) numpy stride value.
+        const EigenIndex n = a.shape(0),
+              stride = a.strides(0) / static_cast<ssize_t>(sizeof(Scalar));
+
+        if (vector) { // Eigen type is a compile-time vector
+            if (fixed && size != n)
+                return false; // Vector size mismatch
+            return {rows == 1 ? 1 : n, cols == 1 ? 1 : n, stride};
+        }
+        else if (fixed) {
+            // The type has a fixed size, but is not a vector: abort
+            return false;
+        }
+        else if (fixed_cols) {
+            // Since this isn't a vector, cols must be != 1.  We allow this only if it exactly
+            // equals the number of elements (rows is Dynamic, and so 1 row is allowed).
+            if (cols != n) return false;
+            return {1, n, stride};
+        }
+        else {
+            // Otherwise it's either fully dynamic, or column dynamic; both become a column vector
+            if (fixed_rows && rows != n) return false;
+            return {n, 1, stride};
+        }
+    }
+
+    static PYBIND11_DESCR descriptor() {
+        constexpr bool show_writeable = is_eigen_dense_map<Type>::value && is_eigen_mutable_map<Type>::value;
+        constexpr bool show_order = is_eigen_dense_map<Type>::value;
+        constexpr bool show_c_contiguous = show_order && requires_row_major;
+        constexpr bool show_f_contiguous = !show_c_contiguous && show_order && requires_col_major;
+
+        return type_descr(_("numpy.ndarray[") + npy_format_descriptor<Scalar>::name() +
+            _("[")  + _<fixed_rows>(_<(size_t) rows>(), _("m")) +
+            _(", ") + _<fixed_cols>(_<(size_t) cols>(), _("n")) +
+            _("]") +
+            // For a reference type (e.g. Ref<MatrixXd>) we have other constraints that might need to be
+            // satisfied: writeable=True (for a mutable reference), and, depending on the map's stride
+            // options, possibly f_contiguous or c_contiguous.  We include them in the descriptor output
+            // to provide some hint as to why a TypeError is occurring (otherwise it can be confusing to
+            // see that a function accepts a 'numpy.ndarray[float64[3,2]]' and an error message that you
+            // *gave* a numpy.ndarray of the right type and dimensions.
+            _<show_writeable>(", flags.writeable", "") +
+            _<show_c_contiguous>(", flags.c_contiguous", "") +
+            _<show_f_contiguous>(", flags.f_contiguous", "") +
+            _("]")
+        );
+    }
+};
+
+// Casts an Eigen type to numpy array.  If given a base, the numpy array references the src data,
+// otherwise it'll make a copy.  writeable lets you turn off the writeable flag for the array.
+template <typename props> handle eigen_array_cast(typename props::Type const &src, handle base = handle(), bool writeable = true) {
+    constexpr ssize_t elem_size = sizeof(typename props::Scalar);
+    array a;
+    if (props::vector)
+        a = array({ src.size() }, { elem_size * src.innerStride() }, src.data(), base);
+    else
+        a = array({ src.rows(), src.cols() }, { elem_size * src.rowStride(), elem_size * src.colStride() },
+                  src.data(), base);
+
+    if (!writeable)
+        array_proxy(a.ptr())->flags &= ~detail::npy_api::NPY_ARRAY_WRITEABLE_;
+
+    return a.release();
+}
+
+// Takes an lvalue ref to some Eigen type and a (python) base object, creating a numpy array that
+// reference the Eigen object's data with `base` as the python-registered base class (if omitted,
+// the base will be set to None, and lifetime management is up to the caller).  The numpy array is
+// non-writeable if the given type is const.
+template <typename props, typename Type>
+handle eigen_ref_array(Type &src, handle parent = none()) {
+    // none here is to get past array's should-we-copy detection, which currently always
+    // copies when there is no base.  Setting the base to None should be harmless.
+    return eigen_array_cast<props>(src, parent, !std::is_const<Type>::value);
+}
+
+// Takes a pointer to some dense, plain Eigen type, builds a capsule around it, then returns a numpy
+// array that references the encapsulated data with a python-side reference to the capsule to tie
+// its destruction to that of any dependent python objects.  Const-ness is determined by whether or
+// not the Type of the pointer given is const.
+template <typename props, typename Type, typename = enable_if_t<is_eigen_dense_plain<Type>::value>>
+handle eigen_encapsulate(Type *src) {
+    capsule base(src, [](void *o) { delete static_cast<Type *>(o); });
+    return eigen_ref_array<props>(*src, base);
+}
+
+// Type caster for regular, dense matrix types (e.g. MatrixXd), but not maps/refs/etc. of dense
+// types.
+template<typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_dense_plain<Type>::value>> {
+    using Scalar = typename Type::Scalar;
+    using props = EigenProps<Type>;
+
+    bool load(handle src, bool convert) {
+        // If we're in no-convert mode, only load if given an array of the correct type
+        if (!convert && !isinstance<array_t<Scalar>>(src))
+            return false;
+
+        // Coerce into an array, but don't do type conversion yet; the copy below handles it.
+        auto buf = array::ensure(src);
+
+        if (!buf)
+            return false;
+
+        auto dims = buf.ndim();
+        if (dims < 1 || dims > 2)
+            return false;
+
+        auto fits = props::conformable(buf);
+        if (!fits)
+            return false;
+
+        // Allocate the new type, then build a numpy reference into it
+        value = Type(fits.rows, fits.cols);
+        auto ref = reinterpret_steal<array>(eigen_ref_array<props>(value));
+        if (dims == 1) ref = ref.squeeze();
+        else if (ref.ndim() == 1) buf = buf.squeeze();
+
+        int result = detail::npy_api::get().PyArray_CopyInto_(ref.ptr(), buf.ptr());
+
+        if (result < 0) { // Copy failed!
+            PyErr_Clear();
+            return false;
+        }
+
+        return true;
+    }
+
+private:
+
+    // Cast implementation
+    template <typename CType>
+    static handle cast_impl(CType *src, return_value_policy policy, handle parent) {
+        switch (policy) {
+            case return_value_policy::take_ownership:
+            case return_value_policy::automatic:
+                return eigen_encapsulate<props>(src);
+            case return_value_policy::move:
+                return eigen_encapsulate<props>(new CType(std::move(*src)));
+            case return_value_policy::copy:
+                return eigen_array_cast<props>(*src);
+            case return_value_policy::reference:
+            case return_value_policy::automatic_reference:
+                return eigen_ref_array<props>(*src);
+            case return_value_policy::reference_internal:
+                return eigen_ref_array<props>(*src, parent);
+            default:
+                throw cast_error("unhandled return_value_policy: should not happen!");
+        };
+    }
+
+public:
+
+    // Normal returned non-reference, non-const value:
+    static handle cast(Type &&src, return_value_policy /* policy */, handle parent) {
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+    // If you return a non-reference const, we mark the numpy array readonly:
+    static handle cast(const Type &&src, return_value_policy /* policy */, handle parent) {
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+    // lvalue reference return; default (automatic) becomes copy
+    static handle cast(Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic || policy == return_value_policy::automatic_reference)
+            policy = return_value_policy::copy;
+        return cast_impl(&src, policy, parent);
+    }
+    // const lvalue reference return; default (automatic) becomes copy
+    static handle cast(const Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic || policy == return_value_policy::automatic_reference)
+            policy = return_value_policy::copy;
+        return cast(&src, policy, parent);
+    }
+    // non-const pointer return
+    static handle cast(Type *src, return_value_policy policy, handle parent) {
+        return cast_impl(src, policy, parent);
+    }
+    // const pointer return
+    static handle cast(const Type *src, return_value_policy policy, handle parent) {
+        return cast_impl(src, policy, parent);
+    }
+
+    static PYBIND11_DESCR name() { return props::descriptor(); }
+
+    operator Type*() { return &value; }
+    operator Type&() { return value; }
+    operator Type&&() && { return std::move(value); }
+    template <typename T> using cast_op_type = movable_cast_op_type<T>;
+
+private:
+    Type value;
+};
+
+// Eigen Ref/Map classes have slightly different policy requirements, meaning we don't want to force
+// `move` when a Ref/Map rvalue is returned; we treat Ref<> sort of like a pointer (we care about
+// the underlying data, not the outer shell).
+template <typename Return>
+struct return_value_policy_override<Return, enable_if_t<is_eigen_dense_map<Return>::value>> {
+    static return_value_policy policy(return_value_policy p) { return p; }
+};
+
+// Base class for casting reference/map/block/etc. objects back to python.
+template <typename MapType> struct eigen_map_caster {
+private:
+    using props = EigenProps<MapType>;
+
+public:
+
+    // Directly referencing a ref/map's data is a bit dangerous (whatever the map/ref points to has
+    // to stay around), but we'll allow it under the assumption that you know what you're doing (and
+    // have an appropriate keep_alive in place).  We return a numpy array pointing directly at the
+    // ref's data (The numpy array ends up read-only if the ref was to a const matrix type.) Note
+    // that this means you need to ensure you don't destroy the object in some other way (e.g. with
+    // an appropriate keep_alive, or with a reference to a statically allocated matrix).
+    static handle cast(const MapType &src, return_value_policy policy, handle parent) {
+        switch (policy) {
+            case return_value_policy::copy:
+                return eigen_array_cast<props>(src);
+            case return_value_policy::reference_internal:
+                return eigen_array_cast<props>(src, parent, is_eigen_mutable_map<MapType>::value);
+            case return_value_policy::reference:
+            case return_value_policy::automatic:
+            case return_value_policy::automatic_reference:
+                return eigen_array_cast<props>(src, none(), is_eigen_mutable_map<MapType>::value);
+            default:
+                // move, take_ownership don't make any sense for a ref/map:
+                pybind11_fail("Invalid return_value_policy for Eigen Map/Ref/Block type");
+        }
+    }
+
+    static PYBIND11_DESCR name() { return props::descriptor(); }
+
+    // Explicitly delete these: support python -> C++ conversion on these (i.e. these can be return
+    // types but not bound arguments).  We still provide them (with an explicitly delete) so that
+    // you end up here if you try anyway.
+    bool load(handle, bool) = delete;
+    operator MapType() = delete;
+    template <typename> using cast_op_type = MapType;
+};
+
+// We can return any map-like object (but can only load Refs, specialized next):
+template <typename Type> struct type_caster<Type, enable_if_t<is_eigen_dense_map<Type>::value>>
+    : eigen_map_caster<Type> {};
+
+// Loader for Ref<...> arguments.  See the documentation for info on how to make this work without
+// copying (it requires some extra effort in many cases).
+template <typename PlainObjectType, typename StrideType>
+struct type_caster<
+    Eigen::Ref<PlainObjectType, 0, StrideType>,
+    enable_if_t<is_eigen_dense_map<Eigen::Ref<PlainObjectType, 0, StrideType>>::value>
+> : public eigen_map_caster<Eigen::Ref<PlainObjectType, 0, StrideType>> {
+private:
+    using Type = Eigen::Ref<PlainObjectType, 0, StrideType>;
+    using props = EigenProps<Type>;
+    using Scalar = typename props::Scalar;
+    using MapType = Eigen::Map<PlainObjectType, 0, StrideType>;
+    using Array = array_t<Scalar, array::forcecast |
+                ((props::row_major ? props::inner_stride : props::outer_stride) == 1 ? array::c_style :
+                 (props::row_major ? props::outer_stride : props::inner_stride) == 1 ? array::f_style : 0)>;
+    static constexpr bool need_writeable = is_eigen_mutable_map<Type>::value;
+    // Delay construction (these have no default constructor)
+    std::unique_ptr<MapType> map;
+    std::unique_ptr<Type> ref;
+    // Our array.  When possible, this is just a numpy array pointing to the source data, but
+    // sometimes we can't avoid copying (e.g. input is not a numpy array at all, has an incompatible
+    // layout, or is an array of a type that needs to be converted).  Using a numpy temporary
+    // (rather than an Eigen temporary) saves an extra copy when we need both type conversion and
+    // storage order conversion.  (Note that we refuse to use this temporary copy when loading an
+    // argument for a Ref<M> with M non-const, i.e. a read-write reference).
+    Array copy_or_ref;
+public:
+    bool load(handle src, bool convert) {
+        // First check whether what we have is already an array of the right type.  If not, we can't
+        // avoid a copy (because the copy is also going to do type conversion).
+        bool need_copy = !isinstance<Array>(src);
+
+        EigenConformable<props::row_major> fits;
+        if (!need_copy) {
+            // We don't need a converting copy, but we also need to check whether the strides are
+            // compatible with the Ref's stride requirements
+            Array aref = reinterpret_borrow<Array>(src);
+
+            if (aref && (!need_writeable || aref.writeable())) {
+                fits = props::conformable(aref);
+                if (!fits) return false; // Incompatible dimensions
+                if (!fits.template stride_compatible<props>())
+                    need_copy = true;
+                else
+                    copy_or_ref = std::move(aref);
+            }
+            else {
+                need_copy = true;
+            }
+        }
+
+        if (need_copy) {
+            // We need to copy: If we need a mutable reference, or we're not supposed to convert
+            // (either because we're in the no-convert overload pass, or because we're explicitly
+            // instructed not to copy (via `py::arg().noconvert()`) we have to fail loading.
+            if (!convert || need_writeable) return false;
+
+            Array copy = Array::ensure(src);
+            if (!copy) return false;
+            fits = props::conformable(copy);
+            if (!fits || !fits.template stride_compatible<props>())
+                return false;
+            copy_or_ref = std::move(copy);
+            loader_life_support::add_patient(copy_or_ref);
+        }
+
+        ref.reset();
+        map.reset(new MapType(data(copy_or_ref), fits.rows, fits.cols, make_stride(fits.stride.outer(), fits.stride.inner())));
+        ref.reset(new Type(*map));
+
+        return true;
+    }
+
+    operator Type*() { return ref.get(); }
+    operator Type&() { return *ref; }
+    template <typename _T> using cast_op_type = pybind11::detail::cast_op_type<_T>;
+
+private:
+    template <typename T = Type, enable_if_t<is_eigen_mutable_map<T>::value, int> = 0>
+    Scalar *data(Array &a) { return a.mutable_data(); }
+
+    template <typename T = Type, enable_if_t<!is_eigen_mutable_map<T>::value, int> = 0>
+    const Scalar *data(Array &a) { return a.data(); }
+
+    // Attempt to figure out a constructor of `Stride` that will work.
+    // If both strides are fixed, use a default constructor:
+    template <typename S> using stride_ctor_default = bool_constant<
+        S::InnerStrideAtCompileTime != Eigen::Dynamic && S::OuterStrideAtCompileTime != Eigen::Dynamic &&
+        std::is_default_constructible<S>::value>;
+    // Otherwise, if there is a two-index constructor, assume it is (outer,inner) like
+    // Eigen::Stride, and use it:
+    template <typename S> using stride_ctor_dual = bool_constant<
+        !stride_ctor_default<S>::value && std::is_constructible<S, EigenIndex, EigenIndex>::value>;
+    // Otherwise, if there is a one-index constructor, and just one of the strides is dynamic, use
+    // it (passing whichever stride is dynamic).
+    template <typename S> using stride_ctor_outer = bool_constant<
+        !any_of<stride_ctor_default<S>, stride_ctor_dual<S>>::value &&
+        S::OuterStrideAtCompileTime == Eigen::Dynamic && S::InnerStrideAtCompileTime != Eigen::Dynamic &&
+        std::is_constructible<S, EigenIndex>::value>;
+    template <typename S> using stride_ctor_inner = bool_constant<
+        !any_of<stride_ctor_default<S>, stride_ctor_dual<S>>::value &&
+        S::InnerStrideAtCompileTime == Eigen::Dynamic && S::OuterStrideAtCompileTime != Eigen::Dynamic &&
+        std::is_constructible<S, EigenIndex>::value>;
+
+    template <typename S = StrideType, enable_if_t<stride_ctor_default<S>::value, int> = 0>
+    static S make_stride(EigenIndex, EigenIndex) { return S(); }
+    template <typename S = StrideType, enable_if_t<stride_ctor_dual<S>::value, int> = 0>
+    static S make_stride(EigenIndex outer, EigenIndex inner) { return S(outer, inner); }
+    template <typename S = StrideType, enable_if_t<stride_ctor_outer<S>::value, int> = 0>
+    static S make_stride(EigenIndex outer, EigenIndex) { return S(outer); }
+    template <typename S = StrideType, enable_if_t<stride_ctor_inner<S>::value, int> = 0>
+    static S make_stride(EigenIndex, EigenIndex inner) { return S(inner); }
+
+};
+
+// type_caster for special matrix types (e.g. DiagonalMatrix), which are EigenBase, but not
+// EigenDense (i.e. they don't have a data(), at least not with the usual matrix layout).
+// load() is not supported, but we can cast them into the python domain by first copying to a
+// regular Eigen::Matrix, then casting that.
+template <typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_other<Type>::value>> {
+protected:
+    using Matrix = Eigen::Matrix<typename Type::Scalar, Type::RowsAtCompileTime, Type::ColsAtCompileTime>;
+    using props = EigenProps<Matrix>;
+public:
+    static handle cast(const Type &src, return_value_policy /* policy */, handle /* parent */) {
+        handle h = eigen_encapsulate<props>(new Matrix(src));
+        return h;
+    }
+    static handle cast(const Type *src, return_value_policy policy, handle parent) { return cast(*src, policy, parent); }
+
+    static PYBIND11_DESCR name() { return props::descriptor(); }
+
+    // Explicitly delete these: support python -> C++ conversion on these (i.e. these can be return
+    // types but not bound arguments).  We still provide them (with an explicitly delete) so that
+    // you end up here if you try anyway.
+    bool load(handle, bool) = delete;
+    operator Type() = delete;
+    template <typename> using cast_op_type = Type;
+};
+
+template<typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_sparse<Type>::value>> {
+    typedef typename Type::Scalar Scalar;
+    typedef remove_reference_t<decltype(*std::declval<Type>().outerIndexPtr())> StorageIndex;
+    typedef typename Type::Index Index;
+    static constexpr bool rowMajor = Type::IsRowMajor;
+
+    bool load(handle src, bool) {
+        if (!src)
+            return false;
+
+        auto obj = reinterpret_borrow<object>(src);
+        object sparse_module = module::import("scipy.sparse");
+        object matrix_type = sparse_module.attr(
+            rowMajor ? "csr_matrix" : "csc_matrix");
+
+        if (!obj.get_type().is(matrix_type)) {
+            try {
+                obj = matrix_type(obj);
+            } catch (const error_already_set &) {
+                return false;
+            }
+        }
+
+        auto values = array_t<Scalar>((object) obj.attr("data"));
+        auto innerIndices = array_t<StorageIndex>((object) obj.attr("indices"));
+        auto outerIndices = array_t<StorageIndex>((object) obj.attr("indptr"));
+        auto shape = pybind11::tuple((pybind11::object) obj.attr("shape"));
+        auto nnz = obj.attr("nnz").cast<Index>();
+
+        if (!values || !innerIndices || !outerIndices)
+            return false;
+
+        value = Eigen::MappedSparseMatrix<Scalar, Type::Flags, StorageIndex>(
+            shape[0].cast<Index>(), shape[1].cast<Index>(), nnz,
+            outerIndices.mutable_data(), innerIndices.mutable_data(), values.mutable_data());
+
+        return true;
+    }
+
+    static handle cast(const Type &src, return_value_policy /* policy */, handle /* parent */) {
+        const_cast<Type&>(src).makeCompressed();
+
+        object matrix_type = module::import("scipy.sparse").attr(
+            rowMajor ? "csr_matrix" : "csc_matrix");
+
+        array data(src.nonZeros(), src.valuePtr());
+        array outerIndices((rowMajor ? src.rows() : src.cols()) + 1, src.outerIndexPtr());
+        array innerIndices(src.nonZeros(), src.innerIndexPtr());
+
+        return matrix_type(
+            std::make_tuple(data, innerIndices, outerIndices),
+            std::make_pair(src.rows(), src.cols())
+        ).release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type, _<(Type::IsRowMajor) != 0>("scipy.sparse.csr_matrix[", "scipy.sparse.csc_matrix[")
+            + npy_format_descriptor<Scalar>::name() + _("]"));
+};
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
+
+#if defined(__GNUG__) || defined(__clang__)
+#  pragma GCC diagnostic pop
+#elif defined(_MSC_VER)
+#  pragma warning(pop)
+#endif
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/embed.h b/ml/dlib/dlib/external/pybind11/include/pybind11/embed.h
new file mode 100644
index 000000000..9abc61c34
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/embed.h
@@ -0,0 +1,194 @@
+/*
+    pybind11/embed.h: Support for embedding the interpreter
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include "eval.h"
+
+#if defined(PYPY_VERSION)
+#  error Embedding the interpreter is not supported with PyPy
+#endif
+
+#if PY_MAJOR_VERSION >= 3
+#  define PYBIND11_EMBEDDED_MODULE_IMPL(name)            \
+      extern "C" PyObject *pybind11_init_impl_##name() { \
+          return pybind11_init_wrapper_##name();         \
+      }
+#else
+#  define PYBIND11_EMBEDDED_MODULE_IMPL(name)            \
+      extern "C" void pybind11_init_impl_##name() {      \
+          pybind11_init_wrapper_##name();                \
+      }
+#endif
+
+/** \rst
+    Add a new module to the table of builtins for the interpreter. Must be
+    defined in global scope. The first macro parameter is the name of the
+    module (without quotes). The second parameter is the variable which will
+    be used as the interface to add functions and classes to the module.
+
+    .. code-block:: cpp
+
+        PYBIND11_EMBEDDED_MODULE(example, m) {
+            // ... initialize functions and classes here
+            m.def("foo", []() {
+                return "Hello, World!";
+            });
+        }
+ \endrst */
+#define PYBIND11_EMBEDDED_MODULE(name, variable)                              \
+    static void PYBIND11_CONCAT(pybind11_init_, name)(pybind11::module &);    \
+    static PyObject PYBIND11_CONCAT(*pybind11_init_wrapper_, name)() {        \
+        auto m = pybind11::module(PYBIND11_TOSTRING(name));                   \
+        try {                                                                 \
+            PYBIND11_CONCAT(pybind11_init_, name)(m);                         \
+            return m.ptr();                                                   \
+        } catch (pybind11::error_already_set &e) {                            \
+            PyErr_SetString(PyExc_ImportError, e.what());                     \
+            return nullptr;                                                   \
+        } catch (const std::exception &e) {                                   \
+            PyErr_SetString(PyExc_ImportError, e.what());                     \
+            return nullptr;                                                   \
+        }                                                                     \
+    }                                                                         \
+    PYBIND11_EMBEDDED_MODULE_IMPL(name)                                       \
+    pybind11::detail::embedded_module name(PYBIND11_TOSTRING(name),           \
+                               PYBIND11_CONCAT(pybind11_init_impl_, name));   \
+    void PYBIND11_CONCAT(pybind11_init_, name)(pybind11::module &variable)
+
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/// Python 2.7/3.x compatible version of `PyImport_AppendInittab` and error checks.
+struct embedded_module {
+#if PY_MAJOR_VERSION >= 3
+    using init_t = PyObject *(*)();
+#else
+    using init_t = void (*)();
+#endif
+    embedded_module(const char *name, init_t init) {
+        if (Py_IsInitialized())
+            pybind11_fail("Can't add new modules after the interpreter has been initialized");
+
+        auto result = PyImport_AppendInittab(name, init);
+        if (result == -1)
+            pybind11_fail("Insufficient memory to add a new module");
+    }
+};
+
+NAMESPACE_END(detail)
+
+/** \rst
+    Initialize the Python interpreter. No other pybind11 or CPython API functions can be
+    called before this is done; with the exception of `PYBIND11_EMBEDDED_MODULE`. The
+    optional parameter can be used to skip the registration of signal handlers (see the
+    Python documentation for details). Calling this function again after the interpreter
+    has already been initialized is a fatal error.
+ \endrst */
+inline void initialize_interpreter(bool init_signal_handlers = true) {
+    if (Py_IsInitialized())
+        pybind11_fail("The interpreter is already running");
+
+    Py_InitializeEx(init_signal_handlers ? 1 : 0);
+
+    // Make .py files in the working directory available by default
+    module::import("sys").attr("path").cast<list>().append(".");
+}
+
+/** \rst
+    Shut down the Python interpreter. No pybind11 or CPython API functions can be called
+    after this. In addition, pybind11 objects must not outlive the interpreter:
+
+    .. code-block:: cpp
+
+        { // BAD
+            py::initialize_interpreter();
+            auto hello = py::str("Hello, World!");
+            py::finalize_interpreter();
+        } // <-- BOOM, hello's destructor is called after interpreter shutdown
+
+        { // GOOD
+            py::initialize_interpreter();
+            { // scoped
+                auto hello = py::str("Hello, World!");
+            } // <-- OK, hello is cleaned up properly
+            py::finalize_interpreter();
+        }
+
+        { // BETTER
+            py::scoped_interpreter guard{};
+            auto hello = py::str("Hello, World!");
+        }
+
+    .. warning::
+
+        The interpreter can be restarted by calling `initialize_interpreter` again.
+        Modules created using pybind11 can be safely re-initialized. However, Python
+        itself cannot completely unload binary extension modules and there are several
+        caveats with regard to interpreter restarting. All the details can be found
+        in the CPython documentation. In short, not all interpreter memory may be
+        freed, either due to reference cycles or user-created global data.
+
+ \endrst */
+inline void finalize_interpreter() {
+    handle builtins(PyEval_GetBuiltins());
+    const char *id = PYBIND11_INTERNALS_ID;
+
+    // Get the internals pointer (without creating it if it doesn't exist).  It's possible for the
+    // internals to be created during Py_Finalize() (e.g. if a py::capsule calls `get_internals()`
+    // during destruction), so we get the pointer-pointer here and check it after Py_Finalize().
+    detail::internals **internals_ptr_ptr = detail::get_internals_pp();
+    // It could also be stashed in builtins, so look there too:
+    if (builtins.contains(id) && isinstance<capsule>(builtins[id]))
+        internals_ptr_ptr = capsule(builtins[id]);
+
+    Py_Finalize();
+
+    if (internals_ptr_ptr) {
+        delete *internals_ptr_ptr;
+        *internals_ptr_ptr = nullptr;
+    }
+}
+
+/** \rst
+    Scope guard version of `initialize_interpreter` and `finalize_interpreter`.
+    This a move-only guard and only a single instance can exist.
+
+    .. code-block:: cpp
+
+        #include <pybind11/embed.h>
+
+        int main() {
+            py::scoped_interpreter guard{};
+            py::print(Hello, World!);
+        } // <-- interpreter shutdown
+ \endrst */
+class scoped_interpreter {
+public:
+    scoped_interpreter(bool init_signal_handlers = true) {
+        initialize_interpreter(init_signal_handlers);
+    }
+
+    scoped_interpreter(const scoped_interpreter &) = delete;
+    scoped_interpreter(scoped_interpreter &&other) noexcept { other.is_valid = false; }
+    scoped_interpreter &operator=(const scoped_interpreter &) = delete;
+    scoped_interpreter &operator=(scoped_interpreter &&) = delete;
+
+    ~scoped_interpreter() {
+        if (is_valid)
+            finalize_interpreter();
+    }
+
+private:
+    bool is_valid = true;
+};
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/eval.h b/ml/dlib/dlib/external/pybind11/include/pybind11/eval.h
new file mode 100644
index 000000000..ea85ba1db
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/eval.h
@@ -0,0 +1,117 @@
+/*
+    pybind11/exec.h: Support for evaluating Python expressions and statements
+    from strings and files
+
+    Copyright (c) 2016 Klemens Morgenstern <klemens.morgenstern@ed-chemnitz.de> and
+                       Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+enum eval_mode {
+    /// Evaluate a string containing an isolated expression
+    eval_expr,
+
+    /// Evaluate a string containing a single statement. Returns \c none
+    eval_single_statement,
+
+    /// Evaluate a string containing a sequence of statement. Returns \c none
+    eval_statements
+};
+
+template <eval_mode mode = eval_expr>
+object eval(str expr, object global = globals(), object local = object()) {
+    if (!local)
+        local = global;
+
+    /* PyRun_String does not accept a PyObject / encoding specifier,
+       this seems to be the only alternative */
+    std::string buffer = "# -*- coding: utf-8 -*-\n" + (std::string) expr;
+
+    int start;
+    switch (mode) {
+        case eval_expr:             start = Py_eval_input;   break;
+        case eval_single_statement: start = Py_single_input; break;
+        case eval_statements:       start = Py_file_input;   break;
+        default: pybind11_fail("invalid evaluation mode");
+    }
+
+    PyObject *result = PyRun_String(buffer.c_str(), start, global.ptr(), local.ptr());
+    if (!result)
+        throw error_already_set();
+    return reinterpret_steal<object>(result);
+}
+
+template <eval_mode mode = eval_expr, size_t N>
+object eval(const char (&s)[N], object global = globals(), object local = object()) {
+    /* Support raw string literals by removing common leading whitespace */
+    auto expr = (s[0] == '\n') ? str(module::import("textwrap").attr("dedent")(s))
+                               : str(s);
+    return eval<mode>(expr, global, local);
+}
+
+inline void exec(str expr, object global = globals(), object local = object()) {
+    eval<eval_statements>(expr, global, local);
+}
+
+template <size_t N>
+void exec(const char (&s)[N], object global = globals(), object local = object()) {
+    eval<eval_statements>(s, global, local);
+}
+
+template <eval_mode mode = eval_statements>
+object eval_file(str fname, object global = globals(), object local = object()) {
+    if (!local)
+        local = global;
+
+    int start;
+    switch (mode) {
+        case eval_expr:             start = Py_eval_input;   break;
+        case eval_single_statement: start = Py_single_input; break;
+        case eval_statements:       start = Py_file_input;   break;
+        default: pybind11_fail("invalid evaluation mode");
+    }
+
+    int closeFile = 1;
+    std::string fname_str = (std::string) fname;
+#if PY_VERSION_HEX >= 0x03040000
+    FILE *f = _Py_fopen_obj(fname.ptr(), "r");
+#elif PY_VERSION_HEX >= 0x03000000
+    FILE *f = _Py_fopen(fname.ptr(), "r");
+#else
+    /* No unicode support in open() :( */
+    auto fobj = reinterpret_steal<object>(PyFile_FromString(
+        const_cast<char *>(fname_str.c_str()),
+        const_cast<char*>("r")));
+    FILE *f = nullptr;
+    if (fobj)
+        f = PyFile_AsFile(fobj.ptr());
+    closeFile = 0;
+#endif
+    if (!f) {
+        PyErr_Clear();
+        pybind11_fail("File \"" + fname_str + "\" could not be opened!");
+    }
+
+#if PY_VERSION_HEX < 0x03000000 && defined(PYPY_VERSION)
+    PyObject *result = PyRun_File(f, fname_str.c_str(), start, global.ptr(),
+                                  local.ptr());
+    (void) closeFile;
+#else
+    PyObject *result = PyRun_FileEx(f, fname_str.c_str(), start, global.ptr(),
+                                    local.ptr(), closeFile);
+#endif
+
+    if (!result)
+        throw error_already_set();
+    return reinterpret_steal<object>(result);
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/functional.h b/ml/dlib/dlib/external/pybind11/include/pybind11/functional.h
new file mode 100644
index 000000000..eda14ba58
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/functional.h
@@ -0,0 +1,85 @@
+/*
+    pybind11/functional.h: std::function<> support
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include <functional>
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+template <typename Return, typename... Args>
+struct type_caster<std::function<Return(Args...)>> {
+    using type = std::function<Return(Args...)>;
+    using retval_type = conditional_t<std::is_same<Return, void>::value, void_type, Return>;
+    using function_type = Return (*) (Args...);
+
+public:
+    bool load(handle src, bool convert) {
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) return false;
+            return true;
+        }
+
+        if (!isinstance<function>(src))
+            return false;
+
+        auto func = reinterpret_borrow<function>(src);
+
+        /*
+           When passing a C++ function as an argument to another C++
+           function via Python, every function call would normally involve
+           a full C++ -> Python -> C++ roundtrip, which can be prohibitive.
+           Here, we try to at least detect the case where the function is
+           stateless (i.e. function pointer or lambda function without
+           captured variables), in which case the roundtrip can be avoided.
+         */
+        if (auto cfunc = func.cpp_function()) {
+            auto c = reinterpret_borrow<capsule>(PyCFunction_GET_SELF(cfunc.ptr()));
+            auto rec = (function_record *) c;
+
+            if (rec && rec->is_stateless &&
+                    same_type(typeid(function_type), *reinterpret_cast<const std::type_info *>(rec->data[1]))) {
+                struct capture { function_type f; };
+                value = ((capture *) &rec->data)->f;
+                return true;
+            }
+        }
+
+        value = [func](Args... args) -> Return {
+            gil_scoped_acquire acq;
+            object retval(func(std::forward<Args>(args)...));
+            /* Visual studio 2015 parser issue: need parentheses around this expression */
+            return (retval.template cast<Return>());
+        };
+        return true;
+    }
+
+    template <typename Func>
+    static handle cast(Func &&f_, return_value_policy policy, handle /* parent */) {
+        if (!f_)
+            return none().inc_ref();
+
+        auto result = f_.template target<function_type>();
+        if (result)
+            return cpp_function(*result, policy).release();
+        else
+            return cpp_function(std::forward<Func>(f_), policy).release();
+    }
+
+    PYBIND11_TYPE_CASTER(type, _("Callable[[") +
+            argument_loader<Args...>::arg_names() + _("], ") +
+            make_caster<retval_type>::name() +
+            _("]"));
+};
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/iostream.h b/ml/dlib/dlib/external/pybind11/include/pybind11/iostream.h
new file mode 100644
index 000000000..a9c27aac1
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/iostream.h
@@ -0,0 +1,200 @@
+/*
+    pybind11/iostream.h -- Tools to assist with redirecting cout and cerr to Python
+
+    Copyright (c) 2017 Henry F. Schreiner
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#include <streambuf>
+#include <ostream>
+#include <string>
+#include <memory>
+#include <iostream>
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+// Buffer that writes to Python instead of C++
+class pythonbuf : public std::streambuf {
+private:
+    using traits_type = std::streambuf::traits_type;
+
+    char d_buffer[1024];
+    object pywrite;
+    object pyflush;
+
+    int overflow(int c) {
+        if (!traits_type::eq_int_type(c, traits_type::eof())) {
+            *pptr() = traits_type::to_char_type(c);
+            pbump(1);
+        }
+        return sync() ? traits_type::not_eof(c) : traits_type::eof();
+    }
+
+    int sync() {
+        if (pbase() != pptr()) {
+            // This subtraction cannot be negative, so dropping the sign
+            str line(pbase(), static_cast<size_t>(pptr() - pbase()));
+
+            pywrite(line);
+            pyflush();
+
+            setp(pbase(), epptr());
+        }
+        return 0;
+    }
+
+public:
+    pythonbuf(object pyostream)
+        : pywrite(pyostream.attr("write")),
+          pyflush(pyostream.attr("flush")) {
+        setp(d_buffer, d_buffer + sizeof(d_buffer) - 1);
+    }
+
+    /// Sync before destroy
+    ~pythonbuf() {
+        sync();
+    }
+};
+
+NAMESPACE_END(detail)
+
+
+/** \rst
+    This a move-only guard that redirects output.
+
+    .. code-block:: cpp
+
+        #include <pybind11/iostream.h>
+
+        ...
+
+        {
+            py::scoped_ostream_redirect output;
+            std::cout << "Hello, World!"; // Python stdout
+        } // <-- return std::cout to normal
+
+    You can explicitly pass the c++ stream and the python object,
+    for example to guard stderr instead.
+
+    .. code-block:: cpp
+
+        {
+            py::scoped_ostream_redirect output{std::cerr, py::module::import("sys").attr("stderr")};
+            std::cerr << "Hello, World!";
+        }
+ \endrst */
+class scoped_ostream_redirect {
+protected:
+    std::streambuf *old;
+    std::ostream &costream;
+    detail::pythonbuf buffer;
+
+public:
+    scoped_ostream_redirect(
+            std::ostream &costream = std::cout,
+            object pyostream = module::import("sys").attr("stdout"))
+        : costream(costream), buffer(pyostream) {
+        old = costream.rdbuf(&buffer);
+    }
+
+    ~scoped_ostream_redirect() {
+        costream.rdbuf(old);
+    }
+
+    scoped_ostream_redirect(const scoped_ostream_redirect &) = delete;
+    scoped_ostream_redirect(scoped_ostream_redirect &&other) = default;
+    scoped_ostream_redirect &operator=(const scoped_ostream_redirect &) = delete;
+    scoped_ostream_redirect &operator=(scoped_ostream_redirect &&) = delete;
+};
+
+
+/** \rst
+    Like `scoped_ostream_redirect`, but redirects cerr by default. This class
+    is provided primary to make ``py::call_guard`` easier to make.
+
+    .. code-block:: cpp
+
+     m.def("noisy_func", &noisy_func,
+           py::call_guard<scoped_ostream_redirect,
+                          scoped_estream_redirect>());
+
+\endrst */
+class scoped_estream_redirect : public scoped_ostream_redirect {
+public:
+    scoped_estream_redirect(
+            std::ostream &costream = std::cerr,
+            object pyostream = module::import("sys").attr("stderr"))
+        : scoped_ostream_redirect(costream,pyostream) {}
+};
+
+
+NAMESPACE_BEGIN(detail)
+
+// Class to redirect output as a context manager. C++ backend.
+class OstreamRedirect {
+    bool do_stdout_;
+    bool do_stderr_;
+    std::unique_ptr<scoped_ostream_redirect> redirect_stdout;
+    std::unique_ptr<scoped_estream_redirect> redirect_stderr;
+
+public:
+    OstreamRedirect(bool do_stdout = true, bool do_stderr = true)
+        : do_stdout_(do_stdout), do_stderr_(do_stderr) {}
+
+    void enter() {
+        if (do_stdout_)
+            redirect_stdout.reset(new scoped_ostream_redirect());
+        if (do_stderr_)
+            redirect_stderr.reset(new scoped_estream_redirect());
+    }
+
+    void exit() {
+        redirect_stdout.reset();
+        redirect_stderr.reset();
+    }
+};
+
+NAMESPACE_END(detail)
+
+/** \rst
+    This is a helper function to add a C++ redirect context manager to Python
+    instead of using a C++ guard. To use it, add the following to your binding code:
+
+    .. code-block:: cpp
+
+        #include <pybind11/iostream.h>
+
+        ...
+
+        py::add_ostream_redirect(m, "ostream_redirect");
+
+    You now have a Python context manager that redirects your output:
+
+    .. code-block:: python
+
+        with m.ostream_redirect():
+            m.print_to_cout_function()
+
+    This manager can optionally be told which streams to operate on:
+
+    .. code-block:: python
+
+        with m.ostream_redirect(stdout=true, stderr=true):
+            m.noisy_function_with_error_printing()
+
+ \endrst */
+inline class_<detail::OstreamRedirect> add_ostream_redirect(module m, std::string name = "ostream_redirect") {
+    return class_<detail::OstreamRedirect>(m, name.c_str(), module_local())
+        .def(init<bool,bool>(), arg("stdout")=true, arg("stderr")=true)
+        .def("__enter__", &detail::OstreamRedirect::enter)
+        .def("__exit__", [](detail::OstreamRedirect &self, args) { self.exit(); });
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/numpy.h b/ml/dlib/dlib/external/pybind11/include/pybind11/numpy.h
new file mode 100644
index 000000000..b1600dc2e
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/numpy.h
@@ -0,0 +1,1600 @@
+/*
+    pybind11/numpy.h: Basic NumPy support, vectorize() wrapper
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include "complex.h"
+#include <numeric>
+#include <algorithm>
+#include <array>
+#include <cstdlib>
+#include <cstring>
+#include <sstream>
+#include <string>
+#include <initializer_list>
+#include <functional>
+#include <utility>
+#include <typeindex>
+
+#if defined(_MSC_VER)
+#  pragma warning(push)
+#  pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#endif
+
+/* This will be true on all flat address space platforms and allows us to reduce the
+   whole npy_intp / ssize_t / Py_intptr_t business down to just ssize_t for all size
+   and dimension types (e.g. shape, strides, indexing), instead of inflicting this
+   upon the library user. */
+static_assert(sizeof(ssize_t) == sizeof(Py_intptr_t), "ssize_t != Py_intptr_t");
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+class array; // Forward declaration
+
+NAMESPACE_BEGIN(detail)
+template <typename type, typename SFINAE = void> struct npy_format_descriptor;
+
+struct PyArrayDescr_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char flags;
+    int type_num;
+    int elsize;
+    int alignment;
+    char *subarray;
+    PyObject *fields;
+    PyObject *names;
+};
+
+struct PyArray_Proxy {
+    PyObject_HEAD
+    char *data;
+    int nd;
+    ssize_t *dimensions;
+    ssize_t *strides;
+    PyObject *base;
+    PyObject *descr;
+    int flags;
+};
+
+struct PyVoidScalarObject_Proxy {
+    PyObject_VAR_HEAD
+    char *obval;
+    PyArrayDescr_Proxy *descr;
+    int flags;
+    PyObject *base;
+};
+
+struct numpy_type_info {
+    PyObject* dtype_ptr;
+    std::string format_str;
+};
+
+struct numpy_internals {
+    std::unordered_map<std::type_index, numpy_type_info> registered_dtypes;
+
+    numpy_type_info *get_type_info(const std::type_info& tinfo, bool throw_if_missing = true) {
+        auto it = registered_dtypes.find(std::type_index(tinfo));
+        if (it != registered_dtypes.end())
+            return &(it->second);
+        if (throw_if_missing)
+            pybind11_fail(std::string("NumPy type info missing for ") + tinfo.name());
+        return nullptr;
+    }
+
+    template<typename T> numpy_type_info *get_type_info(bool throw_if_missing = true) {
+        return get_type_info(typeid(typename std::remove_cv<T>::type), throw_if_missing);
+    }
+};
+
+inline PYBIND11_NOINLINE void load_numpy_internals(numpy_internals* &ptr) {
+    ptr = &get_or_create_shared_data<numpy_internals>("_numpy_internals");
+}
+
+inline numpy_internals& get_numpy_internals() {
+    static numpy_internals* ptr = nullptr;
+    if (!ptr)
+        load_numpy_internals(ptr);
+    return *ptr;
+}
+
+struct npy_api {
+    enum constants {
+        NPY_ARRAY_C_CONTIGUOUS_ = 0x0001,
+        NPY_ARRAY_F_CONTIGUOUS_ = 0x0002,
+        NPY_ARRAY_OWNDATA_ = 0x0004,
+        NPY_ARRAY_FORCECAST_ = 0x0010,
+        NPY_ARRAY_ENSUREARRAY_ = 0x0040,
+        NPY_ARRAY_ALIGNED_ = 0x0100,
+        NPY_ARRAY_WRITEABLE_ = 0x0400,
+        NPY_BOOL_ = 0,
+        NPY_BYTE_, NPY_UBYTE_,
+        NPY_SHORT_, NPY_USHORT_,
+        NPY_INT_, NPY_UINT_,
+        NPY_LONG_, NPY_ULONG_,
+        NPY_LONGLONG_, NPY_ULONGLONG_,
+        NPY_FLOAT_, NPY_DOUBLE_, NPY_LONGDOUBLE_,
+        NPY_CFLOAT_, NPY_CDOUBLE_, NPY_CLONGDOUBLE_,
+        NPY_OBJECT_ = 17,
+        NPY_STRING_, NPY_UNICODE_, NPY_VOID_
+    };
+
+    typedef struct {
+        Py_intptr_t *ptr;
+        int len;
+    } PyArray_Dims;
+
+    static npy_api& get() {
+        static npy_api api = lookup();
+        return api;
+    }
+
+    bool PyArray_Check_(PyObject *obj) const {
+        return (bool) PyObject_TypeCheck(obj, PyArray_Type_);
+    }
+    bool PyArrayDescr_Check_(PyObject *obj) const {
+        return (bool) PyObject_TypeCheck(obj, PyArrayDescr_Type_);
+    }
+
+    unsigned int (*PyArray_GetNDArrayCFeatureVersion_)();
+    PyObject *(*PyArray_DescrFromType_)(int);
+    PyObject *(*PyArray_NewFromDescr_)
+        (PyTypeObject *, PyObject *, int, Py_intptr_t *,
+         Py_intptr_t *, void *, int, PyObject *);
+    PyObject *(*PyArray_DescrNewFromType_)(int);
+    int (*PyArray_CopyInto_)(PyObject *, PyObject *);
+    PyObject *(*PyArray_NewCopy_)(PyObject *, int);
+    PyTypeObject *PyArray_Type_;
+    PyTypeObject *PyVoidArrType_Type_;
+    PyTypeObject *PyArrayDescr_Type_;
+    PyObject *(*PyArray_DescrFromScalar_)(PyObject *);
+    PyObject *(*PyArray_FromAny_) (PyObject *, PyObject *, int, int, int, PyObject *);
+    int (*PyArray_DescrConverter_) (PyObject *, PyObject **);
+    bool (*PyArray_EquivTypes_) (PyObject *, PyObject *);
+    int (*PyArray_GetArrayParamsFromObject_)(PyObject *, PyObject *, char, PyObject **, int *,
+                                             Py_ssize_t *, PyObject **, PyObject *);
+    PyObject *(*PyArray_Squeeze_)(PyObject *);
+    int (*PyArray_SetBaseObject_)(PyObject *, PyObject *);
+    PyObject* (*PyArray_Resize_)(PyObject*, PyArray_Dims*, int, int);
+private:
+    enum functions {
+        API_PyArray_GetNDArrayCFeatureVersion = 211,
+        API_PyArray_Type = 2,
+        API_PyArrayDescr_Type = 3,
+        API_PyVoidArrType_Type = 39,
+        API_PyArray_DescrFromType = 45,
+        API_PyArray_DescrFromScalar = 57,
+        API_PyArray_FromAny = 69,
+        API_PyArray_Resize = 80,
+        API_PyArray_CopyInto = 82,
+        API_PyArray_NewCopy = 85,
+        API_PyArray_NewFromDescr = 94,
+        API_PyArray_DescrNewFromType = 9,
+        API_PyArray_DescrConverter = 174,
+        API_PyArray_EquivTypes = 182,
+        API_PyArray_GetArrayParamsFromObject = 278,
+        API_PyArray_Squeeze = 136,
+        API_PyArray_SetBaseObject = 282
+    };
+
+    static npy_api lookup() {
+        module m = module::import("numpy.core.multiarray");
+        auto c = m.attr("_ARRAY_API");
+#if PY_MAJOR_VERSION >= 3
+        void **api_ptr = (void **) PyCapsule_GetPointer(c.ptr(), NULL);
+#else
+        void **api_ptr = (void **) PyCObject_AsVoidPtr(c.ptr());
+#endif
+        npy_api api;
+#define DECL_NPY_API(Func) api.Func##_ = (decltype(api.Func##_)) api_ptr[API_##Func];
+        DECL_NPY_API(PyArray_GetNDArrayCFeatureVersion);
+        if (api.PyArray_GetNDArrayCFeatureVersion_() < 0x7)
+            pybind11_fail("pybind11 numpy support requires numpy >= 1.7.0");
+        DECL_NPY_API(PyArray_Type);
+        DECL_NPY_API(PyVoidArrType_Type);
+        DECL_NPY_API(PyArrayDescr_Type);
+        DECL_NPY_API(PyArray_DescrFromType);
+        DECL_NPY_API(PyArray_DescrFromScalar);
+        DECL_NPY_API(PyArray_FromAny);
+        DECL_NPY_API(PyArray_Resize);
+        DECL_NPY_API(PyArray_CopyInto);
+        DECL_NPY_API(PyArray_NewCopy);
+        DECL_NPY_API(PyArray_NewFromDescr);
+        DECL_NPY_API(PyArray_DescrNewFromType);
+        DECL_NPY_API(PyArray_DescrConverter);
+        DECL_NPY_API(PyArray_EquivTypes);
+        DECL_NPY_API(PyArray_GetArrayParamsFromObject);
+        DECL_NPY_API(PyArray_Squeeze);
+        DECL_NPY_API(PyArray_SetBaseObject);
+#undef DECL_NPY_API
+        return api;
+    }
+};
+
+inline PyArray_Proxy* array_proxy(void* ptr) {
+    return reinterpret_cast<PyArray_Proxy*>(ptr);
+}
+
+inline const PyArray_Proxy* array_proxy(const void* ptr) {
+    return reinterpret_cast<const PyArray_Proxy*>(ptr);
+}
+
+inline PyArrayDescr_Proxy* array_descriptor_proxy(PyObject* ptr) {
+   return reinterpret_cast<PyArrayDescr_Proxy*>(ptr);
+}
+
+inline const PyArrayDescr_Proxy* array_descriptor_proxy(const PyObject* ptr) {
+   return reinterpret_cast<const PyArrayDescr_Proxy*>(ptr);
+}
+
+inline bool check_flags(const void* ptr, int flag) {
+    return (flag == (array_proxy(ptr)->flags & flag));
+}
+
+template <typename T> struct is_std_array : std::false_type { };
+template <typename T, size_t N> struct is_std_array<std::array<T, N>> : std::true_type { };
+template <typename T> struct is_complex : std::false_type { };
+template <typename T> struct is_complex<std::complex<T>> : std::true_type { };
+
+template <typename T> struct array_info_scalar {
+    typedef T type;
+    static constexpr bool is_array = false;
+    static constexpr bool is_empty = false;
+    static PYBIND11_DESCR extents() { return _(""); }
+    static void append_extents(list& /* shape */) { }
+};
+// Computes underlying type and a comma-separated list of extents for array
+// types (any mix of std::array and built-in arrays). An array of char is
+// treated as scalar because it gets special handling.
+template <typename T> struct array_info : array_info_scalar<T> { };
+template <typename T, size_t N> struct array_info<std::array<T, N>> {
+    using type = typename array_info<T>::type;
+    static constexpr bool is_array = true;
+    static constexpr bool is_empty = (N == 0) || array_info<T>::is_empty;
+    static constexpr size_t extent = N;
+
+    // appends the extents to shape
+    static void append_extents(list& shape) {
+        shape.append(N);
+        array_info<T>::append_extents(shape);
+    }
+
+    template<typename T2 = T, enable_if_t<!array_info<T2>::is_array, int> = 0>
+    static PYBIND11_DESCR extents() {
+        return _<N>();
+    }
+
+    template<typename T2 = T, enable_if_t<array_info<T2>::is_array, int> = 0>
+    static PYBIND11_DESCR extents() {
+        return concat(_<N>(), array_info<T>::extents());
+    }
+};
+// For numpy we have special handling for arrays of characters, so we don't include
+// the size in the array extents.
+template <size_t N> struct array_info<char[N]> : array_info_scalar<char[N]> { };
+template <size_t N> struct array_info<std::array<char, N>> : array_info_scalar<std::array<char, N>> { };
+template <typename T, size_t N> struct array_info<T[N]> : array_info<std::array<T, N>> { };
+template <typename T> using remove_all_extents_t = typename array_info<T>::type;
+
+template <typename T> using is_pod_struct = all_of<
+    std::is_standard_layout<T>,     // since we're accessing directly in memory we need a standard layout type
+#if !defined(__GNUG__) || defined(_LIBCPP_VERSION) || defined(_GLIBCXX_USE_CXX11_ABI)
+    // _GLIBCXX_USE_CXX11_ABI indicates that we're using libstdc++ from GCC 5 or newer, independent
+    // of the actual compiler (Clang can also use libstdc++, but it always defines __GNUC__ == 4).
+    std::is_trivially_copyable<T>,
+#else
+    // GCC 4 doesn't implement is_trivially_copyable, so approximate it
+    std::is_trivially_destructible<T>,
+    satisfies_any_of<T, std::has_trivial_copy_constructor, std::has_trivial_copy_assign>,
+#endif
+    satisfies_none_of<T, std::is_reference, std::is_array, is_std_array, std::is_arithmetic, is_complex, std::is_enum>
+>;
+
+template <ssize_t Dim = 0, typename Strides> ssize_t byte_offset_unsafe(const Strides &) { return 0; }
+template <ssize_t Dim = 0, typename Strides, typename... Ix>
+ssize_t byte_offset_unsafe(const Strides &strides, ssize_t i, Ix... index) {
+    return i * strides[Dim] + byte_offset_unsafe<Dim + 1>(strides, index...);
+}
+
+/**
+ * Proxy class providing unsafe, unchecked const access to array data.  This is constructed through
+ * the `unchecked<T, N>()` method of `array` or the `unchecked<N>()` method of `array_t<T>`.  `Dims`
+ * will be -1 for dimensions determined at runtime.
+ */
+template <typename T, ssize_t Dims>
+class unchecked_reference {
+protected:
+    static constexpr bool Dynamic = Dims < 0;
+    const unsigned char *data_;
+    // Storing the shape & strides in local variables (i.e. these arrays) allows the compiler to
+    // make large performance gains on big, nested loops, but requires compile-time dimensions
+    conditional_t<Dynamic, const ssize_t *, std::array<ssize_t, (size_t) Dims>>
+            shape_, strides_;
+    const ssize_t dims_;
+
+    friend class pybind11::array;
+    // Constructor for compile-time dimensions:
+    template <bool Dyn = Dynamic>
+    unchecked_reference(const void *data, const ssize_t *shape, const ssize_t *strides, enable_if_t<!Dyn, ssize_t>)
+    : data_{reinterpret_cast<const unsigned char *>(data)}, dims_{Dims} {
+        for (size_t i = 0; i < (size_t) dims_; i++) {
+            shape_[i] = shape[i];
+            strides_[i] = strides[i];
+        }
+    }
+    // Constructor for runtime dimensions:
+    template <bool Dyn = Dynamic>
+    unchecked_reference(const void *data, const ssize_t *shape, const ssize_t *strides, enable_if_t<Dyn, ssize_t> dims)
+    : data_{reinterpret_cast<const unsigned char *>(data)}, shape_{shape}, strides_{strides}, dims_{dims} {}
+
+public:
+    /**
+     * Unchecked const reference access to data at the given indices.  For a compile-time known
+     * number of dimensions, this requires the correct number of arguments; for run-time
+     * dimensionality, this is not checked (and so is up to the caller to use safely).
+     */
+    template <typename... Ix> const T &operator()(Ix... index) const {
+        static_assert(ssize_t{sizeof...(Ix)} == Dims || Dynamic,
+                "Invalid number of indices for unchecked array reference");
+        return *reinterpret_cast<const T *>(data_ + byte_offset_unsafe(strides_, ssize_t(index)...));
+    }
+    /**
+     * Unchecked const reference access to data; this operator only participates if the reference
+     * is to a 1-dimensional array.  When present, this is exactly equivalent to `obj(index)`.
+     */
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    const T &operator[](ssize_t index) const { return operator()(index); }
+
+    /// Pointer access to the data at the given indices.
+    template <typename... Ix> const T *data(Ix... ix) const { return &operator()(ssize_t(ix)...); }
+
+    /// Returns the item size, i.e. sizeof(T)
+    constexpr static ssize_t itemsize() { return sizeof(T); }
+
+    /// Returns the shape (i.e. size) of dimension `dim`
+    ssize_t shape(ssize_t dim) const { return shape_[(size_t) dim]; }
+
+    /// Returns the number of dimensions of the array
+    ssize_t ndim() const { return dims_; }
+
+    /// Returns the total number of elements in the referenced array, i.e. the product of the shapes
+    template <bool Dyn = Dynamic>
+    enable_if_t<!Dyn, ssize_t> size() const {
+        return std::accumulate(shape_.begin(), shape_.end(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+    template <bool Dyn = Dynamic>
+    enable_if_t<Dyn, ssize_t> size() const {
+        return std::accumulate(shape_, shape_ + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+
+    /// Returns the total number of bytes used by the referenced data.  Note that the actual span in
+    /// memory may be larger if the referenced array has non-contiguous strides (e.g. for a slice).
+    ssize_t nbytes() const {
+        return size() * itemsize();
+    }
+};
+
+template <typename T, ssize_t Dims>
+class unchecked_mutable_reference : public unchecked_reference<T, Dims> {
+    friend class pybind11::array;
+    using ConstBase = unchecked_reference<T, Dims>;
+    using ConstBase::ConstBase;
+    using ConstBase::Dynamic;
+public:
+    /// Mutable, unchecked access to data at the given indices.
+    template <typename... Ix> T& operator()(Ix... index) {
+        static_assert(ssize_t{sizeof...(Ix)} == Dims || Dynamic,
+                "Invalid number of indices for unchecked array reference");
+        return const_cast<T &>(ConstBase::operator()(index...));
+    }
+    /**
+     * Mutable, unchecked access data at the given index; this operator only participates if the
+     * reference is to a 1-dimensional array (or has runtime dimensions).  When present, this is
+     * exactly equivalent to `obj(index)`.
+     */
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    T &operator[](ssize_t index) { return operator()(index); }
+
+    /// Mutable pointer access to the data at the given indices.
+    template <typename... Ix> T *mutable_data(Ix... ix) { return &operator()(ssize_t(ix)...); }
+};
+
+template <typename T, ssize_t Dim>
+struct type_caster<unchecked_reference<T, Dim>> {
+    static_assert(Dim == 0 && Dim > 0 /* always fail */, "unchecked array proxy object is not castable");
+};
+template <typename T, ssize_t Dim>
+struct type_caster<unchecked_mutable_reference<T, Dim>> : type_caster<unchecked_reference<T, Dim>> {};
+
+NAMESPACE_END(detail)
+
+class dtype : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(dtype, object, detail::npy_api::get().PyArrayDescr_Check_);
+
+    explicit dtype(const buffer_info &info) {
+        dtype descr(_dtype_from_pep3118()(PYBIND11_STR_TYPE(info.format)));
+        // If info.itemsize == 0, use the value calculated from the format string
+        m_ptr = descr.strip_padding(info.itemsize ? info.itemsize : descr.itemsize()).release().ptr();
+    }
+
+    explicit dtype(const std::string &format) {
+        m_ptr = from_args(pybind11::str(format)).release().ptr();
+    }
+
+    dtype(const char *format) : dtype(std::string(format)) { }
+
+    dtype(list names, list formats, list offsets, ssize_t itemsize) {
+        dict args;
+        args["names"] = names;
+        args["formats"] = formats;
+        args["offsets"] = offsets;
+        args["itemsize"] = pybind11::int_(itemsize);
+        m_ptr = from_args(args).release().ptr();
+    }
+
+    /// This is essentially the same as calling numpy.dtype(args) in Python.
+    static dtype from_args(object args) {
+        PyObject *ptr = nullptr;
+        if (!detail::npy_api::get().PyArray_DescrConverter_(args.release().ptr(), &ptr) || !ptr)
+            throw error_already_set();
+        return reinterpret_steal<dtype>(ptr);
+    }
+
+    /// Return dtype associated with a C++ type.
+    template <typename T> static dtype of() {
+        return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::dtype();
+    }
+
+    /// Size of the data type in bytes.
+    ssize_t itemsize() const {
+        return detail::array_descriptor_proxy(m_ptr)->elsize;
+    }
+
+    /// Returns true for structured data types.
+    bool has_fields() const {
+        return detail::array_descriptor_proxy(m_ptr)->names != nullptr;
+    }
+
+    /// Single-character type code.
+    char kind() const {
+        return detail::array_descriptor_proxy(m_ptr)->kind;
+    }
+
+private:
+    static object _dtype_from_pep3118() {
+        static PyObject *obj = module::import("numpy.core._internal")
+            .attr("_dtype_from_pep3118").cast<object>().release().ptr();
+        return reinterpret_borrow<object>(obj);
+    }
+
+    dtype strip_padding(ssize_t itemsize) {
+        // Recursively strip all void fields with empty names that are generated for
+        // padding fields (as of NumPy v1.11).
+        if (!has_fields())
+            return *this;
+
+        struct field_descr { PYBIND11_STR_TYPE name; object format; pybind11::int_ offset; };
+        std::vector<field_descr> field_descriptors;
+
+        for (auto field : attr("fields").attr("items")()) {
+            auto spec = field.cast<tuple>();
+            auto name = spec[0].cast<pybind11::str>();
+            auto format = spec[1].cast<tuple>()[0].cast<dtype>();
+            auto offset = spec[1].cast<tuple>()[1].cast<pybind11::int_>();
+            if (!len(name) && format.kind() == 'V')
+                continue;
+            field_descriptors.push_back({(PYBIND11_STR_TYPE) name, format.strip_padding(format.itemsize()), offset});
+        }
+
+        std::sort(field_descriptors.begin(), field_descriptors.end(),
+                  [](const field_descr& a, const field_descr& b) {
+                      return a.offset.cast<int>() < b.offset.cast<int>();
+                  });
+
+        list names, formats, offsets;
+        for (auto& descr : field_descriptors) {
+            names.append(descr.name);
+            formats.append(descr.format);
+            offsets.append(descr.offset);
+        }
+        return dtype(names, formats, offsets, itemsize);
+    }
+};
+
+class array : public buffer {
+public:
+    PYBIND11_OBJECT_CVT(array, buffer, detail::npy_api::get().PyArray_Check_, raw_array)
+
+    enum {
+        c_style = detail::npy_api::NPY_ARRAY_C_CONTIGUOUS_,
+        f_style = detail::npy_api::NPY_ARRAY_F_CONTIGUOUS_,
+        forcecast = detail::npy_api::NPY_ARRAY_FORCECAST_
+    };
+
+    array() : array({{0}}, static_cast<const double *>(nullptr)) {}
+
+    using ShapeContainer = detail::any_container<ssize_t>;
+    using StridesContainer = detail::any_container<ssize_t>;
+
+    // Constructs an array taking shape/strides from arbitrary container types
+    array(const pybind11::dtype &dt, ShapeContainer shape, StridesContainer strides,
+          const void *ptr = nullptr, handle base = handle()) {
+
+        if (strides->empty())
+            *strides = c_strides(*shape, dt.itemsize());
+
+        auto ndim = shape->size();
+        if (ndim != strides->size())
+            pybind11_fail("NumPy: shape ndim doesn't match strides ndim");
+        auto descr = dt;
+
+        int flags = 0;
+        if (base && ptr) {
+            if (isinstance<array>(base))
+                /* Copy flags from base (except ownership bit) */
+                flags = reinterpret_borrow<array>(base).flags() & ~detail::npy_api::NPY_ARRAY_OWNDATA_;
+            else
+                /* Writable by default, easy to downgrade later on if needed */
+                flags = detail::npy_api::NPY_ARRAY_WRITEABLE_;
+        }
+
+        auto &api = detail::npy_api::get();
+        auto tmp = reinterpret_steal<object>(api.PyArray_NewFromDescr_(
+            api.PyArray_Type_, descr.release().ptr(), (int) ndim, shape->data(), strides->data(),
+            const_cast<void *>(ptr), flags, nullptr));
+        if (!tmp)
+            throw error_already_set();
+        if (ptr) {
+            if (base) {
+                api.PyArray_SetBaseObject_(tmp.ptr(), base.inc_ref().ptr());
+            } else {
+                tmp = reinterpret_steal<object>(api.PyArray_NewCopy_(tmp.ptr(), -1 /* any order */));
+            }
+        }
+        m_ptr = tmp.release().ptr();
+    }
+
+    array(const pybind11::dtype &dt, ShapeContainer shape, const void *ptr = nullptr, handle base = handle())
+        : array(dt, std::move(shape), {}, ptr, base) { }
+
+    template <typename T, typename = detail::enable_if_t<std::is_integral<T>::value && !std::is_same<bool, T>::value>>
+    array(const pybind11::dtype &dt, T count, const void *ptr = nullptr, handle base = handle())
+        : array(dt, {{count}}, ptr, base) { }
+
+    template <typename T>
+    array(ShapeContainer shape, StridesContainer strides, const T *ptr, handle base = handle())
+        : array(pybind11::dtype::of<T>(), std::move(shape), std::move(strides), ptr, base) { }
+
+    template <typename T>
+    array(ShapeContainer shape, const T *ptr, handle base = handle())
+        : array(std::move(shape), {}, ptr, base) { }
+
+    template <typename T>
+    explicit array(ssize_t count, const T *ptr, handle base = handle()) : array({count}, {}, ptr, base) { }
+
+    explicit array(const buffer_info &info)
+    : array(pybind11::dtype(info), info.shape, info.strides, info.ptr) { }
+
+    /// Array descriptor (dtype)
+    pybind11::dtype dtype() const {
+        return reinterpret_borrow<pybind11::dtype>(detail::array_proxy(m_ptr)->descr);
+    }
+
+    /// Total number of elements
+    ssize_t size() const {
+        return std::accumulate(shape(), shape() + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+
+    /// Byte size of a single element
+    ssize_t itemsize() const {
+        return detail::array_descriptor_proxy(detail::array_proxy(m_ptr)->descr)->elsize;
+    }
+
+    /// Total number of bytes
+    ssize_t nbytes() const {
+        return size() * itemsize();
+    }
+
+    /// Number of dimensions
+    ssize_t ndim() const {
+        return detail::array_proxy(m_ptr)->nd;
+    }
+
+    /// Base object
+    object base() const {
+        return reinterpret_borrow<object>(detail::array_proxy(m_ptr)->base);
+    }
+
+    /// Dimensions of the array
+    const ssize_t* shape() const {
+        return detail::array_proxy(m_ptr)->dimensions;
+    }
+
+    /// Dimension along a given axis
+    ssize_t shape(ssize_t dim) const {
+        if (dim >= ndim())
+            fail_dim_check(dim, "invalid axis");
+        return shape()[dim];
+    }
+
+    /// Strides of the array
+    const ssize_t* strides() const {
+        return detail::array_proxy(m_ptr)->strides;
+    }
+
+    /// Stride along a given axis
+    ssize_t strides(ssize_t dim) const {
+        if (dim >= ndim())
+            fail_dim_check(dim, "invalid axis");
+        return strides()[dim];
+    }
+
+    /// Return the NumPy array flags
+    int flags() const {
+        return detail::array_proxy(m_ptr)->flags;
+    }
+
+    /// If set, the array is writeable (otherwise the buffer is read-only)
+    bool writeable() const {
+        return detail::check_flags(m_ptr, detail::npy_api::NPY_ARRAY_WRITEABLE_);
+    }
+
+    /// If set, the array owns the data (will be freed when the array is deleted)
+    bool owndata() const {
+        return detail::check_flags(m_ptr, detail::npy_api::NPY_ARRAY_OWNDATA_);
+    }
+
+    /// Pointer to the contained data. If index is not provided, points to the
+    /// beginning of the buffer. May throw if the index would lead to out of bounds access.
+    template<typename... Ix> const void* data(Ix... index) const {
+        return static_cast<const void *>(detail::array_proxy(m_ptr)->data + offset_at(index...));
+    }
+
+    /// Mutable pointer to the contained data. If index is not provided, points to the
+    /// beginning of the buffer. May throw if the index would lead to out of bounds access.
+    /// May throw if the array is not writeable.
+    template<typename... Ix> void* mutable_data(Ix... index) {
+        check_writeable();
+        return static_cast<void *>(detail::array_proxy(m_ptr)->data + offset_at(index...));
+    }
+
+    /// Byte offset from beginning of the array to a given index (full or partial).
+    /// May throw if the index would lead to out of bounds access.
+    template<typename... Ix> ssize_t offset_at(Ix... index) const {
+        if ((ssize_t) sizeof...(index) > ndim())
+            fail_dim_check(sizeof...(index), "too many indices for an array");
+        return byte_offset(ssize_t(index)...);
+    }
+
+    ssize_t offset_at() const { return 0; }
+
+    /// Item count from beginning of the array to a given index (full or partial).
+    /// May throw if the index would lead to out of bounds access.
+    template<typename... Ix> ssize_t index_at(Ix... index) const {
+        return offset_at(index...) / itemsize();
+    }
+
+    /**
+     * Returns a proxy object that provides access to the array's data without bounds or
+     * dimensionality checking.  Will throw if the array is missing the `writeable` flag.  Use with
+     * care: the array must not be destroyed or reshaped for the duration of the returned object,
+     * and the caller must take care not to access invalid dimensions or dimension indices.
+     */
+    template <typename T, ssize_t Dims = -1> detail::unchecked_mutable_reference<T, Dims> mutable_unchecked() & {
+        if (Dims >= 0 && ndim() != Dims)
+            throw std::domain_error("array has incorrect number of dimensions: " + std::to_string(ndim()) +
+                    "; expected " + std::to_string(Dims));
+        return detail::unchecked_mutable_reference<T, Dims>(mutable_data(), shape(), strides(), ndim());
+    }
+
+    /**
+     * Returns a proxy object that provides const access to the array's data without bounds or
+     * dimensionality checking.  Unlike `mutable_unchecked()`, this does not require that the
+     * underlying array have the `writable` flag.  Use with care: the array must not be destroyed or
+     * reshaped for the duration of the returned object, and the caller must take care not to access
+     * invalid dimensions or dimension indices.
+     */
+    template <typename T, ssize_t Dims = -1> detail::unchecked_reference<T, Dims> unchecked() const & {
+        if (Dims >= 0 && ndim() != Dims)
+            throw std::domain_error("array has incorrect number of dimensions: " + std::to_string(ndim()) +
+                    "; expected " + std::to_string(Dims));
+        return detail::unchecked_reference<T, Dims>(data(), shape(), strides(), ndim());
+    }
+
+    /// Return a new view with all of the dimensions of length 1 removed
+    array squeeze() {
+        auto& api = detail::npy_api::get();
+        return reinterpret_steal<array>(api.PyArray_Squeeze_(m_ptr));
+    }
+
+    /// Resize array to given shape
+    /// If refcheck is true and more that one reference exist to this array
+    /// then resize will succeed only if it makes a reshape, i.e. original size doesn't change
+    void resize(ShapeContainer new_shape, bool refcheck = true) {
+        detail::npy_api::PyArray_Dims d = {
+            new_shape->data(), int(new_shape->size())
+        };
+        // try to resize, set ordering param to -1 cause it's not used anyway
+        object new_array = reinterpret_steal<object>(
+            detail::npy_api::get().PyArray_Resize_(m_ptr, &d, int(refcheck), -1)
+        );
+        if (!new_array) throw error_already_set();
+        if (isinstance<array>(new_array)) { *this = std::move(new_array); }
+    }
+
+    /// Ensure that the argument is a NumPy array
+    /// In case of an error, nullptr is returned and the Python error is cleared.
+    static array ensure(handle h, int ExtraFlags = 0) {
+        auto result = reinterpret_steal<array>(raw_array(h.ptr(), ExtraFlags));
+        if (!result)
+            PyErr_Clear();
+        return result;
+    }
+
+protected:
+    template<typename, typename> friend struct detail::npy_format_descriptor;
+
+    void fail_dim_check(ssize_t dim, const std::string& msg) const {
+        throw index_error(msg + ": " + std::to_string(dim) +
+                          " (ndim = " + std::to_string(ndim()) + ")");
+    }
+
+    template<typename... Ix> ssize_t byte_offset(Ix... index) const {
+        check_dimensions(index...);
+        return detail::byte_offset_unsafe(strides(), ssize_t(index)...);
+    }
+
+    void check_writeable() const {
+        if (!writeable())
+            throw std::domain_error("array is not writeable");
+    }
+
+    // Default, C-style strides
+    static std::vector<ssize_t> c_strides(const std::vector<ssize_t> &shape, ssize_t itemsize) {
+        auto ndim = shape.size();
+        std::vector<ssize_t> strides(ndim, itemsize);
+        for (size_t i = ndim - 1; i > 0; --i)
+            strides[i - 1] = strides[i] * shape[i];
+        return strides;
+    }
+
+    // F-style strides; default when constructing an array_t with `ExtraFlags & f_style`
+    static std::vector<ssize_t> f_strides(const std::vector<ssize_t> &shape, ssize_t itemsize) {
+        auto ndim = shape.size();
+        std::vector<ssize_t> strides(ndim, itemsize);
+        for (size_t i = 1; i < ndim; ++i)
+            strides[i] = strides[i - 1] * shape[i - 1];
+        return strides;
+    }
+
+    template<typename... Ix> void check_dimensions(Ix... index) const {
+        check_dimensions_impl(ssize_t(0), shape(), ssize_t(index)...);
+    }
+
+    void check_dimensions_impl(ssize_t, const ssize_t*) const { }
+
+    template<typename... Ix> void check_dimensions_impl(ssize_t axis, const ssize_t* shape, ssize_t i, Ix... index) const {
+        if (i >= *shape) {
+            throw index_error(std::string("index ") + std::to_string(i) +
+                              " is out of bounds for axis " + std::to_string(axis) +
+                              " with size " + std::to_string(*shape));
+        }
+        check_dimensions_impl(axis + 1, shape + 1, index...);
+    }
+
+    /// Create array from any object -- always returns a new reference
+    static PyObject *raw_array(PyObject *ptr, int ExtraFlags = 0) {
+        if (ptr == nullptr) {
+            PyErr_SetString(PyExc_ValueError, "cannot create a pybind11::array from a nullptr");
+            return nullptr;
+        }
+        return detail::npy_api::get().PyArray_FromAny_(
+            ptr, nullptr, 0, 0, detail::npy_api::NPY_ARRAY_ENSUREARRAY_ | ExtraFlags, nullptr);
+    }
+};
+
+template <typename T, int ExtraFlags = array::forcecast> class array_t : public array {
+private:
+    struct private_ctor {};
+    // Delegating constructor needed when both moving and accessing in the same constructor
+    array_t(private_ctor, ShapeContainer &&shape, StridesContainer &&strides, const T *ptr, handle base)
+        : array(std::move(shape), std::move(strides), ptr, base) {}
+public:
+    static_assert(!detail::array_info<T>::is_array, "Array types cannot be used with array_t");
+
+    using value_type = T;
+
+    array_t() : array(0, static_cast<const T *>(nullptr)) {}
+    array_t(handle h, borrowed_t) : array(h, borrowed_t{}) { }
+    array_t(handle h, stolen_t) : array(h, stolen_t{}) { }
+
+    PYBIND11_DEPRECATED("Use array_t<T>::ensure() instead")
+    array_t(handle h, bool is_borrowed) : array(raw_array_t(h.ptr()), stolen_t{}) {
+        if (!m_ptr) PyErr_Clear();
+        if (!is_borrowed) Py_XDECREF(h.ptr());
+    }
+
+    array_t(const object &o) : array(raw_array_t(o.ptr()), stolen_t{}) {
+        if (!m_ptr) throw error_already_set();
+    }
+
+    explicit array_t(const buffer_info& info) : array(info) { }
+
+    array_t(ShapeContainer shape, StridesContainer strides, const T *ptr = nullptr, handle base = handle())
+        : array(std::move(shape), std::move(strides), ptr, base) { }
+
+    explicit array_t(ShapeContainer shape, const T *ptr = nullptr, handle base = handle())
+        : array_t(private_ctor{}, std::move(shape),
+                ExtraFlags & f_style ? f_strides(*shape, itemsize()) : c_strides(*shape, itemsize()),
+                ptr, base) { }
+
+    explicit array_t(size_t count, const T *ptr = nullptr, handle base = handle())
+        : array({count}, {}, ptr, base) { }
+
+    constexpr ssize_t itemsize() const {
+        return sizeof(T);
+    }
+
+    template<typename... Ix> ssize_t index_at(Ix... index) const {
+        return offset_at(index...) / itemsize();
+    }
+
+    template<typename... Ix> const T* data(Ix... index) const {
+        return static_cast<const T*>(array::data(index...));
+    }
+
+    template<typename... Ix> T* mutable_data(Ix... index) {
+        return static_cast<T*>(array::mutable_data(index...));
+    }
+
+    // Reference to element at a given index
+    template<typename... Ix> const T& at(Ix... index) const {
+        if (sizeof...(index) != ndim())
+            fail_dim_check(sizeof...(index), "index dimension mismatch");
+        return *(static_cast<const T*>(array::data()) + byte_offset(ssize_t(index)...) / itemsize());
+    }
+
+    // Mutable reference to element at a given index
+    template<typename... Ix> T& mutable_at(Ix... index) {
+        if (sizeof...(index) != ndim())
+            fail_dim_check(sizeof...(index), "index dimension mismatch");
+        return *(static_cast<T*>(array::mutable_data()) + byte_offset(ssize_t(index)...) / itemsize());
+    }
+
+    /**
+     * Returns a proxy object that provides access to the array's data without bounds or
+     * dimensionality checking.  Will throw if the array is missing the `writeable` flag.  Use with
+     * care: the array must not be destroyed or reshaped for the duration of the returned object,
+     * and the caller must take care not to access invalid dimensions or dimension indices.
+     */
+    template <ssize_t Dims = -1> detail::unchecked_mutable_reference<T, Dims> mutable_unchecked() & {
+        return array::mutable_unchecked<T, Dims>();
+    }
+
+    /**
+     * Returns a proxy object that provides const access to the array's data without bounds or
+     * dimensionality checking.  Unlike `unchecked()`, this does not require that the underlying
+     * array have the `writable` flag.  Use with care: the array must not be destroyed or reshaped
+     * for the duration of the returned object, and the caller must take care not to access invalid
+     * dimensions or dimension indices.
+     */
+    template <ssize_t Dims = -1> detail::unchecked_reference<T, Dims> unchecked() const & {
+        return array::unchecked<T, Dims>();
+    }
+
+    /// Ensure that the argument is a NumPy array of the correct dtype (and if not, try to convert
+    /// it).  In case of an error, nullptr is returned and the Python error is cleared.
+    static array_t ensure(handle h) {
+        auto result = reinterpret_steal<array_t>(raw_array_t(h.ptr()));
+        if (!result)
+            PyErr_Clear();
+        return result;
+    }
+
+    static bool check_(handle h) {
+        const auto &api = detail::npy_api::get();
+        return api.PyArray_Check_(h.ptr())
+               && api.PyArray_EquivTypes_(detail::array_proxy(h.ptr())->descr, dtype::of<T>().ptr());
+    }
+
+protected:
+    /// Create array from any object -- always returns a new reference
+    static PyObject *raw_array_t(PyObject *ptr) {
+        if (ptr == nullptr) {
+            PyErr_SetString(PyExc_ValueError, "cannot create a pybind11::array_t from a nullptr");
+            return nullptr;
+        }
+        return detail::npy_api::get().PyArray_FromAny_(
+            ptr, dtype::of<T>().release().ptr(), 0, 0,
+            detail::npy_api::NPY_ARRAY_ENSUREARRAY_ | ExtraFlags, nullptr);
+    }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<detail::is_pod_struct<T>::value>> {
+    static std::string format() {
+        return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::format();
+    }
+};
+
+template <size_t N> struct format_descriptor<char[N]> {
+    static std::string format() { return std::to_string(N) + "s"; }
+};
+template <size_t N> struct format_descriptor<std::array<char, N>> {
+    static std::string format() { return std::to_string(N) + "s"; }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<std::is_enum<T>::value>> {
+    static std::string format() {
+        return format_descriptor<
+            typename std::remove_cv<typename std::underlying_type<T>::type>::type>::format();
+    }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<detail::array_info<T>::is_array>> {
+    static std::string format() {
+        using namespace detail;
+        PYBIND11_DESCR extents = _("(") + array_info<T>::extents() + _(")");
+        return extents.text() + format_descriptor<remove_all_extents_t<T>>::format();
+    }
+};
+
+NAMESPACE_BEGIN(detail)
+template <typename T, int ExtraFlags>
+struct pyobject_caster<array_t<T, ExtraFlags>> {
+    using type = array_t<T, ExtraFlags>;
+
+    bool load(handle src, bool convert) {
+        if (!convert && !type::check_(src))
+            return false;
+        value = type::ensure(src);
+        return static_cast<bool>(value);
+    }
+
+    static handle cast(const handle &src, return_value_policy /* policy */, handle /* parent */) {
+        return src.inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(type, handle_type_name<type>::name());
+};
+
+template <typename T>
+struct compare_buffer_info<T, detail::enable_if_t<detail::is_pod_struct<T>::value>> {
+    static bool compare(const buffer_info& b) {
+        return npy_api::get().PyArray_EquivTypes_(dtype::of<T>().ptr(), dtype(b).ptr());
+    }
+};
+
+template <typename T> struct npy_format_descriptor<T, enable_if_t<satisfies_any_of<T, std::is_arithmetic, is_complex>::value>> {
+private:
+    // NB: the order here must match the one in common.h
+    constexpr static const int values[15] = {
+        npy_api::NPY_BOOL_,
+        npy_api::NPY_BYTE_,   npy_api::NPY_UBYTE_,   npy_api::NPY_SHORT_,    npy_api::NPY_USHORT_,
+        npy_api::NPY_INT_,    npy_api::NPY_UINT_,    npy_api::NPY_LONGLONG_, npy_api::NPY_ULONGLONG_,
+        npy_api::NPY_FLOAT_,  npy_api::NPY_DOUBLE_,  npy_api::NPY_LONGDOUBLE_,
+        npy_api::NPY_CFLOAT_, npy_api::NPY_CDOUBLE_, npy_api::NPY_CLONGDOUBLE_
+    };
+
+public:
+    static constexpr int value = values[detail::is_fmt_numeric<T>::index];
+
+    static pybind11::dtype dtype() {
+        if (auto ptr = npy_api::get().PyArray_DescrFromType_(value))
+            return reinterpret_borrow<pybind11::dtype>(ptr);
+        pybind11_fail("Unsupported buffer format!");
+    }
+    template <typename T2 = T, enable_if_t<std::is_integral<T2>::value, int> = 0>
+    static PYBIND11_DESCR name() {
+        return _<std::is_same<T, bool>::value>(_("bool"),
+            _<std::is_signed<T>::value>("int", "uint") + _<sizeof(T)*8>());
+    }
+    template <typename T2 = T, enable_if_t<std::is_floating_point<T2>::value, int> = 0>
+    static PYBIND11_DESCR name() {
+        return _<std::is_same<T, float>::value || std::is_same<T, double>::value>(
+                _("float") + _<sizeof(T)*8>(), _("longdouble"));
+    }
+    template <typename T2 = T, enable_if_t<is_complex<T2>::value, int> = 0>
+    static PYBIND11_DESCR name() {
+        return _<std::is_same<typename T2::value_type, float>::value || std::is_same<typename T2::value_type, double>::value>(
+                _("complex") + _<sizeof(typename T2::value_type)*16>(), _("longcomplex"));
+    }
+};
+
+#define PYBIND11_DECL_CHAR_FMT \
+    static PYBIND11_DESCR name() { return _("S") + _<N>(); } \
+    static pybind11::dtype dtype() { return pybind11::dtype(std::string("S") + std::to_string(N)); }
+template <size_t N> struct npy_format_descriptor<char[N]> { PYBIND11_DECL_CHAR_FMT };
+template <size_t N> struct npy_format_descriptor<std::array<char, N>> { PYBIND11_DECL_CHAR_FMT };
+#undef PYBIND11_DECL_CHAR_FMT
+
+template<typename T> struct npy_format_descriptor<T, enable_if_t<array_info<T>::is_array>> {
+private:
+    using base_descr = npy_format_descriptor<typename array_info<T>::type>;
+public:
+    static_assert(!array_info<T>::is_empty, "Zero-sized arrays are not supported");
+
+    static PYBIND11_DESCR name() { return _("(") + array_info<T>::extents() + _(")") + base_descr::name(); }
+    static pybind11::dtype dtype() {
+        list shape;
+        array_info<T>::append_extents(shape);
+        return pybind11::dtype::from_args(pybind11::make_tuple(base_descr::dtype(), shape));
+    }
+};
+
+template<typename T> struct npy_format_descriptor<T, enable_if_t<std::is_enum<T>::value>> {
+private:
+    using base_descr = npy_format_descriptor<typename std::underlying_type<T>::type>;
+public:
+    static PYBIND11_DESCR name() { return base_descr::name(); }
+    static pybind11::dtype dtype() { return base_descr::dtype(); }
+};
+
+struct field_descriptor {
+    const char *name;
+    ssize_t offset;
+    ssize_t size;
+    std::string format;
+    dtype descr;
+};
+
+inline PYBIND11_NOINLINE void register_structured_dtype(
+    const std::initializer_list<field_descriptor>& fields,
+    const std::type_info& tinfo, ssize_t itemsize,
+    bool (*direct_converter)(PyObject *, void *&)) {
+
+    auto& numpy_internals = get_numpy_internals();
+    if (numpy_internals.get_type_info(tinfo, false))
+        pybind11_fail("NumPy: dtype is already registered");
+
+    list names, formats, offsets;
+    for (auto field : fields) {
+        if (!field.descr)
+            pybind11_fail(std::string("NumPy: unsupported field dtype: `") +
+                            field.name + "` @ " + tinfo.name());
+        names.append(PYBIND11_STR_TYPE(field.name));
+        formats.append(field.descr);
+        offsets.append(pybind11::int_(field.offset));
+    }
+    auto dtype_ptr = pybind11::dtype(names, formats, offsets, itemsize).release().ptr();
+
+    // There is an existing bug in NumPy (as of v1.11): trailing bytes are
+    // not encoded explicitly into the format string. This will supposedly
+    // get fixed in v1.12; for further details, see these:
+    // - https://github.com/numpy/numpy/issues/7797
+    // - https://github.com/numpy/numpy/pull/7798
+    // Because of this, we won't use numpy's logic to generate buffer format
+    // strings and will just do it ourselves.
+    std::vector<field_descriptor> ordered_fields(fields);
+    std::sort(ordered_fields.begin(), ordered_fields.end(),
+        [](const field_descriptor &a, const field_descriptor &b) { return a.offset < b.offset; });
+    ssize_t offset = 0;
+    std::ostringstream oss;
+    // mark the structure as unaligned with '^', because numpy and C++ don't
+    // always agree about alignment (particularly for complex), and we're
+    // explicitly listing all our padding. This depends on none of the fields
+    // overriding the endianness. Putting the ^ in front of individual fields
+    // isn't guaranteed to work due to https://github.com/numpy/numpy/issues/9049
+    oss << "^T{";
+    for (auto& field : ordered_fields) {
+        if (field.offset > offset)
+            oss << (field.offset - offset) << 'x';
+        oss << field.format << ':' << field.name << ':';
+        offset = field.offset + field.size;
+    }
+    if (itemsize > offset)
+        oss << (itemsize - offset) << 'x';
+    oss << '}';
+    auto format_str = oss.str();
+
+    // Sanity check: verify that NumPy properly parses our buffer format string
+    auto& api = npy_api::get();
+    auto arr =  array(buffer_info(nullptr, itemsize, format_str, 1));
+    if (!api.PyArray_EquivTypes_(dtype_ptr, arr.dtype().ptr()))
+        pybind11_fail("NumPy: invalid buffer descriptor!");
+
+    auto tindex = std::type_index(tinfo);
+    numpy_internals.registered_dtypes[tindex] = { dtype_ptr, format_str };
+    get_internals().direct_conversions[tindex].push_back(direct_converter);
+}
+
+template <typename T, typename SFINAE> struct npy_format_descriptor {
+    static_assert(is_pod_struct<T>::value, "Attempt to use a non-POD or unimplemented POD type as a numpy dtype");
+
+    static PYBIND11_DESCR name() { return make_caster<T>::name(); }
+
+    static pybind11::dtype dtype() {
+        return reinterpret_borrow<pybind11::dtype>(dtype_ptr());
+    }
+
+    static std::string format() {
+        static auto format_str = get_numpy_internals().get_type_info<T>(true)->format_str;
+        return format_str;
+    }
+
+    static void register_dtype(const std::initializer_list<field_descriptor>& fields) {
+        register_structured_dtype(fields, typeid(typename std::remove_cv<T>::type),
+                                  sizeof(T), &direct_converter);
+    }
+
+private:
+    static PyObject* dtype_ptr() {
+        static PyObject* ptr = get_numpy_internals().get_type_info<T>(true)->dtype_ptr;
+        return ptr;
+    }
+
+    static bool direct_converter(PyObject *obj, void*& value) {
+        auto& api = npy_api::get();
+        if (!PyObject_TypeCheck(obj, api.PyVoidArrType_Type_))
+            return false;
+        if (auto descr = reinterpret_steal<object>(api.PyArray_DescrFromScalar_(obj))) {
+            if (api.PyArray_EquivTypes_(dtype_ptr(), descr.ptr())) {
+                value = ((PyVoidScalarObject_Proxy *) obj)->obval;
+                return true;
+            }
+        }
+        return false;
+    }
+};
+
+#ifdef __CLION_IDE__ // replace heavy macro with dummy code for the IDE (doesn't affect code)
+# define PYBIND11_NUMPY_DTYPE(Type, ...) ((void)0)
+# define PYBIND11_NUMPY_DTYPE_EX(Type, ...) ((void)0)
+#else
+
+#define PYBIND11_FIELD_DESCRIPTOR_EX(T, Field, Name)                                          \
+    ::pybind11::detail::field_descriptor {                                                    \
+        Name, offsetof(T, Field), sizeof(decltype(std::declval<T>().Field)),                  \
+        ::pybind11::format_descriptor<decltype(std::declval<T>().Field)>::format(),           \
+        ::pybind11::detail::npy_format_descriptor<decltype(std::declval<T>().Field)>::dtype() \
+    }
+
+// Extract name, offset and format descriptor for a struct field
+#define PYBIND11_FIELD_DESCRIPTOR(T, Field) PYBIND11_FIELD_DESCRIPTOR_EX(T, Field, #Field)
+
+// The main idea of this macro is borrowed from https://github.com/swansontec/map-macro
+// (C) William Swanson, Paul Fultz
+#define PYBIND11_EVAL0(...) __VA_ARGS__
+#define PYBIND11_EVAL1(...) PYBIND11_EVAL0 (PYBIND11_EVAL0 (PYBIND11_EVAL0 (__VA_ARGS__)))
+#define PYBIND11_EVAL2(...) PYBIND11_EVAL1 (PYBIND11_EVAL1 (PYBIND11_EVAL1 (__VA_ARGS__)))
+#define PYBIND11_EVAL3(...) PYBIND11_EVAL2 (PYBIND11_EVAL2 (PYBIND11_EVAL2 (__VA_ARGS__)))
+#define PYBIND11_EVAL4(...) PYBIND11_EVAL3 (PYBIND11_EVAL3 (PYBIND11_EVAL3 (__VA_ARGS__)))
+#define PYBIND11_EVAL(...)  PYBIND11_EVAL4 (PYBIND11_EVAL4 (PYBIND11_EVAL4 (__VA_ARGS__)))
+#define PYBIND11_MAP_END(...)
+#define PYBIND11_MAP_OUT
+#define PYBIND11_MAP_COMMA ,
+#define PYBIND11_MAP_GET_END() 0, PYBIND11_MAP_END
+#define PYBIND11_MAP_NEXT0(test, next, ...) next PYBIND11_MAP_OUT
+#define PYBIND11_MAP_NEXT1(test, next) PYBIND11_MAP_NEXT0 (test, next, 0)
+#define PYBIND11_MAP_NEXT(test, next)  PYBIND11_MAP_NEXT1 (PYBIND11_MAP_GET_END test, next)
+#ifdef _MSC_VER // MSVC is not as eager to expand macros, hence this workaround
+#define PYBIND11_MAP_LIST_NEXT1(test, next) \
+    PYBIND11_EVAL0 (PYBIND11_MAP_NEXT0 (test, PYBIND11_MAP_COMMA next, 0))
+#else
+#define PYBIND11_MAP_LIST_NEXT1(test, next) \
+    PYBIND11_MAP_NEXT0 (test, PYBIND11_MAP_COMMA next, 0)
+#endif
+#define PYBIND11_MAP_LIST_NEXT(test, next) \
+    PYBIND11_MAP_LIST_NEXT1 (PYBIND11_MAP_GET_END test, next)
+#define PYBIND11_MAP_LIST0(f, t, x, peek, ...) \
+    f(t, x) PYBIND11_MAP_LIST_NEXT (peek, PYBIND11_MAP_LIST1) (f, t, peek, __VA_ARGS__)
+#define PYBIND11_MAP_LIST1(f, t, x, peek, ...) \
+    f(t, x) PYBIND11_MAP_LIST_NEXT (peek, PYBIND11_MAP_LIST0) (f, t, peek, __VA_ARGS__)
+// PYBIND11_MAP_LIST(f, t, a1, a2, ...) expands to f(t, a1), f(t, a2), ...
+#define PYBIND11_MAP_LIST(f, t, ...) \
+    PYBIND11_EVAL (PYBIND11_MAP_LIST1 (f, t, __VA_ARGS__, (), 0))
+
+#define PYBIND11_NUMPY_DTYPE(Type, ...) \
+    ::pybind11::detail::npy_format_descriptor<Type>::register_dtype \
+        ({PYBIND11_MAP_LIST (PYBIND11_FIELD_DESCRIPTOR, Type, __VA_ARGS__)})
+
+#ifdef _MSC_VER
+#define PYBIND11_MAP2_LIST_NEXT1(test, next) \
+    PYBIND11_EVAL0 (PYBIND11_MAP_NEXT0 (test, PYBIND11_MAP_COMMA next, 0))
+#else
+#define PYBIND11_MAP2_LIST_NEXT1(test, next) \
+    PYBIND11_MAP_NEXT0 (test, PYBIND11_MAP_COMMA next, 0)
+#endif
+#define PYBIND11_MAP2_LIST_NEXT(test, next) \
+    PYBIND11_MAP2_LIST_NEXT1 (PYBIND11_MAP_GET_END test, next)
+#define PYBIND11_MAP2_LIST0(f, t, x1, x2, peek, ...) \
+    f(t, x1, x2) PYBIND11_MAP2_LIST_NEXT (peek, PYBIND11_MAP2_LIST1) (f, t, peek, __VA_ARGS__)
+#define PYBIND11_MAP2_LIST1(f, t, x1, x2, peek, ...) \
+    f(t, x1, x2) PYBIND11_MAP2_LIST_NEXT (peek, PYBIND11_MAP2_LIST0) (f, t, peek, __VA_ARGS__)
+// PYBIND11_MAP2_LIST(f, t, a1, a2, ...) expands to f(t, a1, a2), f(t, a3, a4), ...
+#define PYBIND11_MAP2_LIST(f, t, ...) \
+    PYBIND11_EVAL (PYBIND11_MAP2_LIST1 (f, t, __VA_ARGS__, (), 0))
+
+#define PYBIND11_NUMPY_DTYPE_EX(Type, ...) \
+    ::pybind11::detail::npy_format_descriptor<Type>::register_dtype \
+        ({PYBIND11_MAP2_LIST (PYBIND11_FIELD_DESCRIPTOR_EX, Type, __VA_ARGS__)})
+
+#endif // __CLION_IDE__
+
+template  <class T>
+using array_iterator = typename std::add_pointer<T>::type;
+
+template <class T>
+array_iterator<T> array_begin(const buffer_info& buffer) {
+    return array_iterator<T>(reinterpret_cast<T*>(buffer.ptr));
+}
+
+template <class T>
+array_iterator<T> array_end(const buffer_info& buffer) {
+    return array_iterator<T>(reinterpret_cast<T*>(buffer.ptr) + buffer.size);
+}
+
+class common_iterator {
+public:
+    using container_type = std::vector<ssize_t>;
+    using value_type = container_type::value_type;
+    using size_type = container_type::size_type;
+
+    common_iterator() : p_ptr(0), m_strides() {}
+
+    common_iterator(void* ptr, const container_type& strides, const container_type& shape)
+        : p_ptr(reinterpret_cast<char*>(ptr)), m_strides(strides.size()) {
+        m_strides.back() = static_cast<value_type>(strides.back());
+        for (size_type i = m_strides.size() - 1; i != 0; --i) {
+            size_type j = i - 1;
+            value_type s = static_cast<value_type>(shape[i]);
+            m_strides[j] = strides[j] + m_strides[i] - strides[i] * s;
+        }
+    }
+
+    void increment(size_type dim) {
+        p_ptr += m_strides[dim];
+    }
+
+    void* data() const {
+        return p_ptr;
+    }
+
+private:
+    char* p_ptr;
+    container_type m_strides;
+};
+
+template <size_t N> class multi_array_iterator {
+public:
+    using container_type = std::vector<ssize_t>;
+
+    multi_array_iterator(const std::array<buffer_info, N> &buffers,
+                         const container_type &shape)
+        : m_shape(shape.size()), m_index(shape.size(), 0),
+          m_common_iterator() {
+
+        // Manual copy to avoid conversion warning if using std::copy
+        for (size_t i = 0; i < shape.size(); ++i)
+            m_shape[i] = shape[i];
+
+        container_type strides(shape.size());
+        for (size_t i = 0; i < N; ++i)
+            init_common_iterator(buffers[i], shape, m_common_iterator[i], strides);
+    }
+
+    multi_array_iterator& operator++() {
+        for (size_t j = m_index.size(); j != 0; --j) {
+            size_t i = j - 1;
+            if (++m_index[i] != m_shape[i]) {
+                increment_common_iterator(i);
+                break;
+            } else {
+                m_index[i] = 0;
+            }
+        }
+        return *this;
+    }
+
+    template <size_t K, class T = void> T* data() const {
+        return reinterpret_cast<T*>(m_common_iterator[K].data());
+    }
+
+private:
+
+    using common_iter = common_iterator;
+
+    void init_common_iterator(const buffer_info &buffer,
+                              const container_type &shape,
+                              common_iter &iterator,
+                              container_type &strides) {
+        auto buffer_shape_iter = buffer.shape.rbegin();
+        auto buffer_strides_iter = buffer.strides.rbegin();
+        auto shape_iter = shape.rbegin();
+        auto strides_iter = strides.rbegin();
+
+        while (buffer_shape_iter != buffer.shape.rend()) {
+            if (*shape_iter == *buffer_shape_iter)
+                *strides_iter = *buffer_strides_iter;
+            else
+                *strides_iter = 0;
+
+            ++buffer_shape_iter;
+            ++buffer_strides_iter;
+            ++shape_iter;
+            ++strides_iter;
+        }
+
+        std::fill(strides_iter, strides.rend(), 0);
+        iterator = common_iter(buffer.ptr, strides, shape);
+    }
+
+    void increment_common_iterator(size_t dim) {
+        for (auto &iter : m_common_iterator)
+            iter.increment(dim);
+    }
+
+    container_type m_shape;
+    container_type m_index;
+    std::array<common_iter, N> m_common_iterator;
+};
+
+enum class broadcast_trivial { non_trivial, c_trivial, f_trivial };
+
+// Populates the shape and number of dimensions for the set of buffers.  Returns a broadcast_trivial
+// enum value indicating whether the broadcast is "trivial"--that is, has each buffer being either a
+// singleton or a full-size, C-contiguous (`c_trivial`) or Fortran-contiguous (`f_trivial`) storage
+// buffer; returns `non_trivial` otherwise.
+template <size_t N>
+broadcast_trivial broadcast(const std::array<buffer_info, N> &buffers, ssize_t &ndim, std::vector<ssize_t> &shape) {
+    ndim = std::accumulate(buffers.begin(), buffers.end(), ssize_t(0), [](ssize_t res, const buffer_info &buf) {
+        return std::max(res, buf.ndim);
+    });
+
+    shape.clear();
+    shape.resize((size_t) ndim, 1);
+
+    // Figure out the output size, and make sure all input arrays conform (i.e. are either size 1 or
+    // the full size).
+    for (size_t i = 0; i < N; ++i) {
+        auto res_iter = shape.rbegin();
+        auto end = buffers[i].shape.rend();
+        for (auto shape_iter = buffers[i].shape.rbegin(); shape_iter != end; ++shape_iter, ++res_iter) {
+            const auto &dim_size_in = *shape_iter;
+            auto &dim_size_out = *res_iter;
+
+            // Each input dimension can either be 1 or `n`, but `n` values must match across buffers
+            if (dim_size_out == 1)
+                dim_size_out = dim_size_in;
+            else if (dim_size_in != 1 && dim_size_in != dim_size_out)
+                pybind11_fail("pybind11::vectorize: incompatible size/dimension of inputs!");
+        }
+    }
+
+    bool trivial_broadcast_c = true;
+    bool trivial_broadcast_f = true;
+    for (size_t i = 0; i < N && (trivial_broadcast_c || trivial_broadcast_f); ++i) {
+        if (buffers[i].size == 1)
+            continue;
+
+        // Require the same number of dimensions:
+        if (buffers[i].ndim != ndim)
+            return broadcast_trivial::non_trivial;
+
+        // Require all dimensions be full-size:
+        if (!std::equal(buffers[i].shape.cbegin(), buffers[i].shape.cend(), shape.cbegin()))
+            return broadcast_trivial::non_trivial;
+
+        // Check for C contiguity (but only if previous inputs were also C contiguous)
+        if (trivial_broadcast_c) {
+            ssize_t expect_stride = buffers[i].itemsize;
+            auto end = buffers[i].shape.crend();
+            for (auto shape_iter = buffers[i].shape.crbegin(), stride_iter = buffers[i].strides.crbegin();
+                    trivial_broadcast_c && shape_iter != end; ++shape_iter, ++stride_iter) {
+                if (expect_stride == *stride_iter)
+                    expect_stride *= *shape_iter;
+                else
+                    trivial_broadcast_c = false;
+            }
+        }
+
+        // Check for Fortran contiguity (if previous inputs were also F contiguous)
+        if (trivial_broadcast_f) {
+            ssize_t expect_stride = buffers[i].itemsize;
+            auto end = buffers[i].shape.cend();
+            for (auto shape_iter = buffers[i].shape.cbegin(), stride_iter = buffers[i].strides.cbegin();
+                    trivial_broadcast_f && shape_iter != end; ++shape_iter, ++stride_iter) {
+                if (expect_stride == *stride_iter)
+                    expect_stride *= *shape_iter;
+                else
+                    trivial_broadcast_f = false;
+            }
+        }
+    }
+
+    return
+        trivial_broadcast_c ? broadcast_trivial::c_trivial :
+        trivial_broadcast_f ? broadcast_trivial::f_trivial :
+        broadcast_trivial::non_trivial;
+}
+
+template <typename T>
+struct vectorize_arg {
+    static_assert(!std::is_rvalue_reference<T>::value, "Functions with rvalue reference arguments cannot be vectorized");
+    // The wrapped function gets called with this type:
+    using call_type = remove_reference_t<T>;
+    // Is this a vectorized argument?
+    static constexpr bool vectorize =
+        satisfies_any_of<call_type, std::is_arithmetic, is_complex, std::is_pod>::value &&
+        satisfies_none_of<call_type, std::is_pointer, std::is_array, is_std_array, std::is_enum>::value &&
+        (!std::is_reference<T>::value ||
+         (std::is_lvalue_reference<T>::value && std::is_const<call_type>::value));
+    // Accept this type: an array for vectorized types, otherwise the type as-is:
+    using type = conditional_t<vectorize, array_t<remove_cv_t<call_type>, array::forcecast>, T>;
+};
+
+template <typename Func, typename Return, typename... Args>
+struct vectorize_helper {
+private:
+    static constexpr size_t N = sizeof...(Args);
+    static constexpr size_t NVectorized = constexpr_sum(vectorize_arg<Args>::vectorize...);
+    static_assert(NVectorized >= 1,
+            "pybind11::vectorize(...) requires a function with at least one vectorizable argument");
+
+public:
+    template <typename T>
+    explicit vectorize_helper(T &&f) : f(std::forward<T>(f)) { }
+
+    object operator()(typename vectorize_arg<Args>::type... args) {
+        return run(args...,
+                   make_index_sequence<N>(),
+                   select_indices<vectorize_arg<Args>::vectorize...>(),
+                   make_index_sequence<NVectorized>());
+    }
+
+private:
+    remove_reference_t<Func> f;
+
+    template <size_t Index> using param_n_t = typename pack_element<Index, typename vectorize_arg<Args>::call_type...>::type;
+
+    // Runs a vectorized function given arguments tuple and three index sequences:
+    //     - Index is the full set of 0 ... (N-1) argument indices;
+    //     - VIndex is the subset of argument indices with vectorized parameters, letting us access
+    //       vectorized arguments (anything not in this sequence is passed through)
+    //     - BIndex is a incremental sequence (beginning at 0) of the same size as VIndex, so that
+    //       we can store vectorized buffer_infos in an array (argument VIndex has its buffer at
+    //       index BIndex in the array).
+    template <size_t... Index, size_t... VIndex, size_t... BIndex> object run(
+            typename vectorize_arg<Args>::type &...args,
+            index_sequence<Index...> i_seq, index_sequence<VIndex...> vi_seq, index_sequence<BIndex...> bi_seq) {
+
+        // Pointers to values the function was called with; the vectorized ones set here will start
+        // out as array_t<T> pointers, but they will be changed them to T pointers before we make
+        // call the wrapped function.  Non-vectorized pointers are left as-is.
+        std::array<void *, N> params{{ &args... }};
+
+        // The array of `buffer_info`s of vectorized arguments:
+        std::array<buffer_info, NVectorized> buffers{{ reinterpret_cast<array *>(params[VIndex])->request()... }};
+
+        /* Determine dimensions parameters of output array */
+        ssize_t nd = 0;
+        std::vector<ssize_t> shape(0);
+        auto trivial = broadcast(buffers, nd, shape);
+        size_t ndim = (size_t) nd;
+
+        size_t size = std::accumulate(shape.begin(), shape.end(), (size_t) 1, std::multiplies<size_t>());
+
+        // If all arguments are 0-dimension arrays (i.e. single values) return a plain value (i.e.
+        // not wrapped in an array).
+        if (size == 1 && ndim == 0) {
+            PYBIND11_EXPAND_SIDE_EFFECTS(params[VIndex] = buffers[BIndex].ptr);
+            return cast(f(*reinterpret_cast<param_n_t<Index> *>(params[Index])...));
+        }
+
+        array_t<Return> result;
+        if (trivial == broadcast_trivial::f_trivial) result = array_t<Return, array::f_style>(shape);
+        else result = array_t<Return>(shape);
+
+        if (size == 0) return result;
+
+        /* Call the function */
+        if (trivial == broadcast_trivial::non_trivial)
+            apply_broadcast(buffers, params, result, i_seq, vi_seq, bi_seq);
+        else
+            apply_trivial(buffers, params, result.mutable_data(), size, i_seq, vi_seq, bi_seq);
+
+        return result;
+    }
+
+    template <size_t... Index, size_t... VIndex, size_t... BIndex>
+    void apply_trivial(std::array<buffer_info, NVectorized> &buffers,
+                       std::array<void *, N> &params,
+                       Return *out,
+                       size_t size,
+                       index_sequence<Index...>, index_sequence<VIndex...>, index_sequence<BIndex...>) {
+
+        // Initialize an array of mutable byte references and sizes with references set to the
+        // appropriate pointer in `params`; as we iterate, we'll increment each pointer by its size
+        // (except for singletons, which get an increment of 0).
+        std::array<std::pair<unsigned char *&, const size_t>, NVectorized> vecparams{{
+            std::pair<unsigned char *&, const size_t>(
+                    reinterpret_cast<unsigned char *&>(params[VIndex] = buffers[BIndex].ptr),
+                    buffers[BIndex].size == 1 ? 0 : sizeof(param_n_t<VIndex>)
+            )...
+        }};
+
+        for (size_t i = 0; i < size; ++i) {
+            out[i] = f(*reinterpret_cast<param_n_t<Index> *>(params[Index])...);
+            for (auto &x : vecparams) x.first += x.second;
+        }
+    }
+
+    template <size_t... Index, size_t... VIndex, size_t... BIndex>
+    void apply_broadcast(std::array<buffer_info, NVectorized> &buffers,
+                         std::array<void *, N> &params,
+                         array_t<Return> &output_array,
+                         index_sequence<Index...>, index_sequence<VIndex...>, index_sequence<BIndex...>) {
+
+        buffer_info output = output_array.request();
+        multi_array_iterator<NVectorized> input_iter(buffers, output.shape);
+
+        for (array_iterator<Return> iter = array_begin<Return>(output), end = array_end<Return>(output);
+             iter != end;
+             ++iter, ++input_iter) {
+            PYBIND11_EXPAND_SIDE_EFFECTS((
+                params[VIndex] = input_iter.template data<BIndex>()
+            ));
+            *iter = f(*reinterpret_cast<param_n_t<Index> *>(std::get<Index>(params))...);
+        }
+    }
+};
+
+template <typename Func, typename Return, typename... Args>
+vectorize_helper<Func, Return, Args...>
+vectorize_extractor(const Func &f, Return (*) (Args ...)) {
+    return detail::vectorize_helper<Func, Return, Args...>(f);
+}
+
+template <typename T, int Flags> struct handle_type_name<array_t<T, Flags>> {
+    static PYBIND11_DESCR name() {
+        return _("numpy.ndarray[") + npy_format_descriptor<T>::name() + _("]");
+    }
+};
+
+NAMESPACE_END(detail)
+
+// Vanilla pointer vectorizer:
+template <typename Return, typename... Args>
+detail::vectorize_helper<Return (*)(Args...), Return, Args...>
+vectorize(Return (*f) (Args ...)) {
+    return detail::vectorize_helper<Return (*)(Args...), Return, Args...>(f);
+}
+
+// lambda vectorizer:
+template <typename Func, detail::enable_if_t<detail::is_lambda<Func>::value, int> = 0>
+auto vectorize(Func &&f) -> decltype(
+        detail::vectorize_extractor(std::forward<Func>(f), (detail::function_signature_t<Func> *) nullptr)) {
+    return detail::vectorize_extractor(std::forward<Func>(f), (detail::function_signature_t<Func> *) nullptr);
+}
+
+// Vectorize a class method (non-const):
+template <typename Return, typename Class, typename... Args,
+          typename Helper = detail::vectorize_helper<decltype(std::mem_fn(std::declval<Return (Class::*)(Args...)>())), Return, Class *, Args...>>
+Helper vectorize(Return (Class::*f)(Args...)) {
+    return Helper(std::mem_fn(f));
+}
+
+// Vectorize a class method (non-const):
+template <typename Return, typename Class, typename... Args,
+          typename Helper = detail::vectorize_helper<decltype(std::mem_fn(std::declval<Return (Class::*)(Args...) const>())), Return, const Class *, Args...>>
+Helper vectorize(Return (Class::*f)(Args...) const) {
+    return Helper(std::mem_fn(f));
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
+
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/operators.h b/ml/dlib/dlib/external/pybind11/include/pybind11/operators.h
new file mode 100644
index 000000000..b3dd62c3b
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/operators.h
@@ -0,0 +1,168 @@
+/*
+    pybind11/operator.h: Metatemplates for operator overloading
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#if defined(__clang__) && !defined(__INTEL_COMPILER)
+#  pragma clang diagnostic ignored "-Wunsequenced" // multiple unsequenced modifications to 'self' (when using def(py::self OP Type()))
+#elif defined(_MSC_VER)
+#  pragma warning(push)
+#  pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/// Enumeration with all supported operator types
+enum op_id : int {
+    op_add, op_sub, op_mul, op_div, op_mod, op_divmod, op_pow, op_lshift,
+    op_rshift, op_and, op_xor, op_or, op_neg, op_pos, op_abs, op_invert,
+    op_int, op_long, op_float, op_str, op_cmp, op_gt, op_ge, op_lt, op_le,
+    op_eq, op_ne, op_iadd, op_isub, op_imul, op_idiv, op_imod, op_ilshift,
+    op_irshift, op_iand, op_ixor, op_ior, op_complex, op_bool, op_nonzero,
+    op_repr, op_truediv, op_itruediv, op_hash
+};
+
+enum op_type : int {
+    op_l, /* base type on left */
+    op_r, /* base type on right */
+    op_u  /* unary operator */
+};
+
+struct self_t { };
+static const self_t self = self_t();
+
+/// Type for an unused type slot
+struct undefined_t { };
+
+/// Don't warn about an unused variable
+inline self_t __self() { return self; }
+
+/// base template of operator implementations
+template <op_id, op_type, typename B, typename L, typename R> struct op_impl { };
+
+/// Operator implementation generator
+template <op_id id, op_type ot, typename L, typename R> struct op_ {
+    template <typename Class, typename... Extra> void execute(Class &cl, const Extra&... extra) const {
+        using Base = typename Class::type;
+        using L_type = conditional_t<std::is_same<L, self_t>::value, Base, L>;
+        using R_type = conditional_t<std::is_same<R, self_t>::value, Base, R>;
+        using op = op_impl<id, ot, Base, L_type, R_type>;
+        cl.def(op::name(), &op::execute, is_operator(), extra...);
+        #if PY_MAJOR_VERSION < 3
+        if (id == op_truediv || id == op_itruediv)
+            cl.def(id == op_itruediv ? "__idiv__" : ot == op_l ? "__div__" : "__rdiv__",
+                    &op::execute, is_operator(), extra...);
+        #endif
+    }
+    template <typename Class, typename... Extra> void execute_cast(Class &cl, const Extra&... extra) const {
+        using Base = typename Class::type;
+        using L_type = conditional_t<std::is_same<L, self_t>::value, Base, L>;
+        using R_type = conditional_t<std::is_same<R, self_t>::value, Base, R>;
+        using op = op_impl<id, ot, Base, L_type, R_type>;
+        cl.def(op::name(), &op::execute_cast, is_operator(), extra...);
+        #if PY_MAJOR_VERSION < 3
+        if (id == op_truediv || id == op_itruediv)
+            cl.def(id == op_itruediv ? "__idiv__" : ot == op_l ? "__div__" : "__rdiv__",
+                    &op::execute, is_operator(), extra...);
+        #endif
+    }
+};
+
+#define PYBIND11_BINARY_OPERATOR(id, rid, op, expr)                                    \
+template <typename B, typename L, typename R> struct op_impl<op_##id, op_l, B, L, R> { \
+    static char const* name() { return "__" #id "__"; }                                \
+    static auto execute(const L &l, const R &r) -> decltype(expr) { return (expr); }   \
+    static B execute_cast(const L &l, const R &r) { return B(expr); }                  \
+};                                                                                     \
+template <typename B, typename L, typename R> struct op_impl<op_##id, op_r, B, L, R> { \
+    static char const* name() { return "__" #rid "__"; }                               \
+    static auto execute(const R &r, const L &l) -> decltype(expr) { return (expr); }   \
+    static B execute_cast(const R &r, const L &l) { return B(expr); }                  \
+};                                                                                     \
+inline op_<op_##id, op_l, self_t, self_t> op(const self_t &, const self_t &) {         \
+    return op_<op_##id, op_l, self_t, self_t>();                                       \
+}                                                                                      \
+template <typename T> op_<op_##id, op_l, self_t, T> op(const self_t &, const T &) {    \
+    return op_<op_##id, op_l, self_t, T>();                                            \
+}                                                                                      \
+template <typename T> op_<op_##id, op_r, T, self_t> op(const T &, const self_t &) {    \
+    return op_<op_##id, op_r, T, self_t>();                                            \
+}
+
+#define PYBIND11_INPLACE_OPERATOR(id, op, expr)                                        \
+template <typename B, typename L, typename R> struct op_impl<op_##id, op_l, B, L, R> { \
+    static char const* name() { return "__" #id "__"; }                                \
+    static auto execute(L &l, const R &r) -> decltype(expr) { return expr; }           \
+    static B execute_cast(L &l, const R &r) { return B(expr); }                        \
+};                                                                                     \
+template <typename T> op_<op_##id, op_l, self_t, T> op(const self_t &, const T &) {    \
+    return op_<op_##id, op_l, self_t, T>();                                            \
+}
+
+#define PYBIND11_UNARY_OPERATOR(id, op, expr)                                          \
+template <typename B, typename L> struct op_impl<op_##id, op_u, B, L, undefined_t> {   \
+    static char const* name() { return "__" #id "__"; }                                \
+    static auto execute(const L &l) -> decltype(expr) { return expr; }                 \
+    static B execute_cast(const L &l) { return B(expr); }                              \
+};                                                                                     \
+inline op_<op_##id, op_u, self_t, undefined_t> op(const self_t &) {                    \
+    return op_<op_##id, op_u, self_t, undefined_t>();                                  \
+}
+
+PYBIND11_BINARY_OPERATOR(sub,       rsub,         operator-,    l - r)
+PYBIND11_BINARY_OPERATOR(add,       radd,         operator+,    l + r)
+PYBIND11_BINARY_OPERATOR(mul,       rmul,         operator*,    l * r)
+PYBIND11_BINARY_OPERATOR(truediv,   rtruediv,     operator/,    l / r)
+PYBIND11_BINARY_OPERATOR(mod,       rmod,         operator%,    l % r)
+PYBIND11_BINARY_OPERATOR(lshift,    rlshift,      operator<<,   l << r)
+PYBIND11_BINARY_OPERATOR(rshift,    rrshift,      operator>>,   l >> r)
+PYBIND11_BINARY_OPERATOR(and,       rand,         operator&,    l & r)
+PYBIND11_BINARY_OPERATOR(xor,       rxor,         operator^,    l ^ r)
+PYBIND11_BINARY_OPERATOR(eq,        eq,           operator==,   l == r)
+PYBIND11_BINARY_OPERATOR(ne,        ne,           operator!=,   l != r)
+PYBIND11_BINARY_OPERATOR(or,        ror,          operator|,    l | r)
+PYBIND11_BINARY_OPERATOR(gt,        lt,           operator>,    l > r)
+PYBIND11_BINARY_OPERATOR(ge,        le,           operator>=,   l >= r)
+PYBIND11_BINARY_OPERATOR(lt,        gt,           operator<,    l < r)
+PYBIND11_BINARY_OPERATOR(le,        ge,           operator<=,   l <= r)
+//PYBIND11_BINARY_OPERATOR(pow,       rpow,         pow,          std::pow(l,  r))
+PYBIND11_INPLACE_OPERATOR(iadd,     operator+=,   l += r)
+PYBIND11_INPLACE_OPERATOR(isub,     operator-=,   l -= r)
+PYBIND11_INPLACE_OPERATOR(imul,     operator*=,   l *= r)
+PYBIND11_INPLACE_OPERATOR(itruediv, operator/=,   l /= r)
+PYBIND11_INPLACE_OPERATOR(imod,     operator%=,   l %= r)
+PYBIND11_INPLACE_OPERATOR(ilshift,  operator<<=,  l <<= r)
+PYBIND11_INPLACE_OPERATOR(irshift,  operator>>=,  l >>= r)
+PYBIND11_INPLACE_OPERATOR(iand,     operator&=,   l &= r)
+PYBIND11_INPLACE_OPERATOR(ixor,     operator^=,   l ^= r)
+PYBIND11_INPLACE_OPERATOR(ior,      operator|=,   l |= r)
+PYBIND11_UNARY_OPERATOR(neg,        operator-,    -l)
+PYBIND11_UNARY_OPERATOR(pos,        operator+,    +l)
+PYBIND11_UNARY_OPERATOR(abs,        abs,          std::abs(l))
+PYBIND11_UNARY_OPERATOR(hash,       hash,         std::hash<L>()(l))
+PYBIND11_UNARY_OPERATOR(invert,     operator~,    (~l))
+PYBIND11_UNARY_OPERATOR(bool,       operator!,    !!l)
+PYBIND11_UNARY_OPERATOR(int,        int_,         (int) l)
+PYBIND11_UNARY_OPERATOR(float,      float_,       (double) l)
+
+#undef PYBIND11_BINARY_OPERATOR
+#undef PYBIND11_INPLACE_OPERATOR
+#undef PYBIND11_UNARY_OPERATOR
+NAMESPACE_END(detail)
+
+using detail::self;
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
+
+#if defined(_MSC_VER)
+#  pragma warning(pop)
+#endif
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/options.h b/ml/dlib/dlib/external/pybind11/include/pybind11/options.h
new file mode 100644
index 000000000..cc1e1f6f0
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/options.h
@@ -0,0 +1,65 @@
+/*
+    pybind11/options.h: global settings that are configurable at runtime.
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+class options {
+public:
+
+    // Default RAII constructor, which leaves settings as they currently are.
+    options() : previous_state(global_state()) {}
+
+    // Class is non-copyable.
+    options(const options&) = delete;
+    options& operator=(const options&) = delete;
+
+    // Destructor, which restores settings that were in effect before.
+    ~options() {
+        global_state() = previous_state;
+    }
+
+    // Setter methods (affect the global state):
+
+    options& disable_user_defined_docstrings() & { global_state().show_user_defined_docstrings = false; return *this; }
+
+    options& enable_user_defined_docstrings() & { global_state().show_user_defined_docstrings = true; return *this; }
+
+    options& disable_function_signatures() & { global_state().show_function_signatures = false; return *this; }
+
+    options& enable_function_signatures() & { global_state().show_function_signatures = true; return *this; }
+
+    // Getter methods (return the global state):
+
+    static bool show_user_defined_docstrings() { return global_state().show_user_defined_docstrings; }
+
+    static bool show_function_signatures() { return global_state().show_function_signatures; }
+
+    // This type is not meant to be allocated on the heap.
+    void* operator new(size_t) = delete;
+
+private:
+
+    struct state {
+        bool show_user_defined_docstrings = true;  //< Include user-supplied texts in docstrings.
+        bool show_function_signatures = true;      //< Include auto-generated function signatures in docstrings.
+    };
+
+    static state &global_state() {
+        static state instance;
+        return instance;
+    }
+
+    state previous_state;
+};
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/pybind11.h b/ml/dlib/dlib/external/pybind11/include/pybind11/pybind11.h
new file mode 100644
index 000000000..7723d2a8e
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/pybind11.h
@@ -0,0 +1,1963 @@
+/*
+    pybind11/pybind11.h: Main header file of the C++11 python
+    binding generator library
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#if defined(_MSC_VER)
+#  pragma warning(push)
+#  pragma warning(disable: 4100) // warning C4100: Unreferenced formal parameter
+#  pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#  pragma warning(disable: 4512) // warning C4512: Assignment operator was implicitly defined as deleted
+#  pragma warning(disable: 4800) // warning C4800: 'int': forcing value to bool 'true' or 'false' (performance warning)
+#  pragma warning(disable: 4996) // warning C4996: The POSIX name for this item is deprecated. Instead, use the ISO C and C++ conformant name
+#  pragma warning(disable: 4702) // warning C4702: unreachable code
+#  pragma warning(disable: 4522) // warning C4522: multiple assignment operators specified
+#elif defined(__INTEL_COMPILER)
+#  pragma warning(push)
+#  pragma warning(disable: 68)    // integer conversion resulted in a change of sign
+#  pragma warning(disable: 186)   // pointless comparison of unsigned integer with zero
+#  pragma warning(disable: 878)   // incompatible exception specifications
+#  pragma warning(disable: 1334)  // the "template" keyword used for syntactic disambiguation may only be used within a template
+#  pragma warning(disable: 1682)  // implicit conversion of a 64-bit integral type to a smaller integral type (potential portability problem)
+#  pragma warning(disable: 1875)  // offsetof applied to non-POD (Plain Old Data) types is nonstandard
+#  pragma warning(disable: 2196)  // warning #2196: routine is both "inline" and "noinline"
+#elif defined(__GNUG__) && !defined(__clang__)
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wunused-but-set-parameter"
+#  pragma GCC diagnostic ignored "-Wunused-but-set-variable"
+#  pragma GCC diagnostic ignored "-Wmissing-field-initializers"
+#  pragma GCC diagnostic ignored "-Wstrict-aliasing"
+#  pragma GCC diagnostic ignored "-Wattributes"
+#  if __GNUC__ >= 7
+#    pragma GCC diagnostic ignored "-Wnoexcept-type"
+#  endif
+#endif
+
+#include "attr.h"
+#include "options.h"
+#include "detail/class.h"
+#include "detail/init.h"
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+/// Wraps an arbitrary C++ function/method/lambda function/.. into a callable Python object
+class cpp_function : public function {
+public:
+    cpp_function() { }
+
+    /// Construct a cpp_function from a vanilla function pointer
+    template <typename Return, typename... Args, typename... Extra>
+    cpp_function(Return (*f)(Args...), const Extra&... extra) {
+        initialize(f, f, extra...);
+    }
+
+    /// Construct a cpp_function from a lambda function (possibly with internal state)
+    template <typename Func, typename... Extra,
+              typename = detail::enable_if_t<detail::is_lambda<Func>::value>>
+    cpp_function(Func &&f, const Extra&... extra) {
+        initialize(std::forward<Func>(f),
+                   (detail::function_signature_t<Func> *) nullptr, extra...);
+    }
+
+    /// Construct a cpp_function from a class method (non-const)
+    template <typename Return, typename Class, typename... Arg, typename... Extra>
+    cpp_function(Return (Class::*f)(Arg...), const Extra&... extra) {
+        initialize([f](Class *c, Arg... args) -> Return { return (c->*f)(args...); },
+                   (Return (*) (Class *, Arg...)) nullptr, extra...);
+    }
+
+    /// Construct a cpp_function from a class method (const)
+    template <typename Return, typename Class, typename... Arg, typename... Extra>
+    cpp_function(Return (Class::*f)(Arg...) const, const Extra&... extra) {
+        initialize([f](const Class *c, Arg... args) -> Return { return (c->*f)(args...); },
+                   (Return (*)(const Class *, Arg ...)) nullptr, extra...);
+    }
+
+    /// Return the function name
+    object name() const { return attr("__name__"); }
+
+protected:
+    /// Space optimization: don't inline this frequently instantiated fragment
+    PYBIND11_NOINLINE detail::function_record *make_function_record() {
+        return new detail::function_record();
+    }
+
+    /// Special internal constructor for functors, lambda functions, etc.
+    template <typename Func, typename Return, typename... Args, typename... Extra>
+    void initialize(Func &&f, Return (*)(Args...), const Extra&... extra) {
+        using namespace detail;
+
+        struct capture { remove_reference_t<Func> f; };
+
+        /* Store the function including any extra state it might have (e.g. a lambda capture object) */
+        auto rec = make_function_record();
+
+        /* Store the capture object directly in the function record if there is enough space */
+        if (sizeof(capture) <= sizeof(rec->data)) {
+            /* Without these pragmas, GCC warns that there might not be
+               enough space to use the placement new operator. However, the
+               'if' statement above ensures that this is the case. */
+#if defined(__GNUG__) && !defined(__clang__) && __GNUC__ >= 6
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wplacement-new"
+#endif
+            new ((capture *) &rec->data) capture { std::forward<Func>(f) };
+#if defined(__GNUG__) && !defined(__clang__) && __GNUC__ >= 6
+#  pragma GCC diagnostic pop
+#endif
+            if (!std::is_trivially_destructible<Func>::value)
+                rec->free_data = [](function_record *r) { ((capture *) &r->data)->~capture(); };
+        } else {
+            rec->data[0] = new capture { std::forward<Func>(f) };
+            rec->free_data = [](function_record *r) { delete ((capture *) r->data[0]); };
+        }
+
+        /* Type casters for the function arguments and return value */
+        using cast_in = argument_loader<Args...>;
+        using cast_out = make_caster<
+            conditional_t<std::is_void<Return>::value, void_type, Return>
+        >;
+
+        static_assert(expected_num_args<Extra...>(sizeof...(Args), cast_in::has_args, cast_in::has_kwargs),
+                      "The number of argument annotations does not match the number of function arguments");
+
+        /* Dispatch code which converts function arguments and performs the actual function call */
+        rec->impl = [](function_call &call) -> handle {
+            cast_in args_converter;
+
+            /* Try to cast the function arguments into the C++ domain */
+            if (!args_converter.load_args(call))
+                return PYBIND11_TRY_NEXT_OVERLOAD;
+
+            /* Invoke call policy pre-call hook */
+            process_attributes<Extra...>::precall(call);
+
+            /* Get a pointer to the capture object */
+            auto data = (sizeof(capture) <= sizeof(call.func.data)
+                         ? &call.func.data : call.func.data[0]);
+            capture *cap = const_cast<capture *>(reinterpret_cast<const capture *>(data));
+
+            /* Override policy for rvalues -- usually to enforce rvp::move on an rvalue */
+            const auto policy = return_value_policy_override<Return>::policy(call.func.policy);
+
+            /* Function scope guard -- defaults to the compile-to-nothing `void_type` */
+            using Guard = extract_guard_t<Extra...>;
+
+            /* Perform the function call */
+            handle result = cast_out::cast(
+                std::move(args_converter).template call<Return, Guard>(cap->f), policy, call.parent);
+
+            /* Invoke call policy post-call hook */
+            process_attributes<Extra...>::postcall(call, result);
+
+            return result;
+        };
+
+        /* Process any user-provided function attributes */
+        process_attributes<Extra...>::init(extra..., rec);
+
+        /* Generate a readable signature describing the function's arguments and return value types */
+        PYBIND11_DESCR signature = _("(") + cast_in::arg_names() + _(") -> ") + cast_out::name();
+
+        /* Register the function with Python from generic (non-templated) code */
+        initialize_generic(rec, signature.text(), signature.types(), sizeof...(Args));
+
+        if (cast_in::has_args) rec->has_args = true;
+        if (cast_in::has_kwargs) rec->has_kwargs = true;
+
+        /* Stash some additional information used by an important optimization in 'functional.h' */
+        using FunctionType = Return (*)(Args...);
+        constexpr bool is_function_ptr =
+            std::is_convertible<Func, FunctionType>::value &&
+            sizeof(capture) == sizeof(void *);
+        if (is_function_ptr) {
+            rec->is_stateless = true;
+            rec->data[1] = const_cast<void *>(reinterpret_cast<const void *>(&typeid(FunctionType)));
+        }
+    }
+
+    /// Register a function call with Python (generic non-templated code goes here)
+    void initialize_generic(detail::function_record *rec, const char *text,
+                            const std::type_info *const *types, size_t args) {
+
+        /* Create copies of all referenced C-style strings */
+        rec->name = strdup(rec->name ? rec->name : "");
+        if (rec->doc) rec->doc = strdup(rec->doc);
+        for (auto &a: rec->args) {
+            if (a.name)
+                a.name = strdup(a.name);
+            if (a.descr)
+                a.descr = strdup(a.descr);
+            else if (a.value)
+                a.descr = strdup(a.value.attr("__repr__")().cast<std::string>().c_str());
+        }
+
+        rec->is_constructor = !strcmp(rec->name, "__init__") || !strcmp(rec->name, "__setstate__");
+
+#if !defined(NDEBUG) && !defined(PYBIND11_DISABLE_NEW_STYLE_INIT_WARNING)
+        if (rec->is_constructor && !rec->is_new_style_constructor) {
+            const auto class_name = std::string(((PyTypeObject *) rec->scope.ptr())->tp_name);
+            const auto func_name = std::string(rec->name);
+            PyErr_WarnEx(
+                PyExc_FutureWarning,
+                ("pybind11-bound class '" + class_name + "' is using an old-style "
+                 "placement-new '" + func_name + "' which has been deprecated. See "
+                 "the upgrade guide in pybind11's docs. This message is only visible "
+                 "when compiled in debug mode.").c_str(), 0
+            );
+        }
+#endif
+
+        /* Generate a proper function signature */
+        std::string signature;
+        size_t type_depth = 0, char_index = 0, type_index = 0, arg_index = 0;
+        while (true) {
+            char c = text[char_index++];
+            if (c == '\0')
+                break;
+
+            if (c == '{') {
+                // Write arg name for everything except *args, **kwargs and return type.
+                if (type_depth == 0 && text[char_index] != '*' && arg_index < args) {
+                    if (!rec->args.empty() && rec->args[arg_index].name) {
+                        signature += rec->args[arg_index].name;
+                    } else if (arg_index == 0 && rec->is_method) {
+                        signature += "self";
+                    } else {
+                        signature += "arg" + std::to_string(arg_index - (rec->is_method ? 1 : 0));
+                    }
+                    signature += ": ";
+                }
+                ++type_depth;
+            } else if (c == '}') {
+                --type_depth;
+                if (type_depth == 0) {
+                    if (arg_index < rec->args.size() && rec->args[arg_index].descr) {
+                        signature += "=";
+                        signature += rec->args[arg_index].descr;
+                    }
+                    arg_index++;
+                }
+            } else if (c == '%') {
+                const std::type_info *t = types[type_index++];
+                if (!t)
+                    pybind11_fail("Internal error while parsing type signature (1)");
+                if (auto tinfo = detail::get_type_info(*t)) {
+                    handle th((PyObject *) tinfo->type);
+                    signature +=
+                        th.attr("__module__").cast<std::string>() + "." +
+                        th.attr("__qualname__").cast<std::string>(); // Python 3.3+, but we backport it to earlier versions
+                } else if (rec->is_new_style_constructor && arg_index == 0) {
+                    // A new-style `__init__` takes `self` as `value_and_holder`.
+                    // Rewrite it to the proper class type.
+                    signature +=
+                        rec->scope.attr("__module__").cast<std::string>() + "." +
+                        rec->scope.attr("__qualname__").cast<std::string>();
+                } else {
+                    std::string tname(t->name());
+                    detail::clean_type_id(tname);
+                    signature += tname;
+                }
+            } else {
+                signature += c;
+            }
+        }
+        if (type_depth != 0 || types[type_index] != nullptr)
+            pybind11_fail("Internal error while parsing type signature (2)");
+
+        #if !defined(PYBIND11_CONSTEXPR_DESCR)
+            delete[] types;
+            delete[] text;
+        #endif
+
+#if PY_MAJOR_VERSION < 3
+        if (strcmp(rec->name, "__next__") == 0) {
+            std::free(rec->name);
+            rec->name = strdup("next");
+        } else if (strcmp(rec->name, "__bool__") == 0) {
+            std::free(rec->name);
+            rec->name = strdup("__nonzero__");
+        }
+#endif
+        rec->signature = strdup(signature.c_str());
+        rec->args.shrink_to_fit();
+        rec->nargs = (std::uint16_t) args;
+
+        if (rec->sibling && PYBIND11_INSTANCE_METHOD_CHECK(rec->sibling.ptr()))
+            rec->sibling = PYBIND11_INSTANCE_METHOD_GET_FUNCTION(rec->sibling.ptr());
+
+        detail::function_record *chain = nullptr, *chain_start = rec;
+        if (rec->sibling) {
+            if (PyCFunction_Check(rec->sibling.ptr())) {
+                auto rec_capsule = reinterpret_borrow<capsule>(PyCFunction_GET_SELF(rec->sibling.ptr()));
+                chain = (detail::function_record *) rec_capsule;
+                /* Never append a method to an overload chain of a parent class;
+                   instead, hide the parent's overloads in this case */
+                if (!chain->scope.is(rec->scope))
+                    chain = nullptr;
+            }
+            // Don't trigger for things like the default __init__, which are wrapper_descriptors that we are intentionally replacing
+            else if (!rec->sibling.is_none() && rec->name[0] != '_')
+                pybind11_fail("Cannot overload existing non-function object \"" + std::string(rec->name) +
+                        "\" with a function of the same name");
+        }
+
+        if (!chain) {
+            /* No existing overload was found, create a new function object */
+            rec->def = new PyMethodDef();
+            std::memset(rec->def, 0, sizeof(PyMethodDef));
+            rec->def->ml_name = rec->name;
+            rec->def->ml_meth = reinterpret_cast<PyCFunction>(*dispatcher);
+            rec->def->ml_flags = METH_VARARGS | METH_KEYWORDS;
+
+            capsule rec_capsule(rec, [](void *ptr) {
+                destruct((detail::function_record *) ptr);
+            });
+
+            object scope_module;
+            if (rec->scope) {
+                if (hasattr(rec->scope, "__module__")) {
+                    scope_module = rec->scope.attr("__module__");
+                } else if (hasattr(rec->scope, "__name__")) {
+                    scope_module = rec->scope.attr("__name__");
+                }
+            }
+
+            m_ptr = PyCFunction_NewEx(rec->def, rec_capsule.ptr(), scope_module.ptr());
+            if (!m_ptr)
+                pybind11_fail("cpp_function::cpp_function(): Could not allocate function object");
+        } else {
+            /* Append at the end of the overload chain */
+            m_ptr = rec->sibling.ptr();
+            inc_ref();
+            chain_start = chain;
+            if (chain->is_method != rec->is_method)
+                pybind11_fail("overloading a method with both static and instance methods is not supported; "
+                    #if defined(NDEBUG)
+                        "compile in debug mode for more details"
+                    #else
+                        "error while attempting to bind " + std::string(rec->is_method ? "instance" : "static") + " method " +
+                        std::string(pybind11::str(rec->scope.attr("__name__"))) + "." + std::string(rec->name) + signature
+                    #endif
+                );
+            while (chain->next)
+                chain = chain->next;
+            chain->next = rec;
+        }
+
+        std::string signatures;
+        int index = 0;
+        /* Create a nice pydoc rec including all signatures and
+           docstrings of the functions in the overload chain */
+        if (chain && options::show_function_signatures()) {
+            // First a generic signature
+            signatures += rec->name;
+            signatures += "(*args, **kwargs)\n";
+            signatures += "Overloaded function.\n\n";
+        }
+        // Then specific overload signatures
+        bool first_user_def = true;
+        for (auto it = chain_start; it != nullptr; it = it->next) {
+            if (options::show_function_signatures()) {
+                if (index > 0) signatures += "\n";
+                if (chain)
+                    signatures += std::to_string(++index) + ". ";
+                signatures += rec->name;
+                signatures += it->signature;
+                signatures += "\n";
+            }
+            if (it->doc && strlen(it->doc) > 0 && options::show_user_defined_docstrings()) {
+                // If we're appending another docstring, and aren't printing function signatures, we
+                // need to append a newline first:
+                if (!options::show_function_signatures()) {
+                    if (first_user_def) first_user_def = false;
+                    else signatures += "\n";
+                }
+                if (options::show_function_signatures()) signatures += "\n";
+                signatures += it->doc;
+                if (options::show_function_signatures()) signatures += "\n";
+            }
+        }
+
+        /* Install docstring */
+        PyCFunctionObject *func = (PyCFunctionObject *) m_ptr;
+        if (func->m_ml->ml_doc)
+            std::free(const_cast<char *>(func->m_ml->ml_doc));
+        func->m_ml->ml_doc = strdup(signatures.c_str());
+
+        if (rec->is_method) {
+            m_ptr = PYBIND11_INSTANCE_METHOD_NEW(m_ptr, rec->scope.ptr());
+            if (!m_ptr)
+                pybind11_fail("cpp_function::cpp_function(): Could not allocate instance method object");
+            Py_DECREF(func);
+        }
+    }
+
+    /// When a cpp_function is GCed, release any memory allocated by pybind11
+    static void destruct(detail::function_record *rec) {
+        while (rec) {
+            detail::function_record *next = rec->next;
+            if (rec->free_data)
+                rec->free_data(rec);
+            std::free((char *) rec->name);
+            std::free((char *) rec->doc);
+            std::free((char *) rec->signature);
+            for (auto &arg: rec->args) {
+                std::free(const_cast<char *>(arg.name));
+                std::free(const_cast<char *>(arg.descr));
+                arg.value.dec_ref();
+            }
+            if (rec->def) {
+                std::free(const_cast<char *>(rec->def->ml_doc));
+                delete rec->def;
+            }
+            delete rec;
+            rec = next;
+        }
+    }
+
+    /// Main dispatch logic for calls to functions bound using pybind11
+    static PyObject *dispatcher(PyObject *self, PyObject *args_in, PyObject *kwargs_in) {
+        using namespace detail;
+
+        /* Iterator over the list of potentially admissible overloads */
+        function_record *overloads = (function_record *) PyCapsule_GetPointer(self, nullptr),
+                        *it = overloads;
+
+        /* Need to know how many arguments + keyword arguments there are to pick the right overload */
+        const size_t n_args_in = (size_t) PyTuple_GET_SIZE(args_in);
+
+        handle parent = n_args_in > 0 ? PyTuple_GET_ITEM(args_in, 0) : nullptr,
+               result = PYBIND11_TRY_NEXT_OVERLOAD;
+
+        auto self_value_and_holder = value_and_holder();
+        if (overloads->is_constructor) {
+            const auto tinfo = get_type_info((PyTypeObject *) overloads->scope.ptr());
+            const auto pi = reinterpret_cast<instance *>(parent.ptr());
+            self_value_and_holder = pi->get_value_and_holder(tinfo, false);
+
+            if (!self_value_and_holder.type || !self_value_and_holder.inst) {
+                PyErr_SetString(PyExc_TypeError, "__init__(self, ...) called with invalid `self` argument");
+                return nullptr;
+            }
+
+            // If this value is already registered it must mean __init__ is invoked multiple times;
+            // we really can't support that in C++, so just ignore the second __init__.
+            if (self_value_and_holder.instance_registered())
+                return none().release().ptr();
+        }
+
+        try {
+            // We do this in two passes: in the first pass, we load arguments with `convert=false`;
+            // in the second, we allow conversion (except for arguments with an explicit
+            // py::arg().noconvert()).  This lets us prefer calls without conversion, with
+            // conversion as a fallback.
+            std::vector<function_call> second_pass;
+
+            // However, if there are no overloads, we can just skip the no-convert pass entirely
+            const bool overloaded = it != nullptr && it->next != nullptr;
+
+            for (; it != nullptr; it = it->next) {
+
+                /* For each overload:
+                   1. Copy all positional arguments we were given, also checking to make sure that
+                      named positional arguments weren't *also* specified via kwarg.
+                   2. If we weren't given enough, try to make up the omitted ones by checking
+                      whether they were provided by a kwarg matching the `py::arg("name")` name.  If
+                      so, use it (and remove it from kwargs; if not, see if the function binding
+                      provided a default that we can use.
+                   3. Ensure that either all keyword arguments were "consumed", or that the function
+                      takes a kwargs argument to accept unconsumed kwargs.
+                   4. Any positional arguments still left get put into a tuple (for args), and any
+                      leftover kwargs get put into a dict.
+                   5. Pack everything into a vector; if we have py::args or py::kwargs, they are an
+                      extra tuple or dict at the end of the positional arguments.
+                   6. Call the function call dispatcher (function_record::impl)
+
+                   If one of these fail, move on to the next overload and keep trying until we get a
+                   result other than PYBIND11_TRY_NEXT_OVERLOAD.
+                 */
+
+                function_record &func = *it;
+                size_t pos_args = func.nargs;    // Number of positional arguments that we need
+                if (func.has_args) --pos_args;   // (but don't count py::args
+                if (func.has_kwargs) --pos_args; //  or py::kwargs)
+
+                if (!func.has_args && n_args_in > pos_args)
+                    continue; // Too many arguments for this overload
+
+                if (n_args_in < pos_args && func.args.size() < pos_args)
+                    continue; // Not enough arguments given, and not enough defaults to fill in the blanks
+
+                function_call call(func, parent);
+
+                size_t args_to_copy = std::min(pos_args, n_args_in);
+                size_t args_copied = 0;
+
+                // 0. Inject new-style `self` argument
+                if (func.is_new_style_constructor) {
+                    // The `value` may have been preallocated by an old-style `__init__`
+                    // if it was a preceding candidate for overload resolution.
+                    if (self_value_and_holder)
+                        self_value_and_holder.type->dealloc(self_value_and_holder);
+
+                    call.init_self = PyTuple_GET_ITEM(args_in, 0);
+                    call.args.push_back(reinterpret_cast<PyObject *>(&self_value_and_holder));
+                    call.args_convert.push_back(false);
+                    ++args_copied;
+                }
+
+                // 1. Copy any position arguments given.
+                bool bad_arg = false;
+                for (; args_copied < args_to_copy; ++args_copied) {
+                    argument_record *arg_rec = args_copied < func.args.size() ? &func.args[args_copied] : nullptr;
+                    if (kwargs_in && arg_rec && arg_rec->name && PyDict_GetItemString(kwargs_in, arg_rec->name)) {
+                        bad_arg = true;
+                        break;
+                    }
+
+                    handle arg(PyTuple_GET_ITEM(args_in, args_copied));
+                    if (arg_rec && !arg_rec->none && arg.is_none()) {
+                        bad_arg = true;
+                        break;
+                    }
+                    call.args.push_back(arg);
+                    call.args_convert.push_back(arg_rec ? arg_rec->convert : true);
+                }
+                if (bad_arg)
+                    continue; // Maybe it was meant for another overload (issue #688)
+
+                // We'll need to copy this if we steal some kwargs for defaults
+                dict kwargs = reinterpret_borrow<dict>(kwargs_in);
+
+                // 2. Check kwargs and, failing that, defaults that may help complete the list
+                if (args_copied < pos_args) {
+                    bool copied_kwargs = false;
+
+                    for (; args_copied < pos_args; ++args_copied) {
+                        const auto &arg = func.args[args_copied];
+
+                        handle value;
+                        if (kwargs_in && arg.name)
+                            value = PyDict_GetItemString(kwargs.ptr(), arg.name);
+
+                        if (value) {
+                            // Consume a kwargs value
+                            if (!copied_kwargs) {
+                                kwargs = reinterpret_steal<dict>(PyDict_Copy(kwargs.ptr()));
+                                copied_kwargs = true;
+                            }
+                            PyDict_DelItemString(kwargs.ptr(), arg.name);
+                        } else if (arg.value) {
+                            value = arg.value;
+                        }
+
+                        if (value) {
+                            call.args.push_back(value);
+                            call.args_convert.push_back(arg.convert);
+                        }
+                        else
+                            break;
+                    }
+
+                    if (args_copied < pos_args)
+                        continue; // Not enough arguments, defaults, or kwargs to fill the positional arguments
+                }
+
+                // 3. Check everything was consumed (unless we have a kwargs arg)
+                if (kwargs && kwargs.size() > 0 && !func.has_kwargs)
+                    continue; // Unconsumed kwargs, but no py::kwargs argument to accept them
+
+                // 4a. If we have a py::args argument, create a new tuple with leftovers
+                if (func.has_args) {
+                    tuple extra_args;
+                    if (args_to_copy == 0) {
+                        // We didn't copy out any position arguments from the args_in tuple, so we
+                        // can reuse it directly without copying:
+                        extra_args = reinterpret_borrow<tuple>(args_in);
+                    } else if (args_copied >= n_args_in) {
+                        extra_args = tuple(0);
+                    } else {
+                        size_t args_size = n_args_in - args_copied;
+                        extra_args = tuple(args_size);
+                        for (size_t i = 0; i < args_size; ++i) {
+                            extra_args[i] = PyTuple_GET_ITEM(args_in, args_copied + i);
+                        }
+                    }
+                    call.args.push_back(extra_args);
+                    call.args_convert.push_back(false);
+                    call.args_ref = std::move(extra_args);
+                }
+
+                // 4b. If we have a py::kwargs, pass on any remaining kwargs
+                if (func.has_kwargs) {
+                    if (!kwargs.ptr())
+                        kwargs = dict(); // If we didn't get one, send an empty one
+                    call.args.push_back(kwargs);
+                    call.args_convert.push_back(false);
+                    call.kwargs_ref = std::move(kwargs);
+                }
+
+                // 5. Put everything in a vector.  Not technically step 5, we've been building it
+                // in `call.args` all along.
+                #if !defined(NDEBUG)
+                if (call.args.size() != func.nargs || call.args_convert.size() != func.nargs)
+                    pybind11_fail("Internal error: function call dispatcher inserted wrong number of arguments!");
+                #endif
+
+                std::vector<bool> second_pass_convert;
+                if (overloaded) {
+                    // We're in the first no-convert pass, so swap out the conversion flags for a
+                    // set of all-false flags.  If the call fails, we'll swap the flags back in for
+                    // the conversion-allowed call below.
+                    second_pass_convert.resize(func.nargs, false);
+                    call.args_convert.swap(second_pass_convert);
+                }
+
+                // 6. Call the function.
+                try {
+                    loader_life_support guard{};
+                    result = func.impl(call);
+                } catch (reference_cast_error &) {
+                    result = PYBIND11_TRY_NEXT_OVERLOAD;
+                }
+
+                if (result.ptr() != PYBIND11_TRY_NEXT_OVERLOAD)
+                    break;
+
+                if (overloaded) {
+                    // The (overloaded) call failed; if the call has at least one argument that
+                    // permits conversion (i.e. it hasn't been explicitly specified `.noconvert()`)
+                    // then add this call to the list of second pass overloads to try.
+                    for (size_t i = func.is_method ? 1 : 0; i < pos_args; i++) {
+                        if (second_pass_convert[i]) {
+                            // Found one: swap the converting flags back in and store the call for
+                            // the second pass.
+                            call.args_convert.swap(second_pass_convert);
+                            second_pass.push_back(std::move(call));
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (overloaded && !second_pass.empty() && result.ptr() == PYBIND11_TRY_NEXT_OVERLOAD) {
+                // The no-conversion pass finished without success, try again with conversion allowed
+                for (auto &call : second_pass) {
+                    try {
+                        loader_life_support guard{};
+                        result = call.func.impl(call);
+                    } catch (reference_cast_error &) {
+                        result = PYBIND11_TRY_NEXT_OVERLOAD;
+                    }
+
+                    if (result.ptr() != PYBIND11_TRY_NEXT_OVERLOAD)
+                        break;
+                }
+            }
+        } catch (error_already_set &e) {
+            e.restore();
+            return nullptr;
+        } catch (...) {
+            /* When an exception is caught, give each registered exception
+               translator a chance to translate it to a Python exception
+               in reverse order of registration.
+
+               A translator may choose to do one of the following:
+
+                - catch the exception and call PyErr_SetString or PyErr_SetObject
+                  to set a standard (or custom) Python exception, or
+                - do nothing and let the exception fall through to the next translator, or
+                - delegate translation to the next translator by throwing a new type of exception. */
+
+            auto last_exception = std::current_exception();
+            auto &registered_exception_translators = get_internals().registered_exception_translators;
+            for (auto& translator : registered_exception_translators) {
+                try {
+                    translator(last_exception);
+                } catch (...) {
+                    last_exception = std::current_exception();
+                    continue;
+                }
+                return nullptr;
+            }
+            PyErr_SetString(PyExc_SystemError, "Exception escaped from default exception translator!");
+            return nullptr;
+        }
+
+        auto append_note_if_missing_header_is_suspected = [](std::string &msg) {
+            if (msg.find("std::") != std::string::npos) {
+                msg += "\n\n"
+                       "Did you forget to `#include <pybind11/stl.h>`? Or <pybind11/complex.h>,\n"
+                       "<pybind11/functional.h>, <pybind11/chrono.h>, etc. Some automatic\n"
+                       "conversions are optional and require extra headers to be included\n"
+                       "when compiling your pybind11 module.";
+            }
+        };
+
+        if (result.ptr() == PYBIND11_TRY_NEXT_OVERLOAD) {
+            if (overloads->is_operator)
+                return handle(Py_NotImplemented).inc_ref().ptr();
+
+            std::string msg = std::string(overloads->name) + "(): incompatible " +
+                std::string(overloads->is_constructor ? "constructor" : "function") +
+                " arguments. The following argument types are supported:\n";
+
+            int ctr = 0;
+            for (function_record *it2 = overloads; it2 != nullptr; it2 = it2->next) {
+                msg += "    "+ std::to_string(++ctr) + ". ";
+
+                bool wrote_sig = false;
+                if (overloads->is_constructor) {
+                    // For a constructor, rewrite `(self: Object, arg0, ...) -> NoneType` as `Object(arg0, ...)`
+                    std::string sig = it2->signature;
+                    size_t start = sig.find('(') + 7; // skip "(self: "
+                    if (start < sig.size()) {
+                        // End at the , for the next argument
+                        size_t end = sig.find(", "), next = end + 2;
+                        size_t ret = sig.rfind(" -> ");
+                        // Or the ), if there is no comma:
+                        if (end >= sig.size()) next = end = sig.find(')');
+                        if (start < end && next < sig.size()) {
+                            msg.append(sig, start, end - start);
+                            msg += '(';
+                            msg.append(sig, next, ret - next);
+                            wrote_sig = true;
+                        }
+                    }
+                }
+                if (!wrote_sig) msg += it2->signature;
+
+                msg += "\n";
+            }
+            msg += "\nInvoked with: ";
+            auto args_ = reinterpret_borrow<tuple>(args_in);
+            bool some_args = false;
+            for (size_t ti = overloads->is_constructor ? 1 : 0; ti < args_.size(); ++ti) {
+                if (!some_args) some_args = true;
+                else msg += ", ";
+                msg += pybind11::repr(args_[ti]);
+            }
+            if (kwargs_in) {
+                auto kwargs = reinterpret_borrow<dict>(kwargs_in);
+                if (kwargs.size() > 0) {
+                    if (some_args) msg += "; ";
+                    msg += "kwargs: ";
+                    bool first = true;
+                    for (auto kwarg : kwargs) {
+                        if (first) first = false;
+                        else msg += ", ";
+                        msg += pybind11::str("{}={!r}").format(kwarg.first, kwarg.second);
+                    }
+                }
+            }
+
+            append_note_if_missing_header_is_suspected(msg);
+            PyErr_SetString(PyExc_TypeError, msg.c_str());
+            return nullptr;
+        } else if (!result) {
+            std::string msg = "Unable to convert function return value to a "
+                              "Python type! The signature was\n\t";
+            msg += it->signature;
+            append_note_if_missing_header_is_suspected(msg);
+            PyErr_SetString(PyExc_TypeError, msg.c_str());
+            return nullptr;
+        } else {
+            if (overloads->is_constructor && !self_value_and_holder.holder_constructed()) {
+                auto *pi = reinterpret_cast<instance *>(parent.ptr());
+                self_value_and_holder.type->init_instance(pi, nullptr);
+            }
+            return result.ptr();
+        }
+    }
+};
+
+/// Wrapper for Python extension modules
+class module : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(module, object, PyModule_Check)
+
+    /// Create a new top-level Python module with the given name and docstring
+    explicit module(const char *name, const char *doc = nullptr) {
+        if (!options::show_user_defined_docstrings()) doc = nullptr;
+#if PY_MAJOR_VERSION >= 3
+        PyModuleDef *def = new PyModuleDef();
+        std::memset(def, 0, sizeof(PyModuleDef));
+        def->m_name = name;
+        def->m_doc = doc;
+        def->m_size = -1;
+        Py_INCREF(def);
+        m_ptr = PyModule_Create(def);
+#else
+        m_ptr = Py_InitModule3(name, nullptr, doc);
+#endif
+        if (m_ptr == nullptr)
+            pybind11_fail("Internal error in module::module()");
+        inc_ref();
+    }
+
+    /** \rst
+        Create Python binding for a new function within the module scope. ``Func``
+        can be a plain C++ function, a function pointer, or a lambda function. For
+        details on the ``Extra&& ... extra`` argument, see section :ref:`extras`.
+    \endrst */
+    template <typename Func, typename... Extra>
+    module &def(const char *name_, Func &&f, const Extra& ... extra) {
+        cpp_function func(std::forward<Func>(f), name(name_), scope(*this),
+                          sibling(getattr(*this, name_, none())), extra...);
+        // NB: allow overwriting here because cpp_function sets up a chain with the intention of
+        // overwriting (and has already checked internally that it isn't overwriting non-functions).
+        add_object(name_, func, true /* overwrite */);
+        return *this;
+    }
+
+    /** \rst
+        Create and return a new Python submodule with the given name and docstring.
+        This also works recursively, i.e.
+
+        .. code-block:: cpp
+
+            py::module m("example", "pybind11 example plugin");
+            py::module m2 = m.def_submodule("sub", "A submodule of 'example'");
+            py::module m3 = m2.def_submodule("subsub", "A submodule of 'example.sub'");
+    \endrst */
+    module def_submodule(const char *name, const char *doc = nullptr) {
+        std::string full_name = std::string(PyModule_GetName(m_ptr))
+            + std::string(".") + std::string(name);
+        auto result = reinterpret_borrow<module>(PyImport_AddModule(full_name.c_str()));
+        if (doc && options::show_user_defined_docstrings())
+            result.attr("__doc__") = pybind11::str(doc);
+        attr(name) = result;
+        return result;
+    }
+
+    /// Import and return a module or throws `error_already_set`.
+    static module import(const char *name) {
+        PyObject *obj = PyImport_ImportModule(name);
+        if (!obj)
+            throw error_already_set();
+        return reinterpret_steal<module>(obj);
+    }
+
+    /// Reload the module or throws `error_already_set`.
+    void reload() {
+        PyObject *obj = PyImport_ReloadModule(ptr());
+        if (!obj)
+            throw error_already_set();
+        *this = reinterpret_steal<module>(obj);
+    }
+
+    // Adds an object to the module using the given name.  Throws if an object with the given name
+    // already exists.
+    //
+    // overwrite should almost always be false: attempting to overwrite objects that pybind11 has
+    // established will, in most cases, break things.
+    PYBIND11_NOINLINE void add_object(const char *name, handle obj, bool overwrite = false) {
+        if (!overwrite && hasattr(*this, name))
+            pybind11_fail("Error during initialization: multiple incompatible definitions with name \"" +
+                    std::string(name) + "\"");
+
+        PyModule_AddObject(ptr(), name, obj.inc_ref().ptr() /* steals a reference */);
+    }
+};
+
+/// \ingroup python_builtins
+/// Return a dictionary representing the global variables in the current execution frame,
+/// or ``__main__.__dict__`` if there is no frame (usually when the interpreter is embedded).
+inline dict globals() {
+    PyObject *p = PyEval_GetGlobals();
+    return reinterpret_borrow<dict>(p ? p : module::import("__main__").attr("__dict__").ptr());
+}
+
+NAMESPACE_BEGIN(detail)
+/// Generic support for creating new Python heap types
+class generic_type : public object {
+    template <typename...> friend class class_;
+public:
+    PYBIND11_OBJECT_DEFAULT(generic_type, object, PyType_Check)
+protected:
+    void initialize(const type_record &rec) {
+        if (rec.scope && hasattr(rec.scope, rec.name))
+            pybind11_fail("generic_type: cannot initialize type \"" + std::string(rec.name) +
+                          "\": an object with that name is already defined");
+
+        if (rec.module_local ? get_local_type_info(*rec.type) : get_global_type_info(*rec.type))
+            pybind11_fail("generic_type: type \"" + std::string(rec.name) +
+                          "\" is already registered!");
+
+        m_ptr = make_new_python_type(rec);
+
+        /* Register supplemental type information in C++ dict */
+        auto *tinfo = new detail::type_info();
+        tinfo->type = (PyTypeObject *) m_ptr;
+        tinfo->cpptype = rec.type;
+        tinfo->type_size = rec.type_size;
+        tinfo->operator_new = rec.operator_new;
+        tinfo->holder_size_in_ptrs = size_in_ptrs(rec.holder_size);
+        tinfo->init_instance = rec.init_instance;
+        tinfo->dealloc = rec.dealloc;
+        tinfo->simple_type = true;
+        tinfo->simple_ancestors = true;
+        tinfo->default_holder = rec.default_holder;
+        tinfo->module_local = rec.module_local;
+
+        auto &internals = get_internals();
+        auto tindex = std::type_index(*rec.type);
+        tinfo->direct_conversions = &internals.direct_conversions[tindex];
+        if (rec.module_local)
+            registered_local_types_cpp()[tindex] = tinfo;
+        else
+            internals.registered_types_cpp[tindex] = tinfo;
+        internals.registered_types_py[(PyTypeObject *) m_ptr] = { tinfo };
+
+        if (rec.bases.size() > 1 || rec.multiple_inheritance) {
+            mark_parents_nonsimple(tinfo->type);
+            tinfo->simple_ancestors = false;
+        }
+        else if (rec.bases.size() == 1) {
+            auto parent_tinfo = get_type_info((PyTypeObject *) rec.bases[0].ptr());
+            tinfo->simple_ancestors = parent_tinfo->simple_ancestors;
+        }
+
+        if (rec.module_local) {
+            // Stash the local typeinfo and loader so that external modules can access it.
+            tinfo->module_local_load = &type_caster_generic::local_load;
+            setattr(m_ptr, PYBIND11_MODULE_LOCAL_ID, capsule(tinfo));
+        }
+    }
+
+    /// Helper function which tags all parents of a type using mult. inheritance
+    void mark_parents_nonsimple(PyTypeObject *value) {
+        auto t = reinterpret_borrow<tuple>(value->tp_bases);
+        for (handle h : t) {
+            auto tinfo2 = get_type_info((PyTypeObject *) h.ptr());
+            if (tinfo2)
+                tinfo2->simple_type = false;
+            mark_parents_nonsimple((PyTypeObject *) h.ptr());
+        }
+    }
+
+    void install_buffer_funcs(
+            buffer_info *(*get_buffer)(PyObject *, void *),
+            void *get_buffer_data) {
+        PyHeapTypeObject *type = (PyHeapTypeObject*) m_ptr;
+        auto tinfo = detail::get_type_info(&type->ht_type);
+
+        if (!type->ht_type.tp_as_buffer)
+            pybind11_fail(
+                "To be able to register buffer protocol support for the type '" +
+                std::string(tinfo->type->tp_name) +
+                "' the associated class<>(..) invocation must "
+                "include the pybind11::buffer_protocol() annotation!");
+
+        tinfo->get_buffer = get_buffer;
+        tinfo->get_buffer_data = get_buffer_data;
+    }
+
+    void def_property_static_impl(const char *name,
+                                  handle fget, handle fset,
+                                  detail::function_record *rec_fget) {
+        const auto is_static = !(rec_fget->is_method && rec_fget->scope);
+        const auto has_doc = rec_fget->doc && pybind11::options::show_user_defined_docstrings();
+
+        auto property = handle((PyObject *) (is_static ? get_internals().static_property_type
+                                                       : &PyProperty_Type));
+        attr(name) = property(fget.ptr() ? fget : none(),
+                              fset.ptr() ? fset : none(),
+                              /*deleter*/none(),
+                              pybind11::str(has_doc ? rec_fget->doc : ""));
+    }
+};
+
+/// Set the pointer to operator new if it exists. The cast is needed because it can be overloaded.
+template <typename T, typename = void_t<decltype(static_cast<void *(*)(size_t)>(T::operator new))>>
+void set_operator_new(type_record *r) { r->operator_new = &T::operator new; }
+
+template <typename> void set_operator_new(...) { }
+
+template <typename T, typename SFINAE = void> struct has_operator_delete : std::false_type { };
+template <typename T> struct has_operator_delete<T, void_t<decltype(static_cast<void (*)(void *)>(T::operator delete))>>
+    : std::true_type { };
+template <typename T, typename SFINAE = void> struct has_operator_delete_size : std::false_type { };
+template <typename T> struct has_operator_delete_size<T, void_t<decltype(static_cast<void (*)(void *, size_t)>(T::operator delete))>>
+    : std::true_type { };
+/// Call class-specific delete if it exists or global otherwise. Can also be an overload set.
+template <typename T, enable_if_t<has_operator_delete<T>::value, int> = 0>
+void call_operator_delete(T *p, size_t) { T::operator delete(p); }
+template <typename T, enable_if_t<!has_operator_delete<T>::value && has_operator_delete_size<T>::value, int> = 0>
+void call_operator_delete(T *p, size_t s) { T::operator delete(p, s); }
+
+inline void call_operator_delete(void *p, size_t) { ::operator delete(p); }
+
+NAMESPACE_END(detail)
+
+/// Given a pointer to a member function, cast it to its `Derived` version.
+/// Forward everything else unchanged.
+template <typename /*Derived*/, typename F>
+auto method_adaptor(F &&f) -> decltype(std::forward<F>(f)) { return std::forward<F>(f); }
+
+template <typename Derived, typename Return, typename Class, typename... Args>
+auto method_adaptor(Return (Class::*pmf)(Args...)) -> Return (Derived::*)(Args...) { return pmf; }
+
+template <typename Derived, typename Return, typename Class, typename... Args>
+auto method_adaptor(Return (Class::*pmf)(Args...) const) -> Return (Derived::*)(Args...) const { return pmf; }
+
+template <typename type_, typename... options>
+class class_ : public detail::generic_type {
+    template <typename T> using is_holder = detail::is_holder_type<type_, T>;
+    template <typename T> using is_subtype = detail::is_strict_base_of<type_, T>;
+    template <typename T> using is_base = detail::is_strict_base_of<T, type_>;
+    // struct instead of using here to help MSVC:
+    template <typename T> struct is_valid_class_option :
+        detail::any_of<is_holder<T>, is_subtype<T>, is_base<T>> {};
+
+public:
+    using type = type_;
+    using type_alias = detail::exactly_one_t<is_subtype, void, options...>;
+    constexpr static bool has_alias = !std::is_void<type_alias>::value;
+    using holder_type = detail::exactly_one_t<is_holder, std::unique_ptr<type>, options...>;
+
+    static_assert(detail::all_of<is_valid_class_option<options>...>::value,
+            "Unknown/invalid class_ template parameters provided");
+
+    static_assert(!has_alias || std::is_polymorphic<type>::value,
+            "Cannot use an alias class with a non-polymorphic type");
+
+    PYBIND11_OBJECT(class_, generic_type, PyType_Check)
+
+    template <typename... Extra>
+    class_(handle scope, const char *name, const Extra &... extra) {
+        using namespace detail;
+
+        // MI can only be specified via class_ template options, not constructor parameters
+        static_assert(
+            none_of<is_pyobject<Extra>...>::value || // no base class arguments, or:
+            (   constexpr_sum(is_pyobject<Extra>::value...) == 1 && // Exactly one base
+                constexpr_sum(is_base<options>::value...)   == 0 && // no template option bases
+                none_of<std::is_same<multiple_inheritance, Extra>...>::value), // no multiple_inheritance attr
+            "Error: multiple inheritance bases must be specified via class_ template options");
+
+        type_record record;
+        record.scope = scope;
+        record.name = name;
+        record.type = &typeid(type);
+        record.type_size = sizeof(conditional_t<has_alias, type_alias, type>);
+        record.holder_size = sizeof(holder_type);
+        record.init_instance = init_instance;
+        record.dealloc = dealloc;
+        record.default_holder = std::is_same<holder_type, std::unique_ptr<type>>::value;
+
+        set_operator_new<type>(&record);
+
+        /* Register base classes specified via template arguments to class_, if any */
+        PYBIND11_EXPAND_SIDE_EFFECTS(add_base<options>(record));
+
+        /* Process optional arguments, if any */
+        process_attributes<Extra...>::init(extra..., &record);
+
+        generic_type::initialize(record);
+
+        if (has_alias) {
+            auto &instances = record.module_local ? registered_local_types_cpp() : get_internals().registered_types_cpp;
+            instances[std::type_index(typeid(type_alias))] = instances[std::type_index(typeid(type))];
+        }
+    }
+
+    template <typename Base, detail::enable_if_t<is_base<Base>::value, int> = 0>
+    static void add_base(detail::type_record &rec) {
+        rec.add_base(typeid(Base), [](void *src) -> void * {
+            return static_cast<Base *>(reinterpret_cast<type *>(src));
+        });
+    }
+
+    template <typename Base, detail::enable_if_t<!is_base<Base>::value, int> = 0>
+    static void add_base(detail::type_record &) { }
+
+    template <typename Func, typename... Extra>
+    class_ &def(const char *name_, Func&& f, const Extra&... extra) {
+        cpp_function cf(method_adaptor<type>(std::forward<Func>(f)), name(name_), is_method(*this),
+                        sibling(getattr(*this, name_, none())), extra...);
+        attr(cf.name()) = cf;
+        return *this;
+    }
+
+    template <typename Func, typename... Extra> class_ &
+    def_static(const char *name_, Func &&f, const Extra&... extra) {
+        static_assert(!std::is_member_function_pointer<Func>::value,
+                "def_static(...) called with a non-static member function pointer");
+        cpp_function cf(std::forward<Func>(f), name(name_), scope(*this),
+                        sibling(getattr(*this, name_, none())), extra...);
+        attr(cf.name()) = cf;
+        return *this;
+    }
+
+    template <detail::op_id id, detail::op_type ot, typename L, typename R, typename... Extra>
+    class_ &def(const detail::op_<id, ot, L, R> &op, const Extra&... extra) {
+        op.execute(*this, extra...);
+        return *this;
+    }
+
+    template <detail::op_id id, detail::op_type ot, typename L, typename R, typename... Extra>
+    class_ & def_cast(const detail::op_<id, ot, L, R> &op, const Extra&... extra) {
+        op.execute_cast(*this, extra...);
+        return *this;
+    }
+
+    template <typename... Args, typename... Extra>
+    class_ &def(const detail::initimpl::constructor<Args...> &init, const Extra&... extra) {
+        init.execute(*this, extra...);
+        return *this;
+    }
+
+    template <typename... Args, typename... Extra>
+    class_ &def(const detail::initimpl::alias_constructor<Args...> &init, const Extra&... extra) {
+        init.execute(*this, extra...);
+        return *this;
+    }
+
+    template <typename... Args, typename... Extra>
+    class_ &def(detail::initimpl::factory<Args...> &&init, const Extra&... extra) {
+        std::move(init).execute(*this, extra...);
+        return *this;
+    }
+
+    template <typename... Args, typename... Extra>
+    class_ &def(detail::initimpl::pickle_factory<Args...> &&pf, const Extra &...extra) {
+        std::move(pf).execute(*this, extra...);
+        return *this;
+    }
+
+    template <typename Func> class_& def_buffer(Func &&func) {
+        struct capture { Func func; };
+        capture *ptr = new capture { std::forward<Func>(func) };
+        install_buffer_funcs([](PyObject *obj, void *ptr) -> buffer_info* {
+            detail::make_caster<type> caster;
+            if (!caster.load(obj, false))
+                return nullptr;
+            return new buffer_info(((capture *) ptr)->func(caster));
+        }, ptr);
+        return *this;
+    }
+
+    template <typename Return, typename Class, typename... Args>
+    class_ &def_buffer(Return (Class::*func)(Args...)) {
+        return def_buffer([func] (type &obj) { return (obj.*func)(); });
+    }
+
+    template <typename Return, typename Class, typename... Args>
+    class_ &def_buffer(Return (Class::*func)(Args...) const) {
+        return def_buffer([func] (const type &obj) { return (obj.*func)(); });
+    }
+
+    template <typename C, typename D, typename... Extra>
+    class_ &def_readwrite(const char *name, D C::*pm, const Extra&... extra) {
+        static_assert(std::is_base_of<C, type>::value, "def_readwrite() requires a class member (or base class member)");
+        cpp_function fget([pm](const type &c) -> const D &{ return c.*pm; }, is_method(*this)),
+                     fset([pm](type &c, const D &value) { c.*pm = value; }, is_method(*this));
+        def_property(name, fget, fset, return_value_policy::reference_internal, extra...);
+        return *this;
+    }
+
+    template <typename C, typename D, typename... Extra>
+    class_ &def_readonly(const char *name, const D C::*pm, const Extra& ...extra) {
+        static_assert(std::is_base_of<C, type>::value, "def_readonly() requires a class member (or base class member)");
+        cpp_function fget([pm](const type &c) -> const D &{ return c.*pm; }, is_method(*this));
+        def_property_readonly(name, fget, return_value_policy::reference_internal, extra...);
+        return *this;
+    }
+
+    template <typename D, typename... Extra>
+    class_ &def_readwrite_static(const char *name, D *pm, const Extra& ...extra) {
+        cpp_function fget([pm](object) -> const D &{ return *pm; }, scope(*this)),
+                     fset([pm](object, const D &value) { *pm = value; }, scope(*this));
+        def_property_static(name, fget, fset, return_value_policy::reference, extra...);
+        return *this;
+    }
+
+    template <typename D, typename... Extra>
+    class_ &def_readonly_static(const char *name, const D *pm, const Extra& ...extra) {
+        cpp_function fget([pm](object) -> const D &{ return *pm; }, scope(*this));
+        def_property_readonly_static(name, fget, return_value_policy::reference, extra...);
+        return *this;
+    }
+
+    /// Uses return_value_policy::reference_internal by default
+    template <typename Getter, typename... Extra>
+    class_ &def_property_readonly(const char *name, const Getter &fget, const Extra& ...extra) {
+        return def_property_readonly(name, cpp_function(method_adaptor<type>(fget)),
+                                     return_value_policy::reference_internal, extra...);
+    }
+
+    /// Uses cpp_function's return_value_policy by default
+    template <typename... Extra>
+    class_ &def_property_readonly(const char *name, const cpp_function &fget, const Extra& ...extra) {
+        return def_property(name, fget, cpp_function(), extra...);
+    }
+
+    /// Uses return_value_policy::reference by default
+    template <typename Getter, typename... Extra>
+    class_ &def_property_readonly_static(const char *name, const Getter &fget, const Extra& ...extra) {
+        return def_property_readonly_static(name, cpp_function(fget), return_value_policy::reference, extra...);
+    }
+
+    /// Uses cpp_function's return_value_policy by default
+    template <typename... Extra>
+    class_ &def_property_readonly_static(const char *name, const cpp_function &fget, const Extra& ...extra) {
+        return def_property_static(name, fget, cpp_function(), extra...);
+    }
+
+    /// Uses return_value_policy::reference_internal by default
+    template <typename Getter, typename Setter, typename... Extra>
+    class_ &def_property(const char *name, const Getter &fget, const Setter &fset, const Extra& ...extra) {
+        return def_property(name, fget, cpp_function(method_adaptor<type>(fset)), extra...);
+    }
+    template <typename Getter, typename... Extra>
+    class_ &def_property(const char *name, const Getter &fget, const cpp_function &fset, const Extra& ...extra) {
+        return def_property(name, cpp_function(method_adaptor<type>(fget)), fset,
+                            return_value_policy::reference_internal, extra...);
+    }
+
+    /// Uses cpp_function's return_value_policy by default
+    template <typename... Extra>
+    class_ &def_property(const char *name, const cpp_function &fget, const cpp_function &fset, const Extra& ...extra) {
+        return def_property_static(name, fget, fset, is_method(*this), extra...);
+    }
+
+    /// Uses return_value_policy::reference by default
+    template <typename Getter, typename... Extra>
+    class_ &def_property_static(const char *name, const Getter &fget, const cpp_function &fset, const Extra& ...extra) {
+        return def_property_static(name, cpp_function(fget), fset, return_value_policy::reference, extra...);
+    }
+
+    /// Uses cpp_function's return_value_policy by default
+    template <typename... Extra>
+    class_ &def_property_static(const char *name, const cpp_function &fget, const cpp_function &fset, const Extra& ...extra) {
+        auto rec_fget = get_function_record(fget), rec_fset = get_function_record(fset);
+        char *doc_prev = rec_fget->doc; /* 'extra' field may include a property-specific documentation string */
+        detail::process_attributes<Extra...>::init(extra..., rec_fget);
+        if (rec_fget->doc && rec_fget->doc != doc_prev) {
+            free(doc_prev);
+            rec_fget->doc = strdup(rec_fget->doc);
+        }
+        if (rec_fset) {
+            doc_prev = rec_fset->doc;
+            detail::process_attributes<Extra...>::init(extra..., rec_fset);
+            if (rec_fset->doc && rec_fset->doc != doc_prev) {
+                free(doc_prev);
+                rec_fset->doc = strdup(rec_fset->doc);
+            }
+        }
+        def_property_static_impl(name, fget, fset, rec_fget);
+        return *this;
+    }
+
+private:
+    /// Initialize holder object, variant 1: object derives from enable_shared_from_this
+    template <typename T>
+    static void init_holder(detail::instance *inst, detail::value_and_holder &v_h,
+            const holder_type * /* unused */, const std::enable_shared_from_this<T> * /* dummy */) {
+        try {
+            auto sh = std::dynamic_pointer_cast<typename holder_type::element_type>(
+                    v_h.value_ptr<type>()->shared_from_this());
+            if (sh) {
+                new (&v_h.holder<holder_type>()) holder_type(std::move(sh));
+                v_h.set_holder_constructed();
+            }
+        } catch (const std::bad_weak_ptr &) {}
+
+        if (!v_h.holder_constructed() && inst->owned) {
+            new (&v_h.holder<holder_type>()) holder_type(v_h.value_ptr<type>());
+            v_h.set_holder_constructed();
+        }
+    }
+
+    static void init_holder_from_existing(const detail::value_and_holder &v_h,
+            const holder_type *holder_ptr, std::true_type /*is_copy_constructible*/) {
+        new (&v_h.holder<holder_type>()) holder_type(*reinterpret_cast<const holder_type *>(holder_ptr));
+    }
+
+    static void init_holder_from_existing(const detail::value_and_holder &v_h,
+            const holder_type *holder_ptr, std::false_type /*is_copy_constructible*/) {
+        new (&v_h.holder<holder_type>()) holder_type(std::move(*const_cast<holder_type *>(holder_ptr)));
+    }
+
+    /// Initialize holder object, variant 2: try to construct from existing holder object, if possible
+    static void init_holder(detail::instance *inst, detail::value_and_holder &v_h,
+            const holder_type *holder_ptr, const void * /* dummy -- not enable_shared_from_this<T>) */) {
+        if (holder_ptr) {
+            init_holder_from_existing(v_h, holder_ptr, std::is_copy_constructible<holder_type>());
+            v_h.set_holder_constructed();
+        } else if (inst->owned || detail::always_construct_holder<holder_type>::value) {
+            new (&v_h.holder<holder_type>()) holder_type(v_h.value_ptr<type>());
+            v_h.set_holder_constructed();
+        }
+    }
+
+    /// Performs instance initialization including constructing a holder and registering the known
+    /// instance.  Should be called as soon as the `type` value_ptr is set for an instance.  Takes an
+    /// optional pointer to an existing holder to use; if not specified and the instance is
+    /// `.owned`, a new holder will be constructed to manage the value pointer.
+    static void init_instance(detail::instance *inst, const void *holder_ptr) {
+        auto v_h = inst->get_value_and_holder(detail::get_type_info(typeid(type)));
+        if (!v_h.instance_registered()) {
+            register_instance(inst, v_h.value_ptr(), v_h.type);
+            v_h.set_instance_registered();
+        }
+        init_holder(inst, v_h, (const holder_type *) holder_ptr, v_h.value_ptr<type>());
+    }
+
+    /// Deallocates an instance; via holder, if constructed; otherwise via operator delete.
+    static void dealloc(detail::value_and_holder &v_h) {
+        if (v_h.holder_constructed()) {
+            v_h.holder<holder_type>().~holder_type();
+            v_h.set_holder_constructed(false);
+        }
+        else {
+            detail::call_operator_delete(v_h.value_ptr<type>(), v_h.type->type_size);
+        }
+        v_h.value_ptr() = nullptr;
+    }
+
+    static detail::function_record *get_function_record(handle h) {
+        h = detail::get_function(h);
+        return h ? (detail::function_record *) reinterpret_borrow<capsule>(PyCFunction_GET_SELF(h.ptr()))
+                 : nullptr;
+    }
+};
+
+/// Binds an existing constructor taking arguments Args...
+template <typename... Args> detail::initimpl::constructor<Args...> init() { return {}; }
+/// Like `init<Args...>()`, but the instance is always constructed through the alias class (even
+/// when not inheriting on the Python side).
+template <typename... Args> detail::initimpl::alias_constructor<Args...> init_alias() { return {}; }
+
+/// Binds a factory function as a constructor
+template <typename Func, typename Ret = detail::initimpl::factory<Func>>
+Ret init(Func &&f) { return {std::forward<Func>(f)}; }
+
+/// Dual-argument factory function: the first function is called when no alias is needed, the second
+/// when an alias is needed (i.e. due to python-side inheritance).  Arguments must be identical.
+template <typename CFunc, typename AFunc, typename Ret = detail::initimpl::factory<CFunc, AFunc>>
+Ret init(CFunc &&c, AFunc &&a) {
+    return {std::forward<CFunc>(c), std::forward<AFunc>(a)};
+}
+
+/// Binds pickling functions `__getstate__` and `__setstate__` and ensures that the type
+/// returned by `__getstate__` is the same as the argument accepted by `__setstate__`.
+template <typename GetState, typename SetState>
+detail::initimpl::pickle_factory<GetState, SetState> pickle(GetState &&g, SetState &&s) {
+    return {std::forward<GetState>(g), std::forward<SetState>(s)};
+}
+
+/// Binds C++ enumerations and enumeration classes to Python
+template <typename Type> class enum_ : public class_<Type> {
+public:
+    using class_<Type>::def;
+    using class_<Type>::def_property_readonly_static;
+    using Scalar = typename std::underlying_type<Type>::type;
+
+    template <typename... Extra>
+    enum_(const handle &scope, const char *name, const Extra&... extra)
+      : class_<Type>(scope, name, extra...), m_entries(), m_parent(scope) {
+
+        constexpr bool is_arithmetic = detail::any_of<std::is_same<arithmetic, Extra>...>::value;
+
+        auto m_entries_ptr = m_entries.inc_ref().ptr();
+        def("__repr__", [name, m_entries_ptr](Type value) -> pybind11::str {
+            for (const auto &kv : reinterpret_borrow<dict>(m_entries_ptr)) {
+                if (pybind11::cast<Type>(kv.second) == value)
+                    return pybind11::str("{}.{}").format(name, kv.first);
+            }
+            return pybind11::str("{}.???").format(name);
+        });
+        def_property_readonly_static("__members__", [m_entries_ptr](object /* self */) {
+            dict m;
+            for (const auto &kv : reinterpret_borrow<dict>(m_entries_ptr))
+                m[kv.first] = kv.second;
+            return m;
+        }, return_value_policy::copy);
+        def(init([](Scalar i) { return static_cast<Type>(i); }));
+        def("__int__", [](Type value) { return (Scalar) value; });
+        #if PY_MAJOR_VERSION < 3
+            def("__long__", [](Type value) { return (Scalar) value; });
+        #endif
+        def("__eq__", [](const Type &value, Type *value2) { return value2 && value == *value2; });
+        def("__ne__", [](const Type &value, Type *value2) { return !value2 || value != *value2; });
+        if (is_arithmetic) {
+            def("__lt__", [](const Type &value, Type *value2) { return value2 && value < *value2; });
+            def("__gt__", [](const Type &value, Type *value2) { return value2 && value > *value2; });
+            def("__le__", [](const Type &value, Type *value2) { return value2 && value <= *value2; });
+            def("__ge__", [](const Type &value, Type *value2) { return value2 && value >= *value2; });
+        }
+        if (std::is_convertible<Type, Scalar>::value) {
+            // Don't provide comparison with the underlying type if the enum isn't convertible,
+            // i.e. if Type is a scoped enum, mirroring the C++ behaviour.  (NB: we explicitly
+            // convert Type to Scalar below anyway because this needs to compile).
+            def("__eq__", [](const Type &value, Scalar value2) { return (Scalar) value == value2; });
+            def("__ne__", [](const Type &value, Scalar value2) { return (Scalar) value != value2; });
+            if (is_arithmetic) {
+                def("__lt__", [](const Type &value, Scalar value2) { return (Scalar) value < value2; });
+                def("__gt__", [](const Type &value, Scalar value2) { return (Scalar) value > value2; });
+                def("__le__", [](const Type &value, Scalar value2) { return (Scalar) value <= value2; });
+                def("__ge__", [](const Type &value, Scalar value2) { return (Scalar) value >= value2; });
+                def("__invert__", [](const Type &value) { return ~((Scalar) value); });
+                def("__and__", [](const Type &value, Scalar value2) { return (Scalar) value & value2; });
+                def("__or__", [](const Type &value, Scalar value2) { return (Scalar) value | value2; });
+                def("__xor__", [](const Type &value, Scalar value2) { return (Scalar) value ^ value2; });
+                def("__rand__", [](const Type &value, Scalar value2) { return (Scalar) value & value2; });
+                def("__ror__", [](const Type &value, Scalar value2) { return (Scalar) value | value2; });
+                def("__rxor__", [](const Type &value, Scalar value2) { return (Scalar) value ^ value2; });
+                def("__and__", [](const Type &value, const Type &value2) { return (Scalar) value & (Scalar) value2; });
+                def("__or__", [](const Type &value, const Type &value2) { return (Scalar) value | (Scalar) value2; });
+                def("__xor__", [](const Type &value, const Type &value2) { return (Scalar) value ^ (Scalar) value2; });
+            }
+        }
+        def("__hash__", [](const Type &value) { return (Scalar) value; });
+        // Pickling and unpickling -- needed for use with the 'multiprocessing' module
+        def(pickle([](const Type &value) { return pybind11::make_tuple((Scalar) value); },
+                   [](tuple t) { return static_cast<Type>(t[0].cast<Scalar>()); }));
+    }
+
+    /// Export enumeration entries into the parent scope
+    enum_& export_values() {
+        for (const auto &kv : m_entries)
+            m_parent.attr(kv.first) = kv.second;
+        return *this;
+    }
+
+    /// Add an enumeration entry
+    enum_& value(char const* name, Type value) {
+        auto v = pybind11::cast(value, return_value_policy::copy);
+        this->attr(name) = v;
+        m_entries[pybind11::str(name)] = v;
+        return *this;
+    }
+
+private:
+    dict m_entries;
+    handle m_parent;
+};
+
+NAMESPACE_BEGIN(detail)
+
+
+inline void keep_alive_impl(handle nurse, handle patient) {
+    if (!nurse || !patient)
+        pybind11_fail("Could not activate keep_alive!");
+
+    if (patient.is_none() || nurse.is_none())
+        return; /* Nothing to keep alive or nothing to be kept alive by */
+
+    auto tinfo = all_type_info(Py_TYPE(nurse.ptr()));
+    if (!tinfo.empty()) {
+        /* It's a pybind-registered type, so we can store the patient in the
+         * internal list. */
+        add_patient(nurse.ptr(), patient.ptr());
+    }
+    else {
+        /* Fall back to clever approach based on weak references taken from
+         * Boost.Python. This is not used for pybind-registered types because
+         * the objects can be destroyed out-of-order in a GC pass. */
+        cpp_function disable_lifesupport(
+            [patient](handle weakref) { patient.dec_ref(); weakref.dec_ref(); });
+
+        weakref wr(nurse, disable_lifesupport);
+
+        patient.inc_ref(); /* reference patient and leak the weak reference */
+        (void) wr.release();
+    }
+}
+
+PYBIND11_NOINLINE inline void keep_alive_impl(size_t Nurse, size_t Patient, function_call &call, handle ret) {
+    auto get_arg = [&](size_t n) {
+        if (n == 0)
+            return ret;
+        else if (n == 1 && call.init_self)
+            return call.init_self;
+        else if (n <= call.args.size())
+            return call.args[n - 1];
+        return handle();
+    };
+
+    keep_alive_impl(get_arg(Nurse), get_arg(Patient));
+}
+
+inline std::pair<decltype(internals::registered_types_py)::iterator, bool> all_type_info_get_cache(PyTypeObject *type) {
+    auto res = get_internals().registered_types_py
+#ifdef __cpp_lib_unordered_map_try_emplace
+        .try_emplace(type);
+#else
+        .emplace(type, std::vector<detail::type_info *>());
+#endif
+    if (res.second) {
+        // New cache entry created; set up a weak reference to automatically remove it if the type
+        // gets destroyed:
+        weakref((PyObject *) type, cpp_function([type](handle wr) {
+            get_internals().registered_types_py.erase(type);
+            wr.dec_ref();
+        })).release();
+    }
+
+    return res;
+}
+
+template <typename Iterator, typename Sentinel, bool KeyIterator, return_value_policy Policy>
+struct iterator_state {
+    Iterator it;
+    Sentinel end;
+    bool first_or_done;
+};
+
+NAMESPACE_END(detail)
+
+/// Makes a python iterator from a first and past-the-end C++ InputIterator.
+template <return_value_policy Policy = return_value_policy::reference_internal,
+          typename Iterator,
+          typename Sentinel,
+          typename ValueType = decltype(*std::declval<Iterator>()),
+          typename... Extra>
+iterator make_iterator(Iterator first, Sentinel last, Extra &&... extra) {
+    typedef detail::iterator_state<Iterator, Sentinel, false, Policy> state;
+
+    if (!detail::get_type_info(typeid(state), false)) {
+        class_<state>(handle(), "iterator", pybind11::module_local())
+            .def("__iter__", [](state &s) -> state& { return s; })
+            .def("__next__", [](state &s) -> ValueType {
+                if (!s.first_or_done)
+                    ++s.it;
+                else
+                    s.first_or_done = false;
+                if (s.it == s.end) {
+                    s.first_or_done = true;
+                    throw stop_iteration();
+                }
+                return *s.it;
+            }, std::forward<Extra>(extra)..., Policy);
+    }
+
+    return cast(state{first, last, true});
+}
+
+/// Makes an python iterator over the keys (`.first`) of a iterator over pairs from a
+/// first and past-the-end InputIterator.
+template <return_value_policy Policy = return_value_policy::reference_internal,
+          typename Iterator,
+          typename Sentinel,
+          typename KeyType = decltype((*std::declval<Iterator>()).first),
+          typename... Extra>
+iterator make_key_iterator(Iterator first, Sentinel last, Extra &&... extra) {
+    typedef detail::iterator_state<Iterator, Sentinel, true, Policy> state;
+
+    if (!detail::get_type_info(typeid(state), false)) {
+        class_<state>(handle(), "iterator", pybind11::module_local())
+            .def("__iter__", [](state &s) -> state& { return s; })
+            .def("__next__", [](state &s) -> KeyType {
+                if (!s.first_or_done)
+                    ++s.it;
+                else
+                    s.first_or_done = false;
+                if (s.it == s.end) {
+                    s.first_or_done = true;
+                    throw stop_iteration();
+                }
+                return (*s.it).first;
+            }, std::forward<Extra>(extra)..., Policy);
+    }
+
+    return cast(state{first, last, true});
+}
+
+/// Makes an iterator over values of an stl container or other container supporting
+/// `std::begin()`/`std::end()`
+template <return_value_policy Policy = return_value_policy::reference_internal,
+          typename Type, typename... Extra> iterator make_iterator(Type &value, Extra&&... extra) {
+    return make_iterator<Policy>(std::begin(value), std::end(value), extra...);
+}
+
+/// Makes an iterator over the keys (`.first`) of a stl map-like container supporting
+/// `std::begin()`/`std::end()`
+template <return_value_policy Policy = return_value_policy::reference_internal,
+          typename Type, typename... Extra> iterator make_key_iterator(Type &value, Extra&&... extra) {
+    return make_key_iterator<Policy>(std::begin(value), std::end(value), extra...);
+}
+
+template <typename InputType, typename OutputType> void implicitly_convertible() {
+    struct set_flag {
+        bool &flag;
+        set_flag(bool &flag) : flag(flag) { flag = true; }
+        ~set_flag() { flag = false; }
+    };
+    auto implicit_caster = [](PyObject *obj, PyTypeObject *type) -> PyObject * {
+        static bool currently_used = false;
+        if (currently_used) // implicit conversions are non-reentrant
+            return nullptr;
+        set_flag flag_helper(currently_used);
+        if (!detail::make_caster<InputType>().load(obj, false))
+            return nullptr;
+        tuple args(1);
+        args[0] = obj;
+        PyObject *result = PyObject_Call((PyObject *) type, args.ptr(), nullptr);
+        if (result == nullptr)
+            PyErr_Clear();
+        return result;
+    };
+
+    if (auto tinfo = detail::get_type_info(typeid(OutputType)))
+        tinfo->implicit_conversions.push_back(implicit_caster);
+    else
+        pybind11_fail("implicitly_convertible: Unable to find type " + type_id<OutputType>());
+}
+
+template <typename ExceptionTranslator>
+void register_exception_translator(ExceptionTranslator&& translator) {
+    detail::get_internals().registered_exception_translators.push_front(
+        std::forward<ExceptionTranslator>(translator));
+}
+
+/**
+ * Wrapper to generate a new Python exception type.
+ *
+ * This should only be used with PyErr_SetString for now.
+ * It is not (yet) possible to use as a py::base.
+ * Template type argument is reserved for future use.
+ */
+template <typename type>
+class exception : public object {
+public:
+    exception(handle scope, const char *name, PyObject *base = PyExc_Exception) {
+        std::string full_name = scope.attr("__name__").cast<std::string>() +
+                                std::string(".") + name;
+        m_ptr = PyErr_NewException(const_cast<char *>(full_name.c_str()), base, NULL);
+        if (hasattr(scope, name))
+            pybind11_fail("Error during initialization: multiple incompatible "
+                          "definitions with name \"" + std::string(name) + "\"");
+        scope.attr(name) = *this;
+    }
+
+    // Sets the current python exception to this exception object with the given message
+    void operator()(const char *message) {
+        PyErr_SetString(m_ptr, message);
+    }
+};
+
+/**
+ * Registers a Python exception in `m` of the given `name` and installs an exception translator to
+ * translate the C++ exception to the created Python exception using the exceptions what() method.
+ * This is intended for simple exception translations; for more complex translation, register the
+ * exception object and translator directly.
+ */
+template <typename CppException>
+exception<CppException> &register_exception(handle scope,
+                                            const char *name,
+                                            PyObject *base = PyExc_Exception) {
+    static exception<CppException> ex(scope, name, base);
+    register_exception_translator([](std::exception_ptr p) {
+        if (!p) return;
+        try {
+            std::rethrow_exception(p);
+        } catch (const CppException &e) {
+            ex(e.what());
+        }
+    });
+    return ex;
+}
+
+NAMESPACE_BEGIN(detail)
+PYBIND11_NOINLINE inline void print(tuple args, dict kwargs) {
+    auto strings = tuple(args.size());
+    for (size_t i = 0; i < args.size(); ++i) {
+        strings[i] = str(args[i]);
+    }
+    auto sep = kwargs.contains("sep") ? kwargs["sep"] : cast(" ");
+    auto line = sep.attr("join")(strings);
+
+    object file;
+    if (kwargs.contains("file")) {
+        file = kwargs["file"].cast<object>();
+    } else {
+        try {
+            file = module::import("sys").attr("stdout");
+        } catch (const error_already_set &) {
+            /* If print() is called from code that is executed as
+               part of garbage collection during interpreter shutdown,
+               importing 'sys' can fail. Give up rather than crashing the
+               interpreter in this case. */
+            return;
+        }
+    }
+
+    auto write = file.attr("write");
+    write(line);
+    write(kwargs.contains("end") ? kwargs["end"] : cast("\n"));
+
+    if (kwargs.contains("flush") && kwargs["flush"].cast<bool>())
+        file.attr("flush")();
+}
+NAMESPACE_END(detail)
+
+template <return_value_policy policy = return_value_policy::automatic_reference, typename... Args>
+void print(Args &&...args) {
+    auto c = detail::collect_arguments<policy>(std::forward<Args>(args)...);
+    detail::print(c.args(), c.kwargs());
+}
+
+#if defined(WITH_THREAD) && !defined(PYPY_VERSION)
+
+/* The functions below essentially reproduce the PyGILState_* API using a RAII
+ * pattern, but there are a few important differences:
+ *
+ * 1. When acquiring the GIL from an non-main thread during the finalization
+ *    phase, the GILState API blindly terminates the calling thread, which
+ *    is often not what is wanted. This API does not do this.
+ *
+ * 2. The gil_scoped_release function can optionally cut the relationship
+ *    of a PyThreadState and its associated thread, which allows moving it to
+ *    another thread (this is a fairly rare/advanced use case).
+ *
+ * 3. The reference count of an acquired thread state can be controlled. This
+ *    can be handy to prevent cases where callbacks issued from an external
+ *    thread would otherwise constantly construct and destroy thread state data
+ *    structures.
+ *
+ * See the Python bindings of NanoGUI (http://github.com/wjakob/nanogui) for an
+ * example which uses features 2 and 3 to migrate the Python thread of
+ * execution to another thread (to run the event loop on the original thread,
+ * in this case).
+ */
+
+class gil_scoped_acquire {
+public:
+    PYBIND11_NOINLINE gil_scoped_acquire() {
+        auto const &internals = detail::get_internals();
+        tstate = (PyThreadState *) PyThread_get_key_value(internals.tstate);
+
+        if (!tstate) {
+            tstate = PyThreadState_New(internals.istate);
+            #if !defined(NDEBUG)
+                if (!tstate)
+                    pybind11_fail("scoped_acquire: could not create thread state!");
+            #endif
+            tstate->gilstate_counter = 0;
+            #if PY_MAJOR_VERSION < 3
+                PyThread_delete_key_value(internals.tstate);
+            #endif
+            PyThread_set_key_value(internals.tstate, tstate);
+        } else {
+            release = detail::get_thread_state_unchecked() != tstate;
+        }
+
+        if (release) {
+            /* Work around an annoying assertion in PyThreadState_Swap */
+            #if defined(Py_DEBUG)
+                PyInterpreterState *interp = tstate->interp;
+                tstate->interp = nullptr;
+            #endif
+            PyEval_AcquireThread(tstate);
+            #if defined(Py_DEBUG)
+                tstate->interp = interp;
+            #endif
+        }
+
+        inc_ref();
+    }
+
+    void inc_ref() {
+        ++tstate->gilstate_counter;
+    }
+
+    PYBIND11_NOINLINE void dec_ref() {
+        --tstate->gilstate_counter;
+        #if !defined(NDEBUG)
+            if (detail::get_thread_state_unchecked() != tstate)
+                pybind11_fail("scoped_acquire::dec_ref(): thread state must be current!");
+            if (tstate->gilstate_counter < 0)
+                pybind11_fail("scoped_acquire::dec_ref(): reference count underflow!");
+        #endif
+        if (tstate->gilstate_counter == 0) {
+            #if !defined(NDEBUG)
+                if (!release)
+                    pybind11_fail("scoped_acquire::dec_ref(): internal error!");
+            #endif
+            PyThreadState_Clear(tstate);
+            PyThreadState_DeleteCurrent();
+            PyThread_delete_key_value(detail::get_internals().tstate);
+            release = false;
+        }
+    }
+
+    PYBIND11_NOINLINE ~gil_scoped_acquire() {
+        dec_ref();
+        if (release)
+           PyEval_SaveThread();
+    }
+private:
+    PyThreadState *tstate = nullptr;
+    bool release = true;
+};
+
+class gil_scoped_release {
+public:
+    explicit gil_scoped_release(bool disassoc = false) : disassoc(disassoc) {
+        // `get_internals()` must be called here unconditionally in order to initialize
+        // `internals.tstate` for subsequent `gil_scoped_acquire` calls. Otherwise, an
+        // initialization race could occur as multiple threads try `gil_scoped_acquire`.
+        const auto &internals = detail::get_internals();
+        tstate = PyEval_SaveThread();
+        if (disassoc) {
+            auto key = internals.tstate;
+            #if PY_MAJOR_VERSION < 3
+                PyThread_delete_key_value(key);
+            #else
+                PyThread_set_key_value(key, nullptr);
+            #endif
+        }
+    }
+    ~gil_scoped_release() {
+        if (!tstate)
+            return;
+        PyEval_RestoreThread(tstate);
+        if (disassoc) {
+            auto key = detail::get_internals().tstate;
+            #if PY_MAJOR_VERSION < 3
+                PyThread_delete_key_value(key);
+            #endif
+            PyThread_set_key_value(key, tstate);
+        }
+    }
+private:
+    PyThreadState *tstate;
+    bool disassoc;
+};
+#elif defined(PYPY_VERSION)
+class gil_scoped_acquire {
+    PyGILState_STATE state;
+public:
+    gil_scoped_acquire() { state = PyGILState_Ensure(); }
+    ~gil_scoped_acquire() { PyGILState_Release(state); }
+};
+
+class gil_scoped_release {
+    PyThreadState *state;
+public:
+    gil_scoped_release() { state = PyEval_SaveThread(); }
+    ~gil_scoped_release() { PyEval_RestoreThread(state); }
+};
+#else
+class gil_scoped_acquire { };
+class gil_scoped_release { };
+#endif
+
+error_already_set::~error_already_set() {
+    if (type) {
+        gil_scoped_acquire gil;
+        type.release().dec_ref();
+        value.release().dec_ref();
+        trace.release().dec_ref();
+    }
+}
+
+inline function get_type_overload(const void *this_ptr, const detail::type_info *this_type, const char *name)  {
+    handle self = detail::get_object_handle(this_ptr, this_type);
+    if (!self)
+        return function();
+    handle type = self.get_type();
+    auto key = std::make_pair(type.ptr(), name);
+
+    /* Cache functions that aren't overloaded in Python to avoid
+       many costly Python dictionary lookups below */
+    auto &cache = detail::get_internals().inactive_overload_cache;
+    if (cache.find(key) != cache.end())
+        return function();
+
+    function overload = getattr(self, name, function());
+    if (overload.is_cpp_function()) {
+        cache.insert(key);
+        return function();
+    }
+
+    /* Don't call dispatch code if invoked from overridden function.
+       Unfortunately this doesn't work on PyPy. */
+#if !defined(PYPY_VERSION)
+    PyFrameObject *frame = PyThreadState_Get()->frame;
+    if (frame && (std::string) str(frame->f_code->co_name) == name &&
+        frame->f_code->co_argcount > 0) {
+        PyFrame_FastToLocals(frame);
+        PyObject *self_caller = PyDict_GetItem(
+            frame->f_locals, PyTuple_GET_ITEM(frame->f_code->co_varnames, 0));
+        if (self_caller == self.ptr())
+            return function();
+    }
+#else
+    /* PyPy currently doesn't provide a detailed cpyext emulation of
+       frame objects, so we have to emulate this using Python. This
+       is going to be slow..*/
+    dict d; d["self"] = self; d["name"] = pybind11::str(name);
+    PyObject *result = PyRun_String(
+        "import inspect\n"
+        "frame = inspect.currentframe()\n"
+        "if frame is not None:\n"
+        "    frame = frame.f_back\n"
+        "    if frame is not None and str(frame.f_code.co_name) == name and "
+        "frame.f_code.co_argcount > 0:\n"
+        "        self_caller = frame.f_locals[frame.f_code.co_varnames[0]]\n"
+        "        if self_caller == self:\n"
+        "            self = None\n",
+        Py_file_input, d.ptr(), d.ptr());
+    if (result == nullptr)
+        throw error_already_set();
+    if (d["self"].is_none())
+        return function();
+    Py_DECREF(result);
+#endif
+
+    return overload;
+}
+
+template <class T> function get_overload(const T *this_ptr, const char *name) {
+    auto tinfo = detail::get_type_info(typeid(T));
+    return tinfo ? get_type_overload(this_ptr, tinfo, name) : function();
+}
+
+#define PYBIND11_OVERLOAD_INT(ret_type, cname, name, ...) { \
+        pybind11::gil_scoped_acquire gil; \
+        pybind11::function overload = pybind11::get_overload(static_cast<const cname *>(this), name); \
+        if (overload) { \
+            auto o = overload(__VA_ARGS__); \
+            if (pybind11::detail::cast_is_temporary_value_reference<ret_type>::value) { \
+                static pybind11::detail::overload_caster_t<ret_type> caster; \
+                return pybind11::detail::cast_ref<ret_type>(std::move(o), caster); \
+            } \
+            else return pybind11::detail::cast_safe<ret_type>(std::move(o)); \
+        } \
+    }
+
+#define PYBIND11_OVERLOAD_NAME(ret_type, cname, name, fn, ...) \
+    PYBIND11_OVERLOAD_INT(ret_type, cname, name, __VA_ARGS__) \
+    return cname::fn(__VA_ARGS__)
+
+#define PYBIND11_OVERLOAD_PURE_NAME(ret_type, cname, name, fn, ...) \
+    PYBIND11_OVERLOAD_INT(ret_type, cname, name, __VA_ARGS__) \
+    pybind11::pybind11_fail("Tried to call pure virtual function \"" #cname "::" name "\"");
+
+#define PYBIND11_OVERLOAD(ret_type, cname, fn, ...) \
+    PYBIND11_OVERLOAD_NAME(ret_type, cname, #fn, fn, __VA_ARGS__)
+
+#define PYBIND11_OVERLOAD_PURE(ret_type, cname, fn, ...) \
+    PYBIND11_OVERLOAD_PURE_NAME(ret_type, cname, #fn, fn, __VA_ARGS__)
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
+
+#if defined(_MSC_VER)
+#  pragma warning(pop)
+#elif defined(__INTEL_COMPILER)
+/* Leave ignored warnings on */
+#elif defined(__GNUG__) && !defined(__clang__)
+#  pragma GCC diagnostic pop
+#endif
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/pytypes.h b/ml/dlib/dlib/external/pybind11/include/pybind11/pytypes.h
new file mode 100644
index 000000000..d7fa17775
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/pytypes.h
@@ -0,0 +1,1332 @@
+/*
+    pybind11/pytypes.h: Convenience wrapper classes for basic Python types
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+#include "buffer_info.h"
+#include <utility>
+#include <type_traits>
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+/* A few forward declarations */
+class handle; class object;
+class str; class iterator;
+struct arg; struct arg_v;
+
+NAMESPACE_BEGIN(detail)
+class args_proxy;
+inline bool isinstance_generic(handle obj, const std::type_info &tp);
+
+// Accessor forward declarations
+template <typename Policy> class accessor;
+namespace accessor_policies {
+    struct obj_attr;
+    struct str_attr;
+    struct generic_item;
+    struct sequence_item;
+    struct list_item;
+    struct tuple_item;
+}
+using obj_attr_accessor = accessor<accessor_policies::obj_attr>;
+using str_attr_accessor = accessor<accessor_policies::str_attr>;
+using item_accessor = accessor<accessor_policies::generic_item>;
+using sequence_accessor = accessor<accessor_policies::sequence_item>;
+using list_accessor = accessor<accessor_policies::list_item>;
+using tuple_accessor = accessor<accessor_policies::tuple_item>;
+
+/// Tag and check to identify a class which implements the Python object API
+class pyobject_tag { };
+template <typename T> using is_pyobject = std::is_base_of<pyobject_tag, remove_reference_t<T>>;
+
+/** \rst
+    A mixin class which adds common functions to `handle`, `object` and various accessors.
+    The only requirement for `Derived` is to implement ``PyObject *Derived::ptr() const``.
+\endrst */
+template <typename Derived>
+class object_api : public pyobject_tag {
+    const Derived &derived() const { return static_cast<const Derived &>(*this); }
+
+public:
+    /** \rst
+        Return an iterator equivalent to calling ``iter()`` in Python. The object
+        must be a collection which supports the iteration protocol.
+    \endrst */
+    iterator begin() const;
+    /// Return a sentinel which ends iteration.
+    iterator end() const;
+
+    /** \rst
+        Return an internal functor to invoke the object's sequence protocol. Casting
+        the returned ``detail::item_accessor`` instance to a `handle` or `object`
+        subclass causes a corresponding call to ``__getitem__``. Assigning a `handle`
+        or `object` subclass causes a call to ``__setitem__``.
+    \endrst */
+    item_accessor operator[](handle key) const;
+    /// See above (the only difference is that they key is provided as a string literal)
+    item_accessor operator[](const char *key) const;
+
+    /** \rst
+        Return an internal functor to access the object's attributes. Casting the
+        returned ``detail::obj_attr_accessor`` instance to a `handle` or `object`
+        subclass causes a corresponding call to ``getattr``. Assigning a `handle`
+        or `object` subclass causes a call to ``setattr``.
+    \endrst */
+    obj_attr_accessor attr(handle key) const;
+    /// See above (the only difference is that they key is provided as a string literal)
+    str_attr_accessor attr(const char *key) const;
+
+    /** \rst
+        Matches * unpacking in Python, e.g. to unpack arguments out of a ``tuple``
+        or ``list`` for a function call. Applying another * to the result yields
+        ** unpacking, e.g. to unpack a dict as function keyword arguments.
+        See :ref:`calling_python_functions`.
+    \endrst */
+    args_proxy operator*() const;
+
+    /// Check if the given item is contained within this object, i.e. ``item in obj``.
+    template <typename T> bool contains(T &&item) const;
+
+    /** \rst
+        Assuming the Python object is a function or implements the ``__call__``
+        protocol, ``operator()`` invokes the underlying function, passing an
+        arbitrary set of parameters. The result is returned as a `object` and
+        may need to be converted back into a Python object using `handle::cast()`.
+
+        When some of the arguments cannot be converted to Python objects, the
+        function will throw a `cast_error` exception. When the Python function
+        call fails, a `error_already_set` exception is thrown.
+    \endrst */
+    template <return_value_policy policy = return_value_policy::automatic_reference, typename... Args>
+    object operator()(Args &&...args) const;
+    template <return_value_policy policy = return_value_policy::automatic_reference, typename... Args>
+    PYBIND11_DEPRECATED("call(...) was deprecated in favor of operator()(...)")
+        object call(Args&&... args) const;
+
+    /// Equivalent to ``obj is other`` in Python.
+    bool is(object_api const& other) const { return derived().ptr() == other.derived().ptr(); }
+    /// Equivalent to ``obj is None`` in Python.
+    bool is_none() const { return derived().ptr() == Py_None; }
+    PYBIND11_DEPRECATED("Use py::str(obj) instead")
+    pybind11::str str() const;
+
+    /// Get or set the object's docstring, i.e. ``obj.__doc__``.
+    str_attr_accessor doc() const;
+
+    /// Return the object's current reference count
+    int ref_count() const { return static_cast<int>(Py_REFCNT(derived().ptr())); }
+    /// Return a handle to the Python type object underlying the instance
+    handle get_type() const;
+};
+
+NAMESPACE_END(detail)
+
+/** \rst
+    Holds a reference to a Python object (no reference counting)
+
+    The `handle` class is a thin wrapper around an arbitrary Python object (i.e. a
+    ``PyObject *`` in Python's C API). It does not perform any automatic reference
+    counting and merely provides a basic C++ interface to various Python API functions.
+
+    .. seealso::
+        The `object` class inherits from `handle` and adds automatic reference
+        counting features.
+\endrst */
+class handle : public detail::object_api<handle> {
+public:
+    /// The default constructor creates a handle with a ``nullptr``-valued pointer
+    handle() = default;
+    /// Creates a ``handle`` from the given raw Python object pointer
+    handle(PyObject *ptr) : m_ptr(ptr) { } // Allow implicit conversion from PyObject*
+
+    /// Return the underlying ``PyObject *`` pointer
+    PyObject *ptr() const { return m_ptr; }
+    PyObject *&ptr() { return m_ptr; }
+
+    /** \rst
+        Manually increase the reference count of the Python object. Usually, it is
+        preferable to use the `object` class which derives from `handle` and calls
+        this function automatically. Returns a reference to itself.
+    \endrst */
+    const handle& inc_ref() const & { Py_XINCREF(m_ptr); return *this; }
+
+    /** \rst
+        Manually decrease the reference count of the Python object. Usually, it is
+        preferable to use the `object` class which derives from `handle` and calls
+        this function automatically. Returns a reference to itself.
+    \endrst */
+    const handle& dec_ref() const & { Py_XDECREF(m_ptr); return *this; }
+
+    /** \rst
+        Attempt to cast the Python object into the given C++ type. A `cast_error`
+        will be throw upon failure.
+    \endrst */
+    template <typename T> T cast() const;
+    /// Return ``true`` when the `handle` wraps a valid Python object
+    explicit operator bool() const { return m_ptr != nullptr; }
+    /** \rst
+        Deprecated: Check that the underlying pointers are the same.
+        Equivalent to ``obj1 is obj2`` in Python.
+    \endrst */
+    PYBIND11_DEPRECATED("Use obj1.is(obj2) instead")
+    bool operator==(const handle &h) const { return m_ptr == h.m_ptr; }
+    PYBIND11_DEPRECATED("Use !obj1.is(obj2) instead")
+    bool operator!=(const handle &h) const { return m_ptr != h.m_ptr; }
+    PYBIND11_DEPRECATED("Use handle::operator bool() instead")
+    bool check() const { return m_ptr != nullptr; }
+protected:
+    PyObject *m_ptr = nullptr;
+};
+
+/** \rst
+    Holds a reference to a Python object (with reference counting)
+
+    Like `handle`, the `object` class is a thin wrapper around an arbitrary Python
+    object (i.e. a ``PyObject *`` in Python's C API). In contrast to `handle`, it
+    optionally increases the object's reference count upon construction, and it
+    *always* decreases the reference count when the `object` instance goes out of
+    scope and is destructed. When using `object` instances consistently, it is much
+    easier to get reference counting right at the first attempt.
+\endrst */
+class object : public handle {
+public:
+    object() = default;
+    PYBIND11_DEPRECATED("Use reinterpret_borrow<object>() or reinterpret_steal<object>()")
+    object(handle h, bool is_borrowed) : handle(h) { if (is_borrowed) inc_ref(); }
+    /// Copy constructor; always increases the reference count
+    object(const object &o) : handle(o) { inc_ref(); }
+    /// Move constructor; steals the object from ``other`` and preserves its reference count
+    object(object &&other) noexcept { m_ptr = other.m_ptr; other.m_ptr = nullptr; }
+    /// Destructor; automatically calls `handle::dec_ref()`
+    ~object() { dec_ref(); }
+
+    /** \rst
+        Resets the internal pointer to ``nullptr`` without without decreasing the
+        object's reference count. The function returns a raw handle to the original
+        Python object.
+    \endrst */
+    handle release() {
+      PyObject *tmp = m_ptr;
+      m_ptr = nullptr;
+      return handle(tmp);
+    }
+
+    object& operator=(const object &other) {
+        other.inc_ref();
+        dec_ref();
+        m_ptr = other.m_ptr;
+        return *this;
+    }
+
+    object& operator=(object &&other) noexcept {
+        if (this != &other) {
+            handle temp(m_ptr);
+            m_ptr = other.m_ptr;
+            other.m_ptr = nullptr;
+            temp.dec_ref();
+        }
+        return *this;
+    }
+
+    // Calling cast() on an object lvalue just copies (via handle::cast)
+    template <typename T> T cast() const &;
+    // Calling on an object rvalue does a move, if needed and/or possible
+    template <typename T> T cast() &&;
+
+protected:
+    // Tags for choosing constructors from raw PyObject *
+    struct borrowed_t { };
+    struct stolen_t { };
+
+    template <typename T> friend T reinterpret_borrow(handle);
+    template <typename T> friend T reinterpret_steal(handle);
+
+public:
+    // Only accessible from derived classes and the reinterpret_* functions
+    object(handle h, borrowed_t) : handle(h) { inc_ref(); }
+    object(handle h, stolen_t) : handle(h) { }
+};
+
+/** \rst
+    Declare that a `handle` or ``PyObject *`` is a certain type and borrow the reference.
+    The target type ``T`` must be `object` or one of its derived classes. The function
+    doesn't do any conversions or checks. It's up to the user to make sure that the
+    target type is correct.
+
+    .. code-block:: cpp
+
+        PyObject *p = PyList_GetItem(obj, index);
+        py::object o = reinterpret_borrow<py::object>(p);
+        // or
+        py::tuple t = reinterpret_borrow<py::tuple>(p); // <-- `p` must be already be a `tuple`
+\endrst */
+template <typename T> T reinterpret_borrow(handle h) { return {h, object::borrowed_t{}}; }
+
+/** \rst
+    Like `reinterpret_borrow`, but steals the reference.
+
+     .. code-block:: cpp
+
+        PyObject *p = PyObject_Str(obj);
+        py::str s = reinterpret_steal<py::str>(p); // <-- `p` must be already be a `str`
+\endrst */
+template <typename T> T reinterpret_steal(handle h) { return {h, object::stolen_t{}}; }
+
+NAMESPACE_BEGIN(detail)
+inline std::string error_string();
+NAMESPACE_END(detail)
+
+/// Fetch and hold an error which was already set in Python.  An instance of this is typically
+/// thrown to propagate python-side errors back through C++ which can either be caught manually or
+/// else falls back to the function dispatcher (which then raises the captured error back to
+/// python).
+class error_already_set : public std::runtime_error {
+public:
+    /// Constructs a new exception from the current Python error indicator, if any.  The current
+    /// Python error indicator will be cleared.
+    error_already_set() : std::runtime_error(detail::error_string()) {
+        PyErr_Fetch(&type.ptr(), &value.ptr(), &trace.ptr());
+    }
+
+    inline ~error_already_set();
+
+    /// Give the currently-held error back to Python, if any.  If there is currently a Python error
+    /// already set it is cleared first.  After this call, the current object no longer stores the
+    /// error variables (but the `.what()` string is still available).
+    void restore() { PyErr_Restore(type.release().ptr(), value.release().ptr(), trace.release().ptr()); }
+
+    // Does nothing; provided for backwards compatibility.
+    PYBIND11_DEPRECATED("Use of error_already_set.clear() is deprecated")
+    void clear() {}
+
+    /// Check if the currently trapped error type matches the given Python exception class (or a
+    /// subclass thereof).  May also be passed a tuple to search for any exception class matches in
+    /// the given tuple.
+    bool matches(handle ex) const { return PyErr_GivenExceptionMatches(ex.ptr(), type.ptr()); }
+
+private:
+    object type, value, trace;
+};
+
+/** \defgroup python_builtins _
+    Unless stated otherwise, the following C++ functions behave the same
+    as their Python counterparts.
+ */
+
+/** \ingroup python_builtins
+    \rst
+    Return true if ``obj`` is an instance of ``T``. Type ``T`` must be a subclass of
+    `object` or a class which was exposed to Python as ``py::class_<T>``.
+\endrst */
+template <typename T, detail::enable_if_t<std::is_base_of<object, T>::value, int> = 0>
+bool isinstance(handle obj) { return T::check_(obj); }
+
+template <typename T, detail::enable_if_t<!std::is_base_of<object, T>::value, int> = 0>
+bool isinstance(handle obj) { return detail::isinstance_generic(obj, typeid(T)); }
+
+template <> inline bool isinstance<handle>(handle obj) = delete;
+template <> inline bool isinstance<object>(handle obj) { return obj.ptr() != nullptr; }
+
+/// \ingroup python_builtins
+/// Return true if ``obj`` is an instance of the ``type``.
+inline bool isinstance(handle obj, handle type) {
+    const auto result = PyObject_IsInstance(obj.ptr(), type.ptr());
+    if (result == -1)
+        throw error_already_set();
+    return result != 0;
+}
+
+/// \addtogroup python_builtins
+/// @{
+inline bool hasattr(handle obj, handle name) {
+    return PyObject_HasAttr(obj.ptr(), name.ptr()) == 1;
+}
+
+inline bool hasattr(handle obj, const char *name) {
+    return PyObject_HasAttrString(obj.ptr(), name) == 1;
+}
+
+inline object getattr(handle obj, handle name) {
+    PyObject *result = PyObject_GetAttr(obj.ptr(), name.ptr());
+    if (!result) { throw error_already_set(); }
+    return reinterpret_steal<object>(result);
+}
+
+inline object getattr(handle obj, const char *name) {
+    PyObject *result = PyObject_GetAttrString(obj.ptr(), name);
+    if (!result) { throw error_already_set(); }
+    return reinterpret_steal<object>(result);
+}
+
+inline object getattr(handle obj, handle name, handle default_) {
+    if (PyObject *result = PyObject_GetAttr(obj.ptr(), name.ptr())) {
+        return reinterpret_steal<object>(result);
+    } else {
+        PyErr_Clear();
+        return reinterpret_borrow<object>(default_);
+    }
+}
+
+inline object getattr(handle obj, const char *name, handle default_) {
+    if (PyObject *result = PyObject_GetAttrString(obj.ptr(), name)) {
+        return reinterpret_steal<object>(result);
+    } else {
+        PyErr_Clear();
+        return reinterpret_borrow<object>(default_);
+    }
+}
+
+inline void setattr(handle obj, handle name, handle value) {
+    if (PyObject_SetAttr(obj.ptr(), name.ptr(), value.ptr()) != 0) { throw error_already_set(); }
+}
+
+inline void setattr(handle obj, const char *name, handle value) {
+    if (PyObject_SetAttrString(obj.ptr(), name, value.ptr()) != 0) { throw error_already_set(); }
+}
+
+inline ssize_t hash(handle obj) {
+    auto h = PyObject_Hash(obj.ptr());
+    if (h == -1) { throw error_already_set(); }
+    return h;
+}
+
+/// @} python_builtins
+
+NAMESPACE_BEGIN(detail)
+inline handle get_function(handle value) {
+    if (value) {
+#if PY_MAJOR_VERSION >= 3
+        if (PyInstanceMethod_Check(value.ptr()))
+            value = PyInstanceMethod_GET_FUNCTION(value.ptr());
+        else
+#endif
+        if (PyMethod_Check(value.ptr()))
+            value = PyMethod_GET_FUNCTION(value.ptr());
+    }
+    return value;
+}
+
+// Helper aliases/functions to support implicit casting of values given to python accessors/methods.
+// When given a pyobject, this simply returns the pyobject as-is; for other C++ type, the value goes
+// through pybind11::cast(obj) to convert it to an `object`.
+template <typename T, enable_if_t<is_pyobject<T>::value, int> = 0>
+auto object_or_cast(T &&o) -> decltype(std::forward<T>(o)) { return std::forward<T>(o); }
+// The following casting version is implemented in cast.h:
+template <typename T, enable_if_t<!is_pyobject<T>::value, int> = 0>
+object object_or_cast(T &&o);
+// Match a PyObject*, which we want to convert directly to handle via its converting constructor
+inline handle object_or_cast(PyObject *ptr) { return ptr; }
+
+
+template <typename Policy>
+class accessor : public object_api<accessor<Policy>> {
+    using key_type = typename Policy::key_type;
+
+public:
+    accessor(handle obj, key_type key) : obj(obj), key(std::move(key)) { }
+    accessor(const accessor &) = default;
+    accessor(accessor &&) = default;
+
+    // accessor overload required to override default assignment operator (templates are not allowed
+    // to replace default compiler-generated assignments).
+    void operator=(const accessor &a) && { std::move(*this).operator=(handle(a)); }
+    void operator=(const accessor &a) & { operator=(handle(a)); }
+
+    template <typename T> void operator=(T &&value) && {
+        Policy::set(obj, key, object_or_cast(std::forward<T>(value)));
+    }
+    template <typename T> void operator=(T &&value) & {
+        get_cache() = reinterpret_borrow<object>(object_or_cast(std::forward<T>(value)));
+    }
+
+    template <typename T = Policy>
+    PYBIND11_DEPRECATED("Use of obj.attr(...) as bool is deprecated in favor of pybind11::hasattr(obj, ...)")
+    explicit operator enable_if_t<std::is_same<T, accessor_policies::str_attr>::value ||
+            std::is_same<T, accessor_policies::obj_attr>::value, bool>() const {
+        return hasattr(obj, key);
+    }
+    template <typename T = Policy>
+    PYBIND11_DEPRECATED("Use of obj[key] as bool is deprecated in favor of obj.contains(key)")
+    explicit operator enable_if_t<std::is_same<T, accessor_policies::generic_item>::value, bool>() const {
+        return obj.contains(key);
+    }
+
+    operator object() const { return get_cache(); }
+    PyObject *ptr() const { return get_cache().ptr(); }
+    template <typename T> T cast() const { return get_cache().template cast<T>(); }
+
+private:
+    object &get_cache() const {
+        if (!cache) { cache = Policy::get(obj, key); }
+        return cache;
+    }
+
+private:
+    handle obj;
+    key_type key;
+    mutable object cache;
+};
+
+NAMESPACE_BEGIN(accessor_policies)
+struct obj_attr {
+    using key_type = object;
+    static object get(handle obj, handle key) { return getattr(obj, key); }
+    static void set(handle obj, handle key, handle val) { setattr(obj, key, val); }
+};
+
+struct str_attr {
+    using key_type = const char *;
+    static object get(handle obj, const char *key) { return getattr(obj, key); }
+    static void set(handle obj, const char *key, handle val) { setattr(obj, key, val); }
+};
+
+struct generic_item {
+    using key_type = object;
+
+    static object get(handle obj, handle key) {
+        PyObject *result = PyObject_GetItem(obj.ptr(), key.ptr());
+        if (!result) { throw error_already_set(); }
+        return reinterpret_steal<object>(result);
+    }
+
+    static void set(handle obj, handle key, handle val) {
+        if (PyObject_SetItem(obj.ptr(), key.ptr(), val.ptr()) != 0) { throw error_already_set(); }
+    }
+};
+
+struct sequence_item {
+    using key_type = size_t;
+
+    static object get(handle obj, size_t index) {
+        PyObject *result = PySequence_GetItem(obj.ptr(), static_cast<ssize_t>(index));
+        if (!result) { throw error_already_set(); }
+        return reinterpret_steal<object>(result);
+    }
+
+    static void set(handle obj, size_t index, handle val) {
+        // PySequence_SetItem does not steal a reference to 'val'
+        if (PySequence_SetItem(obj.ptr(), static_cast<ssize_t>(index), val.ptr()) != 0) {
+            throw error_already_set();
+        }
+    }
+};
+
+struct list_item {
+    using key_type = size_t;
+
+    static object get(handle obj, size_t index) {
+        PyObject *result = PyList_GetItem(obj.ptr(), static_cast<ssize_t>(index));
+        if (!result) { throw error_already_set(); }
+        return reinterpret_borrow<object>(result);
+    }
+
+    static void set(handle obj, size_t index, handle val) {
+        // PyList_SetItem steals a reference to 'val'
+        if (PyList_SetItem(obj.ptr(), static_cast<ssize_t>(index), val.inc_ref().ptr()) != 0) {
+            throw error_already_set();
+        }
+    }
+};
+
+struct tuple_item {
+    using key_type = size_t;
+
+    static object get(handle obj, size_t index) {
+        PyObject *result = PyTuple_GetItem(obj.ptr(), static_cast<ssize_t>(index));
+        if (!result) { throw error_already_set(); }
+        return reinterpret_borrow<object>(result);
+    }
+
+    static void set(handle obj, size_t index, handle val) {
+        // PyTuple_SetItem steals a reference to 'val'
+        if (PyTuple_SetItem(obj.ptr(), static_cast<ssize_t>(index), val.inc_ref().ptr()) != 0) {
+            throw error_already_set();
+        }
+    }
+};
+NAMESPACE_END(accessor_policies)
+
+/// STL iterator template used for tuple, list, sequence and dict
+template <typename Policy>
+class generic_iterator : public Policy {
+    using It = generic_iterator;
+
+public:
+    using difference_type = ssize_t;
+    using iterator_category = typename Policy::iterator_category;
+    using value_type = typename Policy::value_type;
+    using reference = typename Policy::reference;
+    using pointer = typename Policy::pointer;
+
+    generic_iterator() = default;
+    generic_iterator(handle seq, ssize_t index) : Policy(seq, index) { }
+
+    reference operator*() const { return Policy::dereference(); }
+    reference operator[](difference_type n) const { return *(*this + n); }
+    pointer operator->() const { return **this; }
+
+    It &operator++() { Policy::increment(); return *this; }
+    It operator++(int) { auto copy = *this; Policy::increment(); return copy; }
+    It &operator--() { Policy::decrement(); return *this; }
+    It operator--(int) { auto copy = *this; Policy::decrement(); return copy; }
+    It &operator+=(difference_type n) { Policy::advance(n); return *this; }
+    It &operator-=(difference_type n) { Policy::advance(-n); return *this; }
+
+    friend It operator+(const It &a, difference_type n) { auto copy = a; return copy += n; }
+    friend It operator+(difference_type n, const It &b) { return b + n; }
+    friend It operator-(const It &a, difference_type n) { auto copy = a; return copy -= n; }
+    friend difference_type operator-(const It &a, const It &b) { return a.distance_to(b); }
+
+    friend bool operator==(const It &a, const It &b) { return a.equal(b); }
+    friend bool operator!=(const It &a, const It &b) { return !(a == b); }
+    friend bool operator< (const It &a, const It &b) { return b - a > 0; }
+    friend bool operator> (const It &a, const It &b) { return b < a; }
+    friend bool operator>=(const It &a, const It &b) { return !(a < b); }
+    friend bool operator<=(const It &a, const It &b) { return !(a > b); }
+};
+
+NAMESPACE_BEGIN(iterator_policies)
+/// Quick proxy class needed to implement ``operator->`` for iterators which can't return pointers
+template <typename T>
+struct arrow_proxy {
+    T value;
+
+    arrow_proxy(T &&value) : value(std::move(value)) { }
+    T *operator->() const { return &value; }
+};
+
+/// Lightweight iterator policy using just a simple pointer: see ``PySequence_Fast_ITEMS``
+class sequence_fast_readonly {
+protected:
+    using iterator_category = std::random_access_iterator_tag;
+    using value_type = handle;
+    using reference = const handle;
+    using pointer = arrow_proxy<const handle>;
+
+    sequence_fast_readonly(handle obj, ssize_t n) : ptr(PySequence_Fast_ITEMS(obj.ptr()) + n) { }
+
+    reference dereference() const { return *ptr; }
+    void increment() { ++ptr; }
+    void decrement() { --ptr; }
+    void advance(ssize_t n) { ptr += n; }
+    bool equal(const sequence_fast_readonly &b) const { return ptr == b.ptr; }
+    ssize_t distance_to(const sequence_fast_readonly &b) const { return ptr - b.ptr; }
+
+private:
+    PyObject **ptr;
+};
+
+/// Full read and write access using the sequence protocol: see ``detail::sequence_accessor``
+class sequence_slow_readwrite {
+protected:
+    using iterator_category = std::random_access_iterator_tag;
+    using value_type = object;
+    using reference = sequence_accessor;
+    using pointer = arrow_proxy<const sequence_accessor>;
+
+    sequence_slow_readwrite(handle obj, ssize_t index) : obj(obj), index(index) { }
+
+    reference dereference() const { return {obj, static_cast<size_t>(index)}; }
+    void increment() { ++index; }
+    void decrement() { --index; }
+    void advance(ssize_t n) { index += n; }
+    bool equal(const sequence_slow_readwrite &b) const { return index == b.index; }
+    ssize_t distance_to(const sequence_slow_readwrite &b) const { return index - b.index; }
+
+private:
+    handle obj;
+    ssize_t index;
+};
+
+/// Python's dictionary protocol permits this to be a forward iterator
+class dict_readonly {
+protected:
+    using iterator_category = std::forward_iterator_tag;
+    using value_type = std::pair<handle, handle>;
+    using reference = const value_type;
+    using pointer = arrow_proxy<const value_type>;
+
+    dict_readonly() = default;
+    dict_readonly(handle obj, ssize_t pos) : obj(obj), pos(pos) { increment(); }
+
+    reference dereference() const { return {key, value}; }
+    void increment() { if (!PyDict_Next(obj.ptr(), &pos, &key, &value)) { pos = -1; } }
+    bool equal(const dict_readonly &b) const { return pos == b.pos; }
+
+private:
+    handle obj;
+    PyObject *key, *value;
+    ssize_t pos = -1;
+};
+NAMESPACE_END(iterator_policies)
+
+#if !defined(PYPY_VERSION)
+using tuple_iterator = generic_iterator<iterator_policies::sequence_fast_readonly>;
+using list_iterator = generic_iterator<iterator_policies::sequence_fast_readonly>;
+#else
+using tuple_iterator = generic_iterator<iterator_policies::sequence_slow_readwrite>;
+using list_iterator = generic_iterator<iterator_policies::sequence_slow_readwrite>;
+#endif
+
+using sequence_iterator = generic_iterator<iterator_policies::sequence_slow_readwrite>;
+using dict_iterator = generic_iterator<iterator_policies::dict_readonly>;
+
+inline bool PyIterable_Check(PyObject *obj) {
+    PyObject *iter = PyObject_GetIter(obj);
+    if (iter) {
+        Py_DECREF(iter);
+        return true;
+    } else {
+        PyErr_Clear();
+        return false;
+    }
+}
+
+inline bool PyNone_Check(PyObject *o) { return o == Py_None; }
+
+inline bool PyUnicode_Check_Permissive(PyObject *o) { return PyUnicode_Check(o) || PYBIND11_BYTES_CHECK(o); }
+
+class kwargs_proxy : public handle {
+public:
+    explicit kwargs_proxy(handle h) : handle(h) { }
+};
+
+class args_proxy : public handle {
+public:
+    explicit args_proxy(handle h) : handle(h) { }
+    kwargs_proxy operator*() const { return kwargs_proxy(*this); }
+};
+
+/// Python argument categories (using PEP 448 terms)
+template <typename T> using is_keyword = std::is_base_of<arg, T>;
+template <typename T> using is_s_unpacking = std::is_same<args_proxy, T>; // * unpacking
+template <typename T> using is_ds_unpacking = std::is_same<kwargs_proxy, T>; // ** unpacking
+template <typename T> using is_positional = satisfies_none_of<T,
+    is_keyword, is_s_unpacking, is_ds_unpacking
+>;
+template <typename T> using is_keyword_or_ds = satisfies_any_of<T, is_keyword, is_ds_unpacking>;
+
+// Call argument collector forward declarations
+template <return_value_policy policy = return_value_policy::automatic_reference>
+class simple_collector;
+template <return_value_policy policy = return_value_policy::automatic_reference>
+class unpacking_collector;
+
+NAMESPACE_END(detail)
+
+// TODO: After the deprecated constructors are removed, this macro can be simplified by
+//       inheriting ctors: `using Parent::Parent`. It's not an option right now because
+//       the `using` statement triggers the parent deprecation warning even if the ctor
+//       isn't even used.
+#define PYBIND11_OBJECT_COMMON(Name, Parent, CheckFun) \
+    public: \
+        PYBIND11_DEPRECATED("Use reinterpret_borrow<"#Name">() or reinterpret_steal<"#Name">()") \
+        Name(handle h, bool is_borrowed) : Parent(is_borrowed ? Parent(h, borrowed_t{}) : Parent(h, stolen_t{})) { } \
+        Name(handle h, borrowed_t) : Parent(h, borrowed_t{}) { } \
+        Name(handle h, stolen_t) : Parent(h, stolen_t{}) { } \
+        PYBIND11_DEPRECATED("Use py::isinstance<py::python_type>(obj) instead") \
+        bool check() const { return m_ptr != nullptr && (bool) CheckFun(m_ptr); } \
+        static bool check_(handle h) { return h.ptr() != nullptr && CheckFun(h.ptr()); }
+
+#define PYBIND11_OBJECT_CVT(Name, Parent, CheckFun, ConvertFun) \
+    PYBIND11_OBJECT_COMMON(Name, Parent, CheckFun) \
+    /* This is deliberately not 'explicit' to allow implicit conversion from object: */ \
+    Name(const object &o) \
+    : Parent(check_(o) ? o.inc_ref().ptr() : ConvertFun(o.ptr()), stolen_t{}) \
+    { if (!m_ptr) throw error_already_set(); } \
+    Name(object &&o) \
+    : Parent(check_(o) ? o.release().ptr() : ConvertFun(o.ptr()), stolen_t{}) \
+    { if (!m_ptr) throw error_already_set(); } \
+    template <typename Policy_> \
+    Name(const ::pybind11::detail::accessor<Policy_> &a) : Name(object(a)) { }
+
+#define PYBIND11_OBJECT(Name, Parent, CheckFun) \
+    PYBIND11_OBJECT_COMMON(Name, Parent, CheckFun) \
+    /* This is deliberately not 'explicit' to allow implicit conversion from object: */ \
+    Name(const object &o) : Parent(o) { } \
+    Name(object &&o) : Parent(std::move(o)) { }
+
+#define PYBIND11_OBJECT_DEFAULT(Name, Parent, CheckFun) \
+    PYBIND11_OBJECT(Name, Parent, CheckFun) \
+    Name() : Parent() { }
+
+/// \addtogroup pytypes
+/// @{
+
+/** \rst
+    Wraps a Python iterator so that it can also be used as a C++ input iterator
+
+    Caveat: copying an iterator does not (and cannot) clone the internal
+    state of the Python iterable. This also applies to the post-increment
+    operator. This iterator should only be used to retrieve the current
+    value using ``operator*()``.
+\endrst */
+class iterator : public object {
+public:
+    using iterator_category = std::input_iterator_tag;
+    using difference_type = ssize_t;
+    using value_type = handle;
+    using reference = const handle;
+    using pointer = const handle *;
+
+    PYBIND11_OBJECT_DEFAULT(iterator, object, PyIter_Check)
+
+    iterator& operator++() {
+        advance();
+        return *this;
+    }
+
+    iterator operator++(int) {
+        auto rv = *this;
+        advance();
+        return rv;
+    }
+
+    reference operator*() const {
+        if (m_ptr && !value.ptr()) {
+            auto& self = const_cast<iterator &>(*this);
+            self.advance();
+        }
+        return value;
+    }
+
+    pointer operator->() const { operator*(); return &value; }
+
+    /** \rst
+         The value which marks the end of the iteration. ``it == iterator::sentinel()``
+         is equivalent to catching ``StopIteration`` in Python.
+
+         .. code-block:: cpp
+
+             void foo(py::iterator it) {
+                 while (it != py::iterator::sentinel()) {
+                    // use `*it`
+                    ++it;
+                 }
+             }
+    \endrst */
+    static iterator sentinel() { return {}; }
+
+    friend bool operator==(const iterator &a, const iterator &b) { return a->ptr() == b->ptr(); }
+    friend bool operator!=(const iterator &a, const iterator &b) { return a->ptr() != b->ptr(); }
+
+private:
+    void advance() {
+        value = reinterpret_steal<object>(PyIter_Next(m_ptr));
+        if (PyErr_Occurred()) { throw error_already_set(); }
+    }
+
+private:
+    object value = {};
+};
+
+class iterable : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(iterable, object, detail::PyIterable_Check)
+};
+
+class bytes;
+
+class str : public object {
+public:
+    PYBIND11_OBJECT_CVT(str, object, detail::PyUnicode_Check_Permissive, raw_str)
+
+    str(const char *c, size_t n)
+        : object(PyUnicode_FromStringAndSize(c, (ssize_t) n), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate string object!");
+    }
+
+    // 'explicit' is explicitly omitted from the following constructors to allow implicit conversion to py::str from C++ string-like objects
+    str(const char *c = "")
+        : object(PyUnicode_FromString(c), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate string object!");
+    }
+
+    str(const std::string &s) : str(s.data(), s.size()) { }
+
+    explicit str(const bytes &b);
+
+    /** \rst
+        Return a string representation of the object. This is analogous to
+        the ``str()`` function in Python.
+    \endrst */
+    explicit str(handle h) : object(raw_str(h.ptr()), stolen_t{}) { }
+
+    operator std::string() const {
+        object temp = *this;
+        if (PyUnicode_Check(m_ptr)) {
+            temp = reinterpret_steal<object>(PyUnicode_AsUTF8String(m_ptr));
+            if (!temp)
+                pybind11_fail("Unable to extract string contents! (encoding issue)");
+        }
+        char *buffer;
+        ssize_t length;
+        if (PYBIND11_BYTES_AS_STRING_AND_SIZE(temp.ptr(), &buffer, &length))
+            pybind11_fail("Unable to extract string contents! (invalid type)");
+        return std::string(buffer, (size_t) length);
+    }
+
+    template <typename... Args>
+    str format(Args &&...args) const {
+        return attr("format")(std::forward<Args>(args)...);
+    }
+
+private:
+    /// Return string representation -- always returns a new reference, even if already a str
+    static PyObject *raw_str(PyObject *op) {
+        PyObject *str_value = PyObject_Str(op);
+#if PY_MAJOR_VERSION < 3
+        if (!str_value) throw error_already_set();
+        PyObject *unicode = PyUnicode_FromEncodedObject(str_value, "utf-8", nullptr);
+        Py_XDECREF(str_value); str_value = unicode;
+#endif
+        return str_value;
+    }
+};
+/// @} pytypes
+
+inline namespace literals {
+/** \rst
+    String literal version of `str`
+ \endrst */
+inline str operator"" _s(const char *s, size_t size) { return {s, size}; }
+}
+
+/// \addtogroup pytypes
+/// @{
+class bytes : public object {
+public:
+    PYBIND11_OBJECT(bytes, object, PYBIND11_BYTES_CHECK)
+
+    // Allow implicit conversion:
+    bytes(const char *c = "")
+        : object(PYBIND11_BYTES_FROM_STRING(c), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate bytes object!");
+    }
+
+    bytes(const char *c, size_t n)
+        : object(PYBIND11_BYTES_FROM_STRING_AND_SIZE(c, (ssize_t) n), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate bytes object!");
+    }
+
+    // Allow implicit conversion:
+    bytes(const std::string &s) : bytes(s.data(), s.size()) { }
+
+    explicit bytes(const pybind11::str &s);
+
+    operator std::string() const {
+        char *buffer;
+        ssize_t length;
+        if (PYBIND11_BYTES_AS_STRING_AND_SIZE(m_ptr, &buffer, &length))
+            pybind11_fail("Unable to extract bytes contents!");
+        return std::string(buffer, (size_t) length);
+    }
+};
+
+inline bytes::bytes(const pybind11::str &s) {
+    object temp = s;
+    if (PyUnicode_Check(s.ptr())) {
+        temp = reinterpret_steal<object>(PyUnicode_AsUTF8String(s.ptr()));
+        if (!temp)
+            pybind11_fail("Unable to extract string contents! (encoding issue)");
+    }
+    char *buffer;
+    ssize_t length;
+    if (PYBIND11_BYTES_AS_STRING_AND_SIZE(temp.ptr(), &buffer, &length))
+        pybind11_fail("Unable to extract string contents! (invalid type)");
+    auto obj = reinterpret_steal<object>(PYBIND11_BYTES_FROM_STRING_AND_SIZE(buffer, length));
+    if (!obj)
+        pybind11_fail("Could not allocate bytes object!");
+    m_ptr = obj.release().ptr();
+}
+
+inline str::str(const bytes& b) {
+    char *buffer;
+    ssize_t length;
+    if (PYBIND11_BYTES_AS_STRING_AND_SIZE(b.ptr(), &buffer, &length))
+        pybind11_fail("Unable to extract bytes contents!");
+    auto obj = reinterpret_steal<object>(PyUnicode_FromStringAndSize(buffer, (ssize_t) length));
+    if (!obj)
+        pybind11_fail("Could not allocate string object!");
+    m_ptr = obj.release().ptr();
+}
+
+class none : public object {
+public:
+    PYBIND11_OBJECT(none, object, detail::PyNone_Check)
+    none() : object(Py_None, borrowed_t{}) { }
+};
+
+class bool_ : public object {
+public:
+    PYBIND11_OBJECT_CVT(bool_, object, PyBool_Check, raw_bool)
+    bool_() : object(Py_False, borrowed_t{}) { }
+    // Allow implicit conversion from and to `bool`:
+    bool_(bool value) : object(value ? Py_True : Py_False, borrowed_t{}) { }
+    operator bool() const { return m_ptr && PyLong_AsLong(m_ptr) != 0; }
+
+private:
+    /// Return the truth value of an object -- always returns a new reference
+    static PyObject *raw_bool(PyObject *op) {
+        const auto value = PyObject_IsTrue(op);
+        if (value == -1) return nullptr;
+        return handle(value ? Py_True : Py_False).inc_ref().ptr();
+    }
+};
+
+NAMESPACE_BEGIN(detail)
+// Converts a value to the given unsigned type.  If an error occurs, you get back (Unsigned) -1;
+// otherwise you get back the unsigned long or unsigned long long value cast to (Unsigned).
+// (The distinction is critically important when casting a returned -1 error value to some other
+// unsigned type: (A)-1 != (B)-1 when A and B are unsigned types of different sizes).
+template <typename Unsigned>
+Unsigned as_unsigned(PyObject *o) {
+    if (sizeof(Unsigned) <= sizeof(unsigned long)
+#if PY_VERSION_HEX < 0x03000000
+            || PyInt_Check(o)
+#endif
+    ) {
+        unsigned long v = PyLong_AsUnsignedLong(o);
+        return v == (unsigned long) -1 && PyErr_Occurred() ? (Unsigned) -1 : (Unsigned) v;
+    }
+    else {
+        unsigned long long v = PyLong_AsUnsignedLongLong(o);
+        return v == (unsigned long long) -1 && PyErr_Occurred() ? (Unsigned) -1 : (Unsigned) v;
+    }
+}
+NAMESPACE_END(detail)
+
+class int_ : public object {
+public:
+    PYBIND11_OBJECT_CVT(int_, object, PYBIND11_LONG_CHECK, PyNumber_Long)
+    int_() : object(PyLong_FromLong(0), stolen_t{}) { }
+    // Allow implicit conversion from C++ integral types:
+    template <typename T,
+              detail::enable_if_t<std::is_integral<T>::value, int> = 0>
+    int_(T value) {
+        if (sizeof(T) <= sizeof(long)) {
+            if (std::is_signed<T>::value)
+                m_ptr = PyLong_FromLong((long) value);
+            else
+                m_ptr = PyLong_FromUnsignedLong((unsigned long) value);
+        } else {
+            if (std::is_signed<T>::value)
+                m_ptr = PyLong_FromLongLong((long long) value);
+            else
+                m_ptr = PyLong_FromUnsignedLongLong((unsigned long long) value);
+        }
+        if (!m_ptr) pybind11_fail("Could not allocate int object!");
+    }
+
+    template <typename T,
+              detail::enable_if_t<std::is_integral<T>::value, int> = 0>
+    operator T() const {
+        return std::is_unsigned<T>::value
+            ? detail::as_unsigned<T>(m_ptr)
+            : sizeof(T) <= sizeof(long)
+              ? (T) PyLong_AsLong(m_ptr)
+              : (T) PYBIND11_LONG_AS_LONGLONG(m_ptr);
+    }
+};
+
+class float_ : public object {
+public:
+    PYBIND11_OBJECT_CVT(float_, object, PyFloat_Check, PyNumber_Float)
+    // Allow implicit conversion from float/double:
+    float_(float value) : object(PyFloat_FromDouble((double) value), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate float object!");
+    }
+    float_(double value = .0) : object(PyFloat_FromDouble((double) value), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate float object!");
+    }
+    operator float() const { return (float) PyFloat_AsDouble(m_ptr); }
+    operator double() const { return (double) PyFloat_AsDouble(m_ptr); }
+};
+
+class weakref : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(weakref, object, PyWeakref_Check)
+    explicit weakref(handle obj, handle callback = {})
+        : object(PyWeakref_NewRef(obj.ptr(), callback.ptr()), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate weak reference!");
+    }
+};
+
+class slice : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(slice, object, PySlice_Check)
+    slice(ssize_t start_, ssize_t stop_, ssize_t step_) {
+        int_ start(start_), stop(stop_), step(step_);
+        m_ptr = PySlice_New(start.ptr(), stop.ptr(), step.ptr());
+        if (!m_ptr) pybind11_fail("Could not allocate slice object!");
+    }
+    bool compute(size_t length, size_t *start, size_t *stop, size_t *step,
+                 size_t *slicelength) const {
+        return PySlice_GetIndicesEx((PYBIND11_SLICE_OBJECT *) m_ptr,
+                                    (ssize_t) length, (ssize_t *) start,
+                                    (ssize_t *) stop, (ssize_t *) step,
+                                    (ssize_t *) slicelength) == 0;
+    }
+};
+
+class capsule : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(capsule, object, PyCapsule_CheckExact)
+    PYBIND11_DEPRECATED("Use reinterpret_borrow<capsule>() or reinterpret_steal<capsule>()")
+    capsule(PyObject *ptr, bool is_borrowed) : object(is_borrowed ? object(ptr, borrowed_t{}) : object(ptr, stolen_t{})) { }
+
+    explicit capsule(const void *value, const char *name = nullptr, void (*destructor)(PyObject *) = nullptr)
+        : object(PyCapsule_New(const_cast<void *>(value), name, destructor), stolen_t{}) {
+        if (!m_ptr)
+            pybind11_fail("Could not allocate capsule object!");
+    }
+
+    PYBIND11_DEPRECATED("Please pass a destructor that takes a void pointer as input")
+    capsule(const void *value, void (*destruct)(PyObject *))
+        : object(PyCapsule_New(const_cast<void*>(value), nullptr, destruct), stolen_t{}) {
+        if (!m_ptr)
+            pybind11_fail("Could not allocate capsule object!");
+    }
+
+    capsule(const void *value, void (*destructor)(void *)) {
+        m_ptr = PyCapsule_New(const_cast<void *>(value), nullptr, [](PyObject *o) {
+            auto destructor = reinterpret_cast<void (*)(void *)>(PyCapsule_GetContext(o));
+            void *ptr = PyCapsule_GetPointer(o, nullptr);
+            destructor(ptr);
+        });
+
+        if (!m_ptr)
+            pybind11_fail("Could not allocate capsule object!");
+
+        if (PyCapsule_SetContext(m_ptr, (void *) destructor) != 0)
+            pybind11_fail("Could not set capsule context!");
+    }
+
+    capsule(void (*destructor)()) {
+        m_ptr = PyCapsule_New(reinterpret_cast<void *>(destructor), nullptr, [](PyObject *o) {
+            auto destructor = reinterpret_cast<void (*)()>(PyCapsule_GetPointer(o, nullptr));
+            destructor();
+        });
+
+        if (!m_ptr)
+            pybind11_fail("Could not allocate capsule object!");
+    }
+
+    template <typename T> operator T *() const {
+        auto name = this->name();
+        T * result = static_cast<T *>(PyCapsule_GetPointer(m_ptr, name));
+        if (!result) pybind11_fail("Unable to extract capsule contents!");
+        return result;
+    }
+
+    const char *name() const { return PyCapsule_GetName(m_ptr); }
+};
+
+class tuple : public object {
+public:
+    PYBIND11_OBJECT_CVT(tuple, object, PyTuple_Check, PySequence_Tuple)
+    explicit tuple(size_t size = 0) : object(PyTuple_New((ssize_t) size), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate tuple object!");
+    }
+    size_t size() const { return (size_t) PyTuple_Size(m_ptr); }
+    detail::tuple_accessor operator[](size_t index) const { return {*this, index}; }
+    detail::tuple_iterator begin() const { return {*this, 0}; }
+    detail::tuple_iterator end() const { return {*this, PyTuple_GET_SIZE(m_ptr)}; }
+};
+
+class dict : public object {
+public:
+    PYBIND11_OBJECT_CVT(dict, object, PyDict_Check, raw_dict)
+    dict() : object(PyDict_New(), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate dict object!");
+    }
+    template <typename... Args,
+              typename = detail::enable_if_t<detail::all_of<detail::is_keyword_or_ds<Args>...>::value>,
+              // MSVC workaround: it can't compile an out-of-line definition, so defer the collector
+              typename collector = detail::deferred_t<detail::unpacking_collector<>, Args...>>
+    explicit dict(Args &&...args) : dict(collector(std::forward<Args>(args)...).kwargs()) { }
+
+    size_t size() const { return (size_t) PyDict_Size(m_ptr); }
+    detail::dict_iterator begin() const { return {*this, 0}; }
+    detail::dict_iterator end() const { return {}; }
+    void clear() const { PyDict_Clear(ptr()); }
+    bool contains(handle key) const { return PyDict_Contains(ptr(), key.ptr()) == 1; }
+    bool contains(const char *key) const { return PyDict_Contains(ptr(), pybind11::str(key).ptr()) == 1; }
+
+private:
+    /// Call the `dict` Python type -- always returns a new reference
+    static PyObject *raw_dict(PyObject *op) {
+        if (PyDict_Check(op))
+            return handle(op).inc_ref().ptr();
+        return PyObject_CallFunctionObjArgs((PyObject *) &PyDict_Type, op, nullptr);
+    }
+};
+
+class sequence : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(sequence, object, PySequence_Check)
+    size_t size() const { return (size_t) PySequence_Size(m_ptr); }
+    detail::sequence_accessor operator[](size_t index) const { return {*this, index}; }
+    detail::sequence_iterator begin() const { return {*this, 0}; }
+    detail::sequence_iterator end() const { return {*this, PySequence_Size(m_ptr)}; }
+};
+
+class list : public object {
+public:
+    PYBIND11_OBJECT_CVT(list, object, PyList_Check, PySequence_List)
+    explicit list(size_t size = 0) : object(PyList_New((ssize_t) size), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate list object!");
+    }
+    size_t size() const { return (size_t) PyList_Size(m_ptr); }
+    detail::list_accessor operator[](size_t index) const { return {*this, index}; }
+    detail::list_iterator begin() const { return {*this, 0}; }
+    detail::list_iterator end() const { return {*this, PyList_GET_SIZE(m_ptr)}; }
+    template <typename T> void append(T &&val) const {
+        PyList_Append(m_ptr, detail::object_or_cast(std::forward<T>(val)).ptr());
+    }
+};
+
+class args : public tuple { PYBIND11_OBJECT_DEFAULT(args, tuple, PyTuple_Check) };
+class kwargs : public dict { PYBIND11_OBJECT_DEFAULT(kwargs, dict, PyDict_Check)  };
+
+class set : public object {
+public:
+    PYBIND11_OBJECT_CVT(set, object, PySet_Check, PySet_New)
+    set() : object(PySet_New(nullptr), stolen_t{}) {
+        if (!m_ptr) pybind11_fail("Could not allocate set object!");
+    }
+    size_t size() const { return (size_t) PySet_Size(m_ptr); }
+    template <typename T> bool add(T &&val) const {
+        return PySet_Add(m_ptr, detail::object_or_cast(std::forward<T>(val)).ptr()) == 0;
+    }
+    void clear() const { PySet_Clear(m_ptr); }
+};
+
+class function : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(function, object, PyCallable_Check)
+    handle cpp_function() const {
+        handle fun = detail::get_function(m_ptr);
+        if (fun && PyCFunction_Check(fun.ptr()))
+            return fun;
+        return handle();
+    }
+    bool is_cpp_function() const { return (bool) cpp_function(); }
+};
+
+class buffer : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(buffer, object, PyObject_CheckBuffer)
+
+    buffer_info request(bool writable = false) {
+        int flags = PyBUF_STRIDES | PyBUF_FORMAT;
+        if (writable) flags |= PyBUF_WRITABLE;
+        Py_buffer *view = new Py_buffer();
+        if (PyObject_GetBuffer(m_ptr, view, flags) != 0) {
+            delete view;
+            throw error_already_set();
+        }
+        return buffer_info(view);
+    }
+};
+
+class memoryview : public object {
+public:
+    explicit memoryview(const buffer_info& info) {
+        static Py_buffer buf { };
+        // Py_buffer uses signed sizes, strides and shape!..
+        static std::vector<Py_ssize_t> py_strides { };
+        static std::vector<Py_ssize_t> py_shape { };
+        buf.buf = info.ptr;
+        buf.itemsize = info.itemsize;
+        buf.format = const_cast<char *>(info.format.c_str());
+        buf.ndim = (int) info.ndim;
+        buf.len = info.size;
+        py_strides.clear();
+        py_shape.clear();
+        for (size_t i = 0; i < (size_t) info.ndim; ++i) {
+            py_strides.push_back(info.strides[i]);
+            py_shape.push_back(info.shape[i]);
+        }
+        buf.strides = py_strides.data();
+        buf.shape = py_shape.data();
+        buf.suboffsets = nullptr;
+        buf.readonly = false;
+        buf.internal = nullptr;
+
+        m_ptr = PyMemoryView_FromBuffer(&buf);
+        if (!m_ptr)
+            pybind11_fail("Unable to create memoryview from buffer descriptor");
+    }
+
+    PYBIND11_OBJECT_CVT(memoryview, object, PyMemoryView_Check, PyMemoryView_FromObject)
+};
+/// @} pytypes
+
+/// \addtogroup python_builtins
+/// @{
+inline size_t len(handle h) {
+    ssize_t result = PyObject_Length(h.ptr());
+    if (result < 0)
+        pybind11_fail("Unable to compute length of object");
+    return (size_t) result;
+}
+
+inline str repr(handle h) {
+    PyObject *str_value = PyObject_Repr(h.ptr());
+    if (!str_value) throw error_already_set();
+#if PY_MAJOR_VERSION < 3
+    PyObject *unicode = PyUnicode_FromEncodedObject(str_value, "utf-8", nullptr);
+    Py_XDECREF(str_value); str_value = unicode;
+    if (!str_value) throw error_already_set();
+#endif
+    return reinterpret_steal<str>(str_value);
+}
+
+inline iterator iter(handle obj) {
+    PyObject *result = PyObject_GetIter(obj.ptr());
+    if (!result) { throw error_already_set(); }
+    return reinterpret_steal<iterator>(result);
+}
+/// @} python_builtins
+
+NAMESPACE_BEGIN(detail)
+template <typename D> iterator object_api<D>::begin() const { return iter(derived()); }
+template <typename D> iterator object_api<D>::end() const { return iterator::sentinel(); }
+template <typename D> item_accessor object_api<D>::operator[](handle key) const {
+    return {derived(), reinterpret_borrow<object>(key)};
+}
+template <typename D> item_accessor object_api<D>::operator[](const char *key) const {
+    return {derived(), pybind11::str(key)};
+}
+template <typename D> obj_attr_accessor object_api<D>::attr(handle key) const {
+    return {derived(), reinterpret_borrow<object>(key)};
+}
+template <typename D> str_attr_accessor object_api<D>::attr(const char *key) const {
+    return {derived(), key};
+}
+template <typename D> args_proxy object_api<D>::operator*() const {
+    return args_proxy(derived().ptr());
+}
+template <typename D> template <typename T> bool object_api<D>::contains(T &&item) const {
+    return attr("__contains__")(std::forward<T>(item)).template cast<bool>();
+}
+
+template <typename D>
+pybind11::str object_api<D>::str() const { return pybind11::str(derived()); }
+
+template <typename D>
+str_attr_accessor object_api<D>::doc() const { return attr("__doc__"); }
+
+template <typename D>
+handle object_api<D>::get_type() const { return (PyObject *) Py_TYPE(derived().ptr()); }
+
+NAMESPACE_END(detail)
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/stl.h b/ml/dlib/dlib/external/pybind11/include/pybind11/stl.h
new file mode 100644
index 000000000..90eb7ea2e
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/stl.h
@@ -0,0 +1,370 @@
+/*
+    pybind11/stl.h: Transparent conversion for STL data types
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include <set>
+#include <unordered_set>
+#include <map>
+#include <unordered_map>
+#include <iostream>
+#include <list>
+#include <valarray>
+
+#if defined(_MSC_VER)
+#pragma warning(push)
+#pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#endif
+
+#ifdef __has_include
+// std::optional (but including it in c++14 mode isn't allowed)
+#  if defined(PYBIND11_CPP17) && __has_include(<optional>)
+#    include <optional>
+#    define PYBIND11_HAS_OPTIONAL 1
+#  endif
+// std::experimental::optional (but not allowed in c++11 mode)
+#  if defined(PYBIND11_CPP14) && (__has_include(<experimental/optional>) && \
+                                 !__has_include(<optional>))
+#    include <experimental/optional>
+#    define PYBIND11_HAS_EXP_OPTIONAL 1
+#  endif
+// std::variant
+#  if defined(PYBIND11_CPP17) && __has_include(<variant>)
+#    include <variant>
+#    define PYBIND11_HAS_VARIANT 1
+#  endif
+#elif defined(_MSC_VER) && defined(PYBIND11_CPP17)
+#  include <optional>
+#  include <variant>
+#  define PYBIND11_HAS_OPTIONAL 1
+#  define PYBIND11_HAS_VARIANT 1
+#endif
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/// Extracts an const lvalue reference or rvalue reference for U based on the type of T (e.g. for
+/// forwarding a container element).  Typically used indirect via forwarded_type(), below.
+template <typename T, typename U>
+using forwarded_type = conditional_t<
+    std::is_lvalue_reference<T>::value, remove_reference_t<U> &, remove_reference_t<U> &&>;
+
+/// Forwards a value U as rvalue or lvalue according to whether T is rvalue or lvalue; typically
+/// used for forwarding a container's elements.
+template <typename T, typename U>
+forwarded_type<T, U> forward_like(U &&u) {
+    return std::forward<detail::forwarded_type<T, U>>(std::forward<U>(u));
+}
+
+template <typename Type, typename Key> struct set_caster {
+    using type = Type;
+    using key_conv = make_caster<Key>;
+
+    bool load(handle src, bool convert) {
+        if (!isinstance<pybind11::set>(src))
+            return false;
+        auto s = reinterpret_borrow<pybind11::set>(src);
+        value.clear();
+        for (auto entry : s) {
+            key_conv conv;
+            if (!conv.load(entry, convert))
+                return false;
+            value.insert(cast_op<Key &&>(std::move(conv)));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        pybind11::set s;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(key_conv::cast(forward_like<T>(value), policy, parent));
+            if (!value_ || !s.add(value_))
+                return handle();
+        }
+        return s.release();
+    }
+
+    PYBIND11_TYPE_CASTER(type, _("Set[") + key_conv::name() + _("]"));
+};
+
+template <typename Type, typename Key, typename Value> struct map_caster {
+    using key_conv   = make_caster<Key>;
+    using value_conv = make_caster<Value>;
+
+    bool load(handle src, bool convert) {
+        if (!isinstance<dict>(src))
+            return false;
+        auto d = reinterpret_borrow<dict>(src);
+        value.clear();
+        for (auto it : d) {
+            key_conv kconv;
+            value_conv vconv;
+            if (!kconv.load(it.first.ptr(), convert) ||
+                !vconv.load(it.second.ptr(), convert))
+                return false;
+            value.emplace(cast_op<Key &&>(std::move(kconv)), cast_op<Value &&>(std::move(vconv)));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        dict d;
+        for (auto &&kv : src) {
+            auto key = reinterpret_steal<object>(key_conv::cast(forward_like<T>(kv.first), policy, parent));
+            auto value = reinterpret_steal<object>(value_conv::cast(forward_like<T>(kv.second), policy, parent));
+            if (!key || !value)
+                return handle();
+            d[key] = value;
+        }
+        return d.release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type, _("Dict[") + key_conv::name() + _(", ") + value_conv::name() + _("]"));
+};
+
+template <typename Type, typename Value> struct list_caster {
+    using value_conv = make_caster<Value>;
+
+    bool load(handle src, bool convert) {
+        if (!isinstance<sequence>(src))
+            return false;
+        auto s = reinterpret_borrow<sequence>(src);
+        value.clear();
+        reserve_maybe(s, &value);
+        for (auto it : s) {
+            value_conv conv;
+            if (!conv.load(it, convert))
+                return false;
+            value.push_back(cast_op<Value &&>(std::move(conv)));
+        }
+        return true;
+    }
+
+private:
+    template <typename T = Type,
+              enable_if_t<std::is_same<decltype(std::declval<T>().reserve(0)), void>::value, int> = 0>
+    void reserve_maybe(sequence s, Type *) { value.reserve(s.size()); }
+    void reserve_maybe(sequence, void *) { }
+
+public:
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        list l(src.size());
+        size_t index = 0;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(value_conv::cast(forward_like<T>(value), policy, parent));
+            if (!value_)
+                return handle();
+            PyList_SET_ITEM(l.ptr(), (ssize_t) index++, value_.release().ptr()); // steals a reference
+        }
+        return l.release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type, _("List[") + value_conv::name() + _("]"));
+};
+
+template <typename Type, typename Alloc> struct type_caster<std::vector<Type, Alloc>>
+ : list_caster<std::vector<Type, Alloc>, Type> { };
+
+template <typename Type, typename Alloc> struct type_caster<std::list<Type, Alloc>>
+ : list_caster<std::list<Type, Alloc>, Type> { };
+
+template <typename ArrayType, typename Value, bool Resizable, size_t Size = 0> struct array_caster {
+    using value_conv = make_caster<Value>;
+
+private:
+    template <bool R = Resizable>
+    bool require_size(enable_if_t<R, size_t> size) {
+        if (value.size() != size)
+            value.resize(size);
+        return true;
+    }
+    template <bool R = Resizable>
+    bool require_size(enable_if_t<!R, size_t> size) {
+        return size == Size;
+    }
+
+public:
+    bool load(handle src, bool convert) {
+        if (!isinstance<list>(src))
+            return false;
+        auto l = reinterpret_borrow<list>(src);
+        if (!require_size(l.size()))
+            return false;
+        size_t ctr = 0;
+        for (auto it : l) {
+            value_conv conv;
+            if (!conv.load(it, convert))
+                return false;
+            value[ctr++] = cast_op<Value &&>(std::move(conv));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        list l(src.size());
+        size_t index = 0;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(value_conv::cast(forward_like<T>(value), policy, parent));
+            if (!value_)
+                return handle();
+            PyList_SET_ITEM(l.ptr(), (ssize_t) index++, value_.release().ptr()); // steals a reference
+        }
+        return l.release();
+    }
+
+    PYBIND11_TYPE_CASTER(ArrayType, _("List[") + value_conv::name() + _<Resizable>(_(""), _("[") + _<Size>() + _("]")) + _("]"));
+};
+
+template <typename Type, size_t Size> struct type_caster<std::array<Type, Size>>
+ : array_caster<std::array<Type, Size>, Type, false, Size> { };
+
+template <typename Type> struct type_caster<std::valarray<Type>>
+ : array_caster<std::valarray<Type>, Type, true> { };
+
+template <typename Key, typename Compare, typename Alloc> struct type_caster<std::set<Key, Compare, Alloc>>
+  : set_caster<std::set<Key, Compare, Alloc>, Key> { };
+
+template <typename Key, typename Hash, typename Equal, typename Alloc> struct type_caster<std::unordered_set<Key, Hash, Equal, Alloc>>
+  : set_caster<std::unordered_set<Key, Hash, Equal, Alloc>, Key> { };
+
+template <typename Key, typename Value, typename Compare, typename Alloc> struct type_caster<std::map<Key, Value, Compare, Alloc>>
+  : map_caster<std::map<Key, Value, Compare, Alloc>, Key, Value> { };
+
+template <typename Key, typename Value, typename Hash, typename Equal, typename Alloc> struct type_caster<std::unordered_map<Key, Value, Hash, Equal, Alloc>>
+  : map_caster<std::unordered_map<Key, Value, Hash, Equal, Alloc>, Key, Value> { };
+
+// This type caster is intended to be used for std::optional and std::experimental::optional
+template<typename T> struct optional_caster {
+    using value_conv = make_caster<typename T::value_type>;
+
+    template <typename T_>
+    static handle cast(T_ &&src, return_value_policy policy, handle parent) {
+        if (!src)
+            return none().inc_ref();
+        return value_conv::cast(*std::forward<T_>(src), policy, parent);
+    }
+
+    bool load(handle src, bool convert) {
+        if (!src) {
+            return false;
+        } else if (src.is_none()) {
+            return true;  // default-constructed value is already empty
+        }
+        value_conv inner_caster;
+        if (!inner_caster.load(src, convert))
+            return false;
+
+        value.emplace(cast_op<typename T::value_type &&>(std::move(inner_caster)));
+        return true;
+    }
+
+    PYBIND11_TYPE_CASTER(T, _("Optional[") + value_conv::name() + _("]"));
+};
+
+#if PYBIND11_HAS_OPTIONAL
+template<typename T> struct type_caster<std::optional<T>>
+    : public optional_caster<std::optional<T>> {};
+
+template<> struct type_caster<std::nullopt_t>
+    : public void_caster<std::nullopt_t> {};
+#endif
+
+#if PYBIND11_HAS_EXP_OPTIONAL
+template<typename T> struct type_caster<std::experimental::optional<T>>
+    : public optional_caster<std::experimental::optional<T>> {};
+
+template<> struct type_caster<std::experimental::nullopt_t>
+    : public void_caster<std::experimental::nullopt_t> {};
+#endif
+
+/// Visit a variant and cast any found type to Python
+struct variant_caster_visitor {
+    return_value_policy policy;
+    handle parent;
+
+    using result_type = handle; // required by boost::variant in C++11
+
+    template <typename T>
+    result_type operator()(T &&src) const {
+        return make_caster<T>::cast(std::forward<T>(src), policy, parent);
+    }
+};
+
+/// Helper class which abstracts away variant's `visit` function. `std::variant` and similar
+/// `namespace::variant` types which provide a `namespace::visit()` function are handled here
+/// automatically using argument-dependent lookup. Users can provide specializations for other
+/// variant-like classes, e.g. `boost::variant` and `boost::apply_visitor`.
+template <template<typename...> class Variant>
+struct visit_helper {
+    template <typename... Args>
+    static auto call(Args &&...args) -> decltype(visit(std::forward<Args>(args)...)) {
+        return visit(std::forward<Args>(args)...);
+    }
+};
+
+/// Generic variant caster
+template <typename Variant> struct variant_caster;
+
+template <template<typename...> class V, typename... Ts>
+struct variant_caster<V<Ts...>> {
+    static_assert(sizeof...(Ts) > 0, "Variant must consist of at least one alternative.");
+
+    template <typename U, typename... Us>
+    bool load_alternative(handle src, bool convert, type_list<U, Us...>) {
+        auto caster = make_caster<U>();
+        if (caster.load(src, convert)) {
+            value = cast_op<U>(caster);
+            return true;
+        }
+        return load_alternative(src, convert, type_list<Us...>{});
+    }
+
+    bool load_alternative(handle, bool, type_list<>) { return false; }
+
+    bool load(handle src, bool convert) {
+        // Do a first pass without conversions to improve constructor resolution.
+        // E.g. `py::int_(1).cast<variant<double, int>>()` needs to fill the `int`
+        // slot of the variant. Without two-pass loading `double` would be filled
+        // because it appears first and a conversion is possible.
+        if (convert && load_alternative(src, false, type_list<Ts...>{}))
+            return true;
+        return load_alternative(src, convert, type_list<Ts...>{});
+    }
+
+    template <typename Variant>
+    static handle cast(Variant &&src, return_value_policy policy, handle parent) {
+        return visit_helper<V>::call(variant_caster_visitor{policy, parent},
+                                     std::forward<Variant>(src));
+    }
+
+    using Type = V<Ts...>;
+    PYBIND11_TYPE_CASTER(Type, _("Union[") + detail::concat(make_caster<Ts>::name()...) + _("]"));
+};
+
+#if PYBIND11_HAS_VARIANT
+template <typename... Ts>
+struct type_caster<std::variant<Ts...>> : variant_caster<std::variant<Ts...>> { };
+#endif
+NAMESPACE_END(detail)
+
+inline std::ostream &operator<<(std::ostream &os, const handle &obj) {
+    os << (std::string) str(obj);
+    return os;
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
+
+#if defined(_MSC_VER)
+#pragma warning(pop)
+#endif
diff --git a/ml/dlib/dlib/external/pybind11/include/pybind11/stl_bind.h b/ml/dlib/dlib/external/pybind11/include/pybind11/stl_bind.h
new file mode 100644
index 000000000..38dd68f69
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/include/pybind11/stl_bind.h
@@ -0,0 +1,599 @@
+/*
+    pybind11/std_bind.h: Binding generators for STL data types
+
+    Copyright (c) 2016 Sergey Lyskov and Wenzel Jakob
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+#include "operators.h"
+
+#include <algorithm>
+#include <sstream>
+
+NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+NAMESPACE_BEGIN(detail)
+
+/* SFINAE helper class used by 'is_comparable */
+template <typename T>  struct container_traits {
+    template <typename T2> static std::true_type test_comparable(decltype(std::declval<const T2 &>() == std::declval<const T2 &>())*);
+    template <typename T2> static std::false_type test_comparable(...);
+    template <typename T2> static std::true_type test_value(typename T2::value_type *);
+    template <typename T2> static std::false_type test_value(...);
+    template <typename T2> static std::true_type test_pair(typename T2::first_type *, typename T2::second_type *);
+    template <typename T2> static std::false_type test_pair(...);
+
+    static constexpr const bool is_comparable = std::is_same<std::true_type, decltype(test_comparable<T>(nullptr))>::value;
+    static constexpr const bool is_pair = std::is_same<std::true_type, decltype(test_pair<T>(nullptr, nullptr))>::value;
+    static constexpr const bool is_vector = std::is_same<std::true_type, decltype(test_value<T>(nullptr))>::value;
+    static constexpr const bool is_element = !is_pair && !is_vector;
+};
+
+/* Default: is_comparable -> std::false_type */
+template <typename T, typename SFINAE = void>
+struct is_comparable : std::false_type { };
+
+/* For non-map data structures, check whether operator== can be instantiated */
+template <typename T>
+struct is_comparable<
+    T, enable_if_t<container_traits<T>::is_element &&
+                   container_traits<T>::is_comparable>>
+    : std::true_type { };
+
+/* For a vector/map data structure, recursively check the value type (which is std::pair for maps) */
+template <typename T>
+struct is_comparable<T, enable_if_t<container_traits<T>::is_vector>> {
+    static constexpr const bool value =
+        is_comparable<typename T::value_type>::value;
+};
+
+/* For pairs, recursively check the two data types */
+template <typename T>
+struct is_comparable<T, enable_if_t<container_traits<T>::is_pair>> {
+    static constexpr const bool value =
+        is_comparable<typename T::first_type>::value &&
+        is_comparable<typename T::second_type>::value;
+};
+
+/* Fallback functions */
+template <typename, typename, typename... Args> void vector_if_copy_constructible(const Args &...) { }
+template <typename, typename, typename... Args> void vector_if_equal_operator(const Args &...) { }
+template <typename, typename, typename... Args> void vector_if_insertion_operator(const Args &...) { }
+template <typename, typename, typename... Args> void vector_modifiers(const Args &...) { }
+
+template<typename Vector, typename Class_>
+void vector_if_copy_constructible(enable_if_t<is_copy_constructible<Vector>::value, Class_> &cl) {
+    cl.def(init<const Vector &>(), "Copy constructor");
+}
+
+template<typename Vector, typename Class_>
+void vector_if_equal_operator(enable_if_t<is_comparable<Vector>::value, Class_> &cl) {
+    using T = typename Vector::value_type;
+
+    cl.def(self == self);
+    cl.def(self != self);
+
+    cl.def("count",
+        [](const Vector &v, const T &x) {
+            return std::count(v.begin(), v.end(), x);
+        },
+        arg("x"),
+        "Return the number of times ``x`` appears in the list"
+    );
+
+    cl.def("remove", [](Vector &v, const T &x) {
+            auto p = std::find(v.begin(), v.end(), x);
+            if (p != v.end())
+                v.erase(p);
+            else
+                throw value_error();
+        },
+        arg("x"),
+        "Remove the first item from the list whose value is x. "
+        "It is an error if there is no such item."
+    );
+
+    cl.def("__contains__",
+        [](const Vector &v, const T &x) {
+            return std::find(v.begin(), v.end(), x) != v.end();
+        },
+        arg("x"),
+        "Return true the container contains ``x``"
+    );
+}
+
+// Vector modifiers -- requires a copyable vector_type:
+// (Technically, some of these (pop and __delitem__) don't actually require copyability, but it seems
+// silly to allow deletion but not insertion, so include them here too.)
+template <typename Vector, typename Class_>
+void vector_modifiers(enable_if_t<is_copy_constructible<typename Vector::value_type>::value, Class_> &cl) {
+    using T = typename Vector::value_type;
+    using SizeType = typename Vector::size_type;
+    using DiffType = typename Vector::difference_type;
+
+    cl.def("append",
+           [](Vector &v, const T &value) { v.push_back(value); },
+           arg("x"),
+           "Add an item to the end of the list");
+
+    cl.def(init([](iterable it) {
+        auto v = std::unique_ptr<Vector>(new Vector());
+        v->reserve(len(it));
+        for (handle h : it)
+           v->push_back(h.cast<T>());
+        return v.release();
+    }));
+
+    cl.def("extend",
+       [](Vector &v, const Vector &src) {
+           v.insert(v.end(), src.begin(), src.end());
+       },
+       arg("L"),
+       "Extend the list by appending all the items in the given list"
+    );
+
+    cl.def("insert",
+        [](Vector &v, SizeType i, const T &x) {
+            if (i > v.size())
+                throw index_error();
+            v.insert(v.begin() + (DiffType) i, x);
+        },
+        arg("i") , arg("x"),
+        "Insert an item at a given position."
+    );
+
+    cl.def("pop",
+        [](Vector &v) {
+            if (v.empty())
+                throw index_error();
+            T t = v.back();
+            v.pop_back();
+            return t;
+        },
+        "Remove and return the last item"
+    );
+
+    cl.def("pop",
+        [](Vector &v, SizeType i) {
+            if (i >= v.size())
+                throw index_error();
+            T t = v[i];
+            v.erase(v.begin() + (DiffType) i);
+            return t;
+        },
+        arg("i"),
+        "Remove and return the item at index ``i``"
+    );
+
+    cl.def("__setitem__",
+        [](Vector &v, SizeType i, const T &t) {
+            if (i >= v.size())
+                throw index_error();
+            v[i] = t;
+        }
+    );
+
+    /// Slicing protocol
+    cl.def("__getitem__",
+        [](const Vector &v, slice slice) -> Vector * {
+            size_t start, stop, step, slicelength;
+
+            if (!slice.compute(v.size(), &start, &stop, &step, &slicelength))
+                throw error_already_set();
+
+            Vector *seq = new Vector();
+            seq->reserve((size_t) slicelength);
+
+            for (size_t i=0; i<slicelength; ++i) {
+                seq->push_back(v[start]);
+                start += step;
+            }
+            return seq;
+        },
+        arg("s"),
+        "Retrieve list elements using a slice object"
+    );
+
+    cl.def("__setitem__",
+        [](Vector &v, slice slice,  const Vector &value) {
+            size_t start, stop, step, slicelength;
+            if (!slice.compute(v.size(), &start, &stop, &step, &slicelength))
+                throw error_already_set();
+
+            if (slicelength != value.size())
+                throw std::runtime_error("Left and right hand size of slice assignment have different sizes!");
+
+            for (size_t i=0; i<slicelength; ++i) {
+                v[start] = value[i];
+                start += step;
+            }
+        },
+        "Assign list elements using a slice object"
+    );
+
+    cl.def("__delitem__",
+        [](Vector &v, SizeType i) {
+            if (i >= v.size())
+                throw index_error();
+            v.erase(v.begin() + DiffType(i));
+        },
+        "Delete the list elements at index ``i``"
+    );
+
+    cl.def("__delitem__",
+        [](Vector &v, slice slice) {
+            size_t start, stop, step, slicelength;
+
+            if (!slice.compute(v.size(), &start, &stop, &step, &slicelength))
+                throw error_already_set();
+
+            if (step == 1 && false) {
+                v.erase(v.begin() + (DiffType) start, v.begin() + DiffType(start + slicelength));
+            } else {
+                for (size_t i = 0; i < slicelength; ++i) {
+                    v.erase(v.begin() + DiffType(start));
+                    start += step - 1;
+                }
+            }
+        },
+        "Delete list elements using a slice object"
+    );
+
+}
+
+// If the type has an operator[] that doesn't return a reference (most notably std::vector<bool>),
+// we have to access by copying; otherwise we return by reference.
+template <typename Vector> using vector_needs_copy = negation<
+    std::is_same<decltype(std::declval<Vector>()[typename Vector::size_type()]), typename Vector::value_type &>>;
+
+// The usual case: access and iterate by reference
+template <typename Vector, typename Class_>
+void vector_accessor(enable_if_t<!vector_needs_copy<Vector>::value, Class_> &cl) {
+    using T = typename Vector::value_type;
+    using SizeType = typename Vector::size_type;
+    using ItType   = typename Vector::iterator;
+
+    cl.def("__getitem__",
+        [](Vector &v, SizeType i) -> T & {
+            if (i >= v.size())
+                throw index_error();
+            return v[i];
+        },
+        return_value_policy::reference_internal // ref + keepalive
+    );
+
+    cl.def("__iter__",
+           [](Vector &v) {
+               return make_iterator<
+                   return_value_policy::reference_internal, ItType, ItType, T&>(
+                   v.begin(), v.end());
+           },
+           keep_alive<0, 1>() /* Essential: keep list alive while iterator exists */
+    );
+}
+
+// The case for special objects, like std::vector<bool>, that have to be returned-by-copy:
+template <typename Vector, typename Class_>
+void vector_accessor(enable_if_t<vector_needs_copy<Vector>::value, Class_> &cl) {
+    using T = typename Vector::value_type;
+    using SizeType = typename Vector::size_type;
+    using ItType   = typename Vector::iterator;
+    cl.def("__getitem__",
+        [](const Vector &v, SizeType i) -> T {
+            if (i >= v.size())
+                throw index_error();
+            return v[i];
+        }
+    );
+
+    cl.def("__iter__",
+           [](Vector &v) {
+               return make_iterator<
+                   return_value_policy::copy, ItType, ItType, T>(
+                   v.begin(), v.end());
+           },
+           keep_alive<0, 1>() /* Essential: keep list alive while iterator exists */
+    );
+}
+
+template <typename Vector, typename Class_> auto vector_if_insertion_operator(Class_ &cl, std::string const &name)
+    -> decltype(std::declval<std::ostream&>() << std::declval<typename Vector::value_type>(), void()) {
+    using size_type = typename Vector::size_type;
+
+    cl.def("__repr__",
+           [name](Vector &v) {
+            std::ostringstream s;
+            s << name << '[';
+            for (size_type i=0; i < v.size(); ++i) {
+                s << v[i];
+                if (i != v.size() - 1)
+                    s << ", ";
+            }
+            s << ']';
+            return s.str();
+        },
+        "Return the canonical string representation of this list."
+    );
+}
+
+// Provide the buffer interface for vectors if we have data() and we have a format for it
+// GCC seems to have "void std::vector<bool>::data()" - doing SFINAE on the existence of data() is insufficient, we need to check it returns an appropriate pointer
+template <typename Vector, typename = void>
+struct vector_has_data_and_format : std::false_type {};
+template <typename Vector>
+struct vector_has_data_and_format<Vector, enable_if_t<std::is_same<decltype(format_descriptor<typename Vector::value_type>::format(), std::declval<Vector>().data()), typename Vector::value_type*>::value>> : std::true_type {};
+
+// Add the buffer interface to a vector
+template <typename Vector, typename Class_, typename... Args>
+enable_if_t<detail::any_of<std::is_same<Args, buffer_protocol>...>::value>
+vector_buffer(Class_& cl) {
+    using T = typename Vector::value_type;
+
+    static_assert(vector_has_data_and_format<Vector>::value, "There is not an appropriate format descriptor for this vector");
+
+    // numpy.h declares this for arbitrary types, but it may raise an exception and crash hard at runtime if PYBIND11_NUMPY_DTYPE hasn't been called, so check here
+    format_descriptor<T>::format();
+
+    cl.def_buffer([](Vector& v) -> buffer_info {
+        return buffer_info(v.data(), static_cast<ssize_t>(sizeof(T)), format_descriptor<T>::format(), 1, {v.size()}, {sizeof(T)});
+    });
+
+    cl.def(init([](buffer buf) {
+        auto info = buf.request();
+        if (info.ndim != 1 || info.strides[0] % static_cast<ssize_t>(sizeof(T)))
+            throw type_error("Only valid 1D buffers can be copied to a vector");
+        if (!detail::compare_buffer_info<T>::compare(info) || (ssize_t) sizeof(T) != info.itemsize)
+            throw type_error("Format mismatch (Python: " + info.format + " C++: " + format_descriptor<T>::format() + ")");
+
+        auto vec = std::unique_ptr<Vector>(new Vector());
+        vec->reserve((size_t) info.shape[0]);
+        T *p = static_cast<T*>(info.ptr);
+        ssize_t step = info.strides[0] / static_cast<ssize_t>(sizeof(T));
+        T *end = p + info.shape[0] * step;
+        for (; p != end; p += step)
+            vec->push_back(*p);
+        return vec.release();
+    }));
+
+    return;
+}
+
+template <typename Vector, typename Class_, typename... Args>
+enable_if_t<!detail::any_of<std::is_same<Args, buffer_protocol>...>::value> vector_buffer(Class_&) {}
+
+NAMESPACE_END(detail)
+
+//
+// std::vector
+//
+template <typename Vector, typename holder_type = std::unique_ptr<Vector>, typename... Args>
+class_<Vector, holder_type> bind_vector(handle scope, std::string const &name, Args&&... args) {
+    using Class_ = class_<Vector, holder_type>;
+
+    // If the value_type is unregistered (e.g. a converting type) or is itself registered
+    // module-local then make the vector binding module-local as well:
+    using vtype = typename Vector::value_type;
+    auto vtype_info = detail::get_type_info(typeid(vtype));
+    bool local = !vtype_info || vtype_info->module_local;
+
+    Class_ cl(scope, name.c_str(), pybind11::module_local(local), std::forward<Args>(args)...);
+
+    // Declare the buffer interface if a buffer_protocol() is passed in
+    detail::vector_buffer<Vector, Class_, Args...>(cl);
+
+    cl.def(init<>());
+
+    // Register copy constructor (if possible)
+    detail::vector_if_copy_constructible<Vector, Class_>(cl);
+
+    // Register comparison-related operators and functions (if possible)
+    detail::vector_if_equal_operator<Vector, Class_>(cl);
+
+    // Register stream insertion operator (if possible)
+    detail::vector_if_insertion_operator<Vector, Class_>(cl, name);
+
+    // Modifiers require copyable vector value type
+    detail::vector_modifiers<Vector, Class_>(cl);
+
+    // Accessor and iterator; return by value if copyable, otherwise we return by ref + keep-alive
+    detail::vector_accessor<Vector, Class_>(cl);
+
+    cl.def("__bool__",
+        [](const Vector &v) -> bool {
+            return !v.empty();
+        },
+        "Check whether the list is nonempty"
+    );
+
+    cl.def("__len__", &Vector::size);
+
+
+
+
+#if 0
+    // C++ style functions deprecated, leaving it here as an example
+    cl.def(init<size_type>());
+
+    cl.def("resize",
+         (void (Vector::*) (size_type count)) & Vector::resize,
+         "changes the number of elements stored");
+
+    cl.def("erase",
+        [](Vector &v, SizeType i) {
+        if (i >= v.size())
+            throw index_error();
+        v.erase(v.begin() + i);
+    }, "erases element at index ``i``");
+
+    cl.def("empty",         &Vector::empty,         "checks whether the container is empty");
+    cl.def("size",          &Vector::size,          "returns the number of elements");
+    cl.def("push_back", (void (Vector::*)(const T&)) &Vector::push_back, "adds an element to the end");
+    cl.def("pop_back",                               &Vector::pop_back, "removes the last element");
+
+    cl.def("max_size",      &Vector::max_size,      "returns the maximum possible number of elements");
+    cl.def("reserve",       &Vector::reserve,       "reserves storage");
+    cl.def("capacity",      &Vector::capacity,      "returns the number of elements that can be held in currently allocated storage");
+    cl.def("shrink_to_fit", &Vector::shrink_to_fit, "reduces memory usage by freeing unused memory");
+
+    cl.def("clear", &Vector::clear, "clears the contents");
+    cl.def("swap",   &Vector::swap, "swaps the contents");
+
+    cl.def("front", [](Vector &v) {
+        if (v.size()) return v.front();
+        else throw index_error();
+    }, "access the first element");
+
+    cl.def("back", [](Vector &v) {
+        if (v.size()) return v.back();
+        else throw index_error();
+    }, "access the last element ");
+
+#endif
+
+    return cl;
+}
+
+
+
+//
+// std::map, std::unordered_map
+//
+
+NAMESPACE_BEGIN(detail)
+
+/* Fallback functions */
+template <typename, typename, typename... Args> void map_if_insertion_operator(const Args &...) { }
+template <typename, typename, typename... Args> void map_assignment(const Args &...) { }
+
+// Map assignment when copy-assignable: just copy the value
+template <typename Map, typename Class_>
+void map_assignment(enable_if_t<std::is_copy_assignable<typename Map::mapped_type>::value, Class_> &cl) {
+    using KeyType = typename Map::key_type;
+    using MappedType = typename Map::mapped_type;
+
+    cl.def("__setitem__",
+           [](Map &m, const KeyType &k, const MappedType &v) {
+               auto it = m.find(k);
+               if (it != m.end()) it->second = v;
+               else m.emplace(k, v);
+           }
+    );
+}
+
+// Not copy-assignable, but still copy-constructible: we can update the value by erasing and reinserting
+template<typename Map, typename Class_>
+void map_assignment(enable_if_t<
+        !std::is_copy_assignable<typename Map::mapped_type>::value &&
+        is_copy_constructible<typename Map::mapped_type>::value,
+        Class_> &cl) {
+    using KeyType = typename Map::key_type;
+    using MappedType = typename Map::mapped_type;
+
+    cl.def("__setitem__",
+           [](Map &m, const KeyType &k, const MappedType &v) {
+               // We can't use m[k] = v; because value type might not be default constructable
+               auto r = m.emplace(k, v);
+               if (!r.second) {
+                   // value type is not copy assignable so the only way to insert it is to erase it first...
+                   m.erase(r.first);
+                   m.emplace(k, v);
+               }
+           }
+    );
+}
+
+
+template <typename Map, typename Class_> auto map_if_insertion_operator(Class_ &cl, std::string const &name)
+-> decltype(std::declval<std::ostream&>() << std::declval<typename Map::key_type>() << std::declval<typename Map::mapped_type>(), void()) {
+
+    cl.def("__repr__",
+           [name](Map &m) {
+            std::ostringstream s;
+            s << name << '{';
+            bool f = false;
+            for (auto const &kv : m) {
+                if (f)
+                    s << ", ";
+                s << kv.first << ": " << kv.second;
+                f = true;
+            }
+            s << '}';
+            return s.str();
+        },
+        "Return the canonical string representation of this map."
+    );
+}
+
+
+NAMESPACE_END(detail)
+
+template <typename Map, typename holder_type = std::unique_ptr<Map>, typename... Args>
+class_<Map, holder_type> bind_map(handle scope, const std::string &name, Args&&... args) {
+    using KeyType = typename Map::key_type;
+    using MappedType = typename Map::mapped_type;
+    using Class_ = class_<Map, holder_type>;
+
+    // If either type is a non-module-local bound type then make the map binding non-local as well;
+    // otherwise (e.g. both types are either module-local or converting) the map will be
+    // module-local.
+    auto tinfo = detail::get_type_info(typeid(MappedType));
+    bool local = !tinfo || tinfo->module_local;
+    if (local) {
+        tinfo = detail::get_type_info(typeid(KeyType));
+        local = !tinfo || tinfo->module_local;
+    }
+
+    Class_ cl(scope, name.c_str(), pybind11::module_local(local), std::forward<Args>(args)...);
+
+    cl.def(init<>());
+
+    // Register stream insertion operator (if possible)
+    detail::map_if_insertion_operator<Map, Class_>(cl, name);
+
+    cl.def("__bool__",
+        [](const Map &m) -> bool { return !m.empty(); },
+        "Check whether the map is nonempty"
+    );
+
+    cl.def("__iter__",
+           [](Map &m) { return make_key_iterator(m.begin(), m.end()); },
+           keep_alive<0, 1>() /* Essential: keep list alive while iterator exists */
+    );
+
+    cl.def("items",
+           [](Map &m) { return make_iterator(m.begin(), m.end()); },
+           keep_alive<0, 1>() /* Essential: keep list alive while iterator exists */
+    );
+
+    cl.def("__getitem__",
+        [](Map &m, const KeyType &k) -> MappedType & {
+            auto it = m.find(k);
+            if (it == m.end())
+              throw key_error();
+           return it->second;
+        },
+        return_value_policy::reference_internal // ref + keepalive
+    );
+
+    // Assignment provided only if the type is copyable
+    detail::map_assignment<Map, Class_>(cl);
+
+    cl.def("__delitem__",
+           [](Map &m, const KeyType &k) {
+               auto it = m.find(k);
+               if (it == m.end())
+                   throw key_error();
+               m.erase(it);
+           }
+    );
+
+    cl.def("__len__", &Map::size);
+
+    return cl;
+}
+
+NAMESPACE_END(PYBIND11_NAMESPACE)
diff --git a/ml/dlib/dlib/external/pybind11/tools/FindCatch.cmake b/ml/dlib/dlib/external/pybind11/tools/FindCatch.cmake
new file mode 100644
index 000000000..9d490c5aa
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/FindCatch.cmake
@@ -0,0 +1,57 @@
+# - Find the Catch test framework or download it (single header)
+#
+# This is a quick module for internal use. It assumes that Catch is
+# REQUIRED and that a minimum version is provided (not EXACT). If
+# a suitable version isn't found locally, the single header file
+# will be downloaded and placed in the build dir: PROJECT_BINARY_DIR.
+#
+# This code sets the following variables:
+#  CATCH_INCLUDE_DIR      - path to catch.hpp
+#  CATCH_VERSION          - version number
+
+if(NOT Catch_FIND_VERSION)
+  message(FATAL_ERROR "A version number must be specified.")
+elseif(Catch_FIND_REQUIRED)
+  message(FATAL_ERROR "This module assumes Catch is not required.")
+elseif(Catch_FIND_VERSION_EXACT)
+  message(FATAL_ERROR "Exact version numbers are not supported, only minimum.")
+endif()
+
+# Extract the version number from catch.hpp
+function(_get_catch_version)
+  file(STRINGS "${CATCH_INCLUDE_DIR}/catch.hpp" version_line REGEX "Catch v.*" LIMIT_COUNT 1)
+  if(version_line MATCHES "Catch v([0-9]+)\\.([0-9]+)\\.([0-9]+)")
+    set(CATCH_VERSION "${CMAKE_MATCH_1}.${CMAKE_MATCH_2}.${CMAKE_MATCH_3}" PARENT_SCOPE)
+  endif()
+endfunction()
+
+# Download the single-header version of Catch
+function(_download_catch version destination_dir)
+  message(STATUS "Downloading catch v${version}...")
+  set(url https://github.com/philsquared/Catch/releases/download/v${version}/catch.hpp)
+  file(DOWNLOAD ${url} "${destination_dir}/catch.hpp" STATUS status)
+  list(GET status 0 error)
+  if(error)
+    message(FATAL_ERROR "Could not download ${url}")
+  endif()
+  set(CATCH_INCLUDE_DIR "${destination_dir}" CACHE INTERNAL "")
+endfunction()
+
+# Look for catch locally
+find_path(CATCH_INCLUDE_DIR NAMES catch.hpp PATH_SUFFIXES catch)
+if(CATCH_INCLUDE_DIR)
+  _get_catch_version()
+endif()
+
+# Download the header if it wasn't found or if it's outdated
+if(NOT CATCH_VERSION OR CATCH_VERSION VERSION_LESS ${Catch_FIND_VERSION})
+  if(DOWNLOAD_CATCH)
+    _download_catch(${Catch_FIND_VERSION} "${PROJECT_BINARY_DIR}/catch/")
+    _get_catch_version()
+  else()
+    set(CATCH_FOUND FALSE)
+    return()
+  endif()
+endif()
+
+set(CATCH_FOUND TRUE)
diff --git a/ml/dlib/dlib/external/pybind11/tools/FindEigen3.cmake b/ml/dlib/dlib/external/pybind11/tools/FindEigen3.cmake
new file mode 100644
index 000000000..9c546a05d
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/FindEigen3.cmake
@@ -0,0 +1,81 @@
+# - Try to find Eigen3 lib
+#
+# This module supports requiring a minimum version, e.g. you can do
+#   find_package(Eigen3 3.1.2)
+# to require version 3.1.2 or newer of Eigen3.
+#
+# Once done this will define
+#
+#  EIGEN3_FOUND - system has eigen lib with correct version
+#  EIGEN3_INCLUDE_DIR - the eigen include directory
+#  EIGEN3_VERSION - eigen version
+
+# Copyright (c) 2006, 2007 Montel Laurent, <montel@kde.org>
+# Copyright (c) 2008, 2009 Gael Guennebaud, <g.gael@free.fr>
+# Copyright (c) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+# Redistribution and use is allowed according to the terms of the 2-clause BSD license.
+
+if(NOT Eigen3_FIND_VERSION)
+  if(NOT Eigen3_FIND_VERSION_MAJOR)
+    set(Eigen3_FIND_VERSION_MAJOR 2)
+  endif(NOT Eigen3_FIND_VERSION_MAJOR)
+  if(NOT Eigen3_FIND_VERSION_MINOR)
+    set(Eigen3_FIND_VERSION_MINOR 91)
+  endif(NOT Eigen3_FIND_VERSION_MINOR)
+  if(NOT Eigen3_FIND_VERSION_PATCH)
+    set(Eigen3_FIND_VERSION_PATCH 0)
+  endif(NOT Eigen3_FIND_VERSION_PATCH)
+
+  set(Eigen3_FIND_VERSION "${Eigen3_FIND_VERSION_MAJOR}.${Eigen3_FIND_VERSION_MINOR}.${Eigen3_FIND_VERSION_PATCH}")
+endif(NOT Eigen3_FIND_VERSION)
+
+macro(_eigen3_check_version)
+  file(READ "${EIGEN3_INCLUDE_DIR}/Eigen/src/Core/util/Macros.h" _eigen3_version_header)
+
+  string(REGEX MATCH "define[ \t]+EIGEN_WORLD_VERSION[ \t]+([0-9]+)" _eigen3_world_version_match "${_eigen3_version_header}")
+  set(EIGEN3_WORLD_VERSION "${CMAKE_MATCH_1}")
+  string(REGEX MATCH "define[ \t]+EIGEN_MAJOR_VERSION[ \t]+([0-9]+)" _eigen3_major_version_match "${_eigen3_version_header}")
+  set(EIGEN3_MAJOR_VERSION "${CMAKE_MATCH_1}")
+  string(REGEX MATCH "define[ \t]+EIGEN_MINOR_VERSION[ \t]+([0-9]+)" _eigen3_minor_version_match "${_eigen3_version_header}")
+  set(EIGEN3_MINOR_VERSION "${CMAKE_MATCH_1}")
+
+  set(EIGEN3_VERSION ${EIGEN3_WORLD_VERSION}.${EIGEN3_MAJOR_VERSION}.${EIGEN3_MINOR_VERSION})
+  if(${EIGEN3_VERSION} VERSION_LESS ${Eigen3_FIND_VERSION})
+    set(EIGEN3_VERSION_OK FALSE)
+  else(${EIGEN3_VERSION} VERSION_LESS ${Eigen3_FIND_VERSION})
+    set(EIGEN3_VERSION_OK TRUE)
+  endif(${EIGEN3_VERSION} VERSION_LESS ${Eigen3_FIND_VERSION})
+
+  if(NOT EIGEN3_VERSION_OK)
+
+    message(STATUS "Eigen3 version ${EIGEN3_VERSION} found in ${EIGEN3_INCLUDE_DIR}, "
+                   "but at least version ${Eigen3_FIND_VERSION} is required")
+  endif(NOT EIGEN3_VERSION_OK)
+endmacro(_eigen3_check_version)
+
+if (EIGEN3_INCLUDE_DIR)
+
+  # in cache already
+  _eigen3_check_version()
+  set(EIGEN3_FOUND ${EIGEN3_VERSION_OK})
+
+else (EIGEN3_INCLUDE_DIR)
+
+  find_path(EIGEN3_INCLUDE_DIR NAMES signature_of_eigen3_matrix_library
+      PATHS
+      ${CMAKE_INSTALL_PREFIX}/include
+      ${KDE4_INCLUDE_DIR}
+      PATH_SUFFIXES eigen3 eigen
+    )
+
+  if(EIGEN3_INCLUDE_DIR)
+    _eigen3_check_version()
+  endif(EIGEN3_INCLUDE_DIR)
+
+  include(FindPackageHandleStandardArgs)
+  find_package_handle_standard_args(Eigen3 DEFAULT_MSG EIGEN3_INCLUDE_DIR EIGEN3_VERSION_OK)
+
+  mark_as_advanced(EIGEN3_INCLUDE_DIR)
+
+endif(EIGEN3_INCLUDE_DIR)
+
diff --git a/ml/dlib/dlib/external/pybind11/tools/FindPythonLibsNew.cmake b/ml/dlib/dlib/external/pybind11/tools/FindPythonLibsNew.cmake
new file mode 100644
index 000000000..b29b287de
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/FindPythonLibsNew.cmake
@@ -0,0 +1,195 @@
+# - Find python libraries
+# This module finds the libraries corresponding to the Python interpreter
+# FindPythonInterp provides.
+# This code sets the following variables:
+#
+#  PYTHONLIBS_FOUND           - have the Python libs been found
+#  PYTHON_PREFIX              - path to the Python installation
+#  PYTHON_LIBRARIES           - path to the python library
+#  PYTHON_INCLUDE_DIRS        - path to where Python.h is found
+#  PYTHON_MODULE_EXTENSION    - lib extension, e.g. '.so' or '.pyd'
+#  PYTHON_MODULE_PREFIX       - lib name prefix: usually an empty string
+#  PYTHON_SITE_PACKAGES       - path to installation site-packages
+#  PYTHON_IS_DEBUG            - whether the Python interpreter is a debug build
+#
+# Thanks to talljimbo for the patch adding the 'LDVERSION' config
+# variable usage.
+
+#=============================================================================
+# Copyright 2001-2009 Kitware, Inc.
+# Copyright 2012 Continuum Analytics, Inc.
+#
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#
+# * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright
+# notice, this list of conditions and the following disclaimer in the
+# documentation and/or other materials provided with the distribution.
+#
+# * Neither the names of Kitware, Inc., the Insight Software Consortium,
+# nor the names of their contributors may be used to endorse or promote
+# products derived from this software without specific prior written
+# permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# # A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+#=============================================================================
+
+# Checking for the extension makes sure that `LibsNew` was found and not just `Libs`.
+if(PYTHONLIBS_FOUND AND PYTHON_MODULE_EXTENSION)
+    return()
+endif()
+
+# Use the Python interpreter to find the libs.
+if(PythonLibsNew_FIND_REQUIRED)
+    find_package(PythonInterp ${PythonLibsNew_FIND_VERSION} REQUIRED)
+else()
+    find_package(PythonInterp ${PythonLibsNew_FIND_VERSION})
+endif()
+
+if(NOT PYTHONINTERP_FOUND)
+    set(PYTHONLIBS_FOUND FALSE)
+    return()
+endif()
+
+# According to http://stackoverflow.com/questions/646518/python-how-to-detect-debug-interpreter
+# testing whether sys has the gettotalrefcount function is a reliable, cross-platform
+# way to detect a CPython debug interpreter.
+#
+# The library suffix is from the config var LDVERSION sometimes, otherwise
+# VERSION. VERSION will typically be like "2.7" on unix, and "27" on windows.
+execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
+    "from distutils import sysconfig as s;import sys;import struct;
+print('.'.join(str(v) for v in sys.version_info));
+print(sys.prefix);
+print(s.get_python_inc(plat_specific=True));
+print(s.get_python_lib(plat_specific=True));
+print(s.get_config_var('SO'));
+print(hasattr(sys, 'gettotalrefcount')+0);
+print(struct.calcsize('@P'));
+print(s.get_config_var('LDVERSION') or s.get_config_var('VERSION'));
+print(s.get_config_var('LIBDIR') or '');
+print(s.get_config_var('MULTIARCH') or '');
+"
+    RESULT_VARIABLE _PYTHON_SUCCESS
+    OUTPUT_VARIABLE _PYTHON_VALUES
+    ERROR_VARIABLE _PYTHON_ERROR_VALUE)
+
+if(NOT _PYTHON_SUCCESS MATCHES 0)
+    if(PythonLibsNew_FIND_REQUIRED)
+        message(FATAL_ERROR
+            "Python config failure:\n${_PYTHON_ERROR_VALUE}")
+    endif()
+    set(PYTHONLIBS_FOUND FALSE)
+    return()
+endif()
+
+# Convert the process output into a list
+string(REGEX REPLACE ";" "\\\\;" _PYTHON_VALUES ${_PYTHON_VALUES})
+string(REGEX REPLACE "\n" ";" _PYTHON_VALUES ${_PYTHON_VALUES})
+list(GET _PYTHON_VALUES 0 _PYTHON_VERSION_LIST)
+list(GET _PYTHON_VALUES 1 PYTHON_PREFIX)
+list(GET _PYTHON_VALUES 2 PYTHON_INCLUDE_DIR)
+list(GET _PYTHON_VALUES 3 PYTHON_SITE_PACKAGES)
+list(GET _PYTHON_VALUES 4 PYTHON_MODULE_EXTENSION)
+list(GET _PYTHON_VALUES 5 PYTHON_IS_DEBUG)
+list(GET _PYTHON_VALUES 6 PYTHON_SIZEOF_VOID_P)
+list(GET _PYTHON_VALUES 7 PYTHON_LIBRARY_SUFFIX)
+list(GET _PYTHON_VALUES 8 PYTHON_LIBDIR)
+list(GET _PYTHON_VALUES 9 PYTHON_MULTIARCH)
+
+# Make sure the Python has the same pointer-size as the chosen compiler
+# Skip if CMAKE_SIZEOF_VOID_P is not defined
+if(CMAKE_SIZEOF_VOID_P AND (NOT "${PYTHON_SIZEOF_VOID_P}" STREQUAL "${CMAKE_SIZEOF_VOID_P}"))
+    if(PythonLibsNew_FIND_REQUIRED)
+        math(EXPR _PYTHON_BITS "${PYTHON_SIZEOF_VOID_P} * 8")
+        math(EXPR _CMAKE_BITS "${CMAKE_SIZEOF_VOID_P} * 8")
+        message(FATAL_ERROR
+            "Python config failure: Python is ${_PYTHON_BITS}-bit, "
+            "chosen compiler is  ${_CMAKE_BITS}-bit")
+    endif()
+    set(PYTHONLIBS_FOUND FALSE)
+    return()
+endif()
+
+# The built-in FindPython didn't always give the version numbers
+string(REGEX REPLACE "\\." ";" _PYTHON_VERSION_LIST ${_PYTHON_VERSION_LIST})
+list(GET _PYTHON_VERSION_LIST 0 PYTHON_VERSION_MAJOR)
+list(GET _PYTHON_VERSION_LIST 1 PYTHON_VERSION_MINOR)
+list(GET _PYTHON_VERSION_LIST 2 PYTHON_VERSION_PATCH)
+
+# Make sure all directory separators are '/'
+string(REGEX REPLACE "\\\\" "/" PYTHON_PREFIX ${PYTHON_PREFIX})
+string(REGEX REPLACE "\\\\" "/" PYTHON_INCLUDE_DIR ${PYTHON_INCLUDE_DIR})
+string(REGEX REPLACE "\\\\" "/" PYTHON_SITE_PACKAGES ${PYTHON_SITE_PACKAGES})
+
+if(CMAKE_HOST_WIN32)
+    set(PYTHON_LIBRARY
+        "${PYTHON_PREFIX}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
+
+    # when run in a venv, PYTHON_PREFIX points to it. But the libraries remain in the
+    # original python installation. They may be found relative to PYTHON_INCLUDE_DIR.
+    if(NOT EXISTS "${PYTHON_LIBRARY}")
+        get_filename_component(_PYTHON_ROOT ${PYTHON_INCLUDE_DIR} DIRECTORY)
+        set(PYTHON_LIBRARY
+            "${_PYTHON_ROOT}/libs/Python${PYTHON_LIBRARY_SUFFIX}.lib")
+    endif()
+
+    # raise an error if the python libs are still not found.
+    if(NOT EXISTS "${PYTHON_LIBRARY}")
+        message(FATAL_ERROR "Python libraries not found")
+    endif()
+
+else()
+    if(PYTHON_MULTIARCH)
+        set(_PYTHON_LIBS_SEARCH "${PYTHON_LIBDIR}/${PYTHON_MULTIARCH}" "${PYTHON_LIBDIR}")
+    else()
+        set(_PYTHON_LIBS_SEARCH "${PYTHON_LIBDIR}")
+    endif()
+    #message(STATUS "Searching for Python libs in ${_PYTHON_LIBS_SEARCH}")
+    # Probably this needs to be more involved. It would be nice if the config
+    # information the python interpreter itself gave us were more complete.
+    find_library(PYTHON_LIBRARY
+        NAMES "python${PYTHON_LIBRARY_SUFFIX}"
+        PATHS ${_PYTHON_LIBS_SEARCH}
+        NO_DEFAULT_PATH)
+
+    # If all else fails, just set the name/version and let the linker figure out the path.
+    if(NOT PYTHON_LIBRARY)
+        set(PYTHON_LIBRARY python${PYTHON_LIBRARY_SUFFIX})
+    endif()
+endif()
+
+MARK_AS_ADVANCED(
+  PYTHON_LIBRARY
+  PYTHON_INCLUDE_DIR
+)
+
+# We use PYTHON_INCLUDE_DIR, PYTHON_LIBRARY and PYTHON_DEBUG_LIBRARY for the
+# cache entries because they are meant to specify the location of a single
+# library. We now set the variables listed by the documentation for this
+# module.
+SET(PYTHON_INCLUDE_DIRS "${PYTHON_INCLUDE_DIR}")
+SET(PYTHON_LIBRARIES "${PYTHON_LIBRARY}")
+SET(PYTHON_DEBUG_LIBRARIES "${PYTHON_DEBUG_LIBRARY}")
+
+find_package_message(PYTHON
+    "Found PythonLibs: ${PYTHON_LIBRARY}"
+    "${PYTHON_EXECUTABLE}${PYTHON_VERSION}")
+
+set(PYTHONLIBS_FOUND TRUE)
diff --git a/ml/dlib/dlib/external/pybind11/tools/check-style.sh b/ml/dlib/dlib/external/pybind11/tools/check-style.sh
new file mode 100755
index 000000000..0a9f7d24f
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/check-style.sh
@@ -0,0 +1,70 @@
+#!/bin/bash
+#
+# Script to check include/test code for common pybind11 code style errors.
+#
+# This script currently checks for
+#
+# 1. use of tabs instead of spaces
+# 2. MSDOS-style CRLF endings
+# 3. trailing spaces
+# 4. missing space between keyword and parenthesis, e.g.: for(, if(, while(
+# 5. Missing space between right parenthesis and brace, e.g. 'for (...){'
+# 6. opening brace on its own line. It should always be on the same line as the
+#    if/while/for/do statement.
+#
+# Invoke as: tools/check-style.sh
+#
+
+check_style_errors=0
+IFS=$'\n'
+
+found="$( GREP_COLORS='mt=41' GREP_COLOR='41' grep $'\t' include tests/*.{cpp,py,h} docs/*.rst -rn --color=always )"
+if [ -n "$found" ]; then
+    # The mt=41 sets a red background for matched tabs:
+    echo -e '\033[31;01mError: found tab characters in the following files:\033[0m'
+    check_style_errors=1
+    echo "$found" | sed -e 's/^/    /'
+fi
+
+
+found="$( grep -IUlr $'\r' include tests/*.{cpp,py,h} docs/*.rst --color=always )"
+if [ -n "$found" ]; then
+    echo -e '\033[31;01mError: found CRLF characters in the following files:\033[0m'
+    check_style_errors=1
+    echo "$found" | sed -e 's/^/    /'
+fi
+
+found="$(GREP_COLORS='mt=41' GREP_COLOR='41' grep '[[:blank:]]\+$' include tests/*.{cpp,py,h} docs/*.rst -rn --color=always )"
+if [ -n "$found" ]; then
+    # The mt=41 sets a red background for matched trailing spaces
+    echo -e '\033[31;01mError: found trailing spaces in the following files:\033[0m'
+    check_style_errors=1
+    echo "$found" | sed -e 's/^/    /'
+fi
+
+found="$(grep '\<\(if\|for\|while\|catch\)(\|){' include tests/*.{cpp,h} -rn --color=always)"
+if [ -n "$found" ]; then
+    echo -e '\033[31;01mError: found the following coding style problems:\033[0m'
+    check_style_errors=1
+    echo "$found" | sed -e 's/^/    /'
+fi
+
+found="$(awk '
+function prefix(filename, lineno) {
+    return "    \033[35m" filename "\033[36m:\033[32m" lineno "\033[36m:\033[0m"
+}
+function mark(pattern, string) { sub(pattern, "\033[01;31m&\033[0m", string); return string }
+last && /^\s*{/ {
+    print prefix(FILENAME, FNR-1) mark("\\)\\s*$", last)
+    print prefix(FILENAME, FNR)   mark("^\\s*{", $0)
+    last=""
+}
+{ last = /(if|for|while|catch|switch)\s*\(.*\)\s*$/ ? $0 : "" }
+' $(find include -type f) tests/*.{cpp,h} docs/*.rst)"
+if [ -n "$found" ]; then
+    check_style_errors=1
+    echo -e '\033[31;01mError: braces should occur on the same line as the if/while/.. statement. Found issues in the following files:\033[0m'
+    echo "$found"
+fi
+
+exit $check_style_errors
diff --git a/ml/dlib/dlib/external/pybind11/tools/libsize.py b/ml/dlib/dlib/external/pybind11/tools/libsize.py
new file mode 100644
index 000000000..5dcb8b0d0
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/libsize.py
@@ -0,0 +1,38 @@
+from __future__ import print_function, division
+import os
+import sys
+
+# Internal build script for generating debugging test .so size.
+# Usage:
+#     python libsize.py file.so save.txt -- displays the size of file.so and, if save.txt exists, compares it to the
+#                                           size in it, then overwrites save.txt with the new size for future runs.
+
+if len(sys.argv) != 3:
+    sys.exit("Invalid arguments: usage: python libsize.py file.so save.txt")
+
+lib = sys.argv[1]
+save = sys.argv[2]
+
+if not os.path.exists(lib):
+    sys.exit("Error: requested file ({}) does not exist".format(lib))
+
+libsize = os.path.getsize(lib)
+
+print("------", os.path.basename(lib), "file size:", libsize, end='')
+
+if os.path.exists(save):
+    with open(save) as sf:
+        oldsize = int(sf.readline())
+
+    if oldsize > 0:
+        change = libsize - oldsize
+        if change == 0:
+            print(" (no change)")
+        else:
+            print(" (change of {:+} bytes = {:+.2%})".format(change, change / oldsize))
+else:
+    print()
+
+with open(save, 'w') as sf:
+    sf.write(str(libsize))
+
diff --git a/ml/dlib/dlib/external/pybind11/tools/mkdoc.py b/ml/dlib/dlib/external/pybind11/tools/mkdoc.py
new file mode 100644
index 000000000..1fd8cceed
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/mkdoc.py
@@ -0,0 +1,304 @@
+#!/usr/bin/env python3
+#
+#  Syntax: mkdoc.py [-I<path> ..] [.. a list of header files ..]
+#
+#  Extract documentation from C++ header files to use it in Python bindings
+#
+
+import os
+import sys
+import platform
+import re
+import textwrap
+
+from clang import cindex
+from clang.cindex import CursorKind
+from collections import OrderedDict
+from threading import Thread, Semaphore
+from multiprocessing import cpu_count
+
+RECURSE_LIST = [
+    CursorKind.TRANSLATION_UNIT,
+    CursorKind.NAMESPACE,
+    CursorKind.CLASS_DECL,
+    CursorKind.STRUCT_DECL,
+    CursorKind.ENUM_DECL,
+    CursorKind.CLASS_TEMPLATE
+]
+
+PRINT_LIST = [
+    CursorKind.CLASS_DECL,
+    CursorKind.STRUCT_DECL,
+    CursorKind.ENUM_DECL,
+    CursorKind.ENUM_CONSTANT_DECL,
+    CursorKind.CLASS_TEMPLATE,
+    CursorKind.FUNCTION_DECL,
+    CursorKind.FUNCTION_TEMPLATE,
+    CursorKind.CONVERSION_FUNCTION,
+    CursorKind.CXX_METHOD,
+    CursorKind.CONSTRUCTOR,
+    CursorKind.FIELD_DECL
+]
+
+CPP_OPERATORS = {
+    '<=': 'le', '>=': 'ge', '==': 'eq', '!=': 'ne', '[]': 'array',
+    '+=': 'iadd', '-=': 'isub', '*=': 'imul', '/=': 'idiv', '%=':
+    'imod', '&=': 'iand', '|=': 'ior', '^=': 'ixor', '<<=': 'ilshift',
+    '>>=': 'irshift', '++': 'inc', '--': 'dec', '<<': 'lshift', '>>':
+    'rshift', '&&': 'land', '||': 'lor', '!': 'lnot', '~': 'bnot',
+    '&': 'band', '|': 'bor', '+': 'add', '-': 'sub', '*': 'mul', '/':
+    'div', '%': 'mod', '<': 'lt', '>': 'gt', '=': 'assign', '()': 'call'
+}
+
+CPP_OPERATORS = OrderedDict(
+    sorted(CPP_OPERATORS.items(), key=lambda t: -len(t[0])))
+
+job_count = cpu_count()
+job_semaphore = Semaphore(job_count)
+
+output = []
+
+def d(s):
+    return s.decode('utf8')
+
+
+def sanitize_name(name):
+    name = re.sub(r'type-parameter-0-([0-9]+)', r'T\1', name)
+    for k, v in CPP_OPERATORS.items():
+        name = name.replace('operator%s' % k, 'operator_%s' % v)
+    name = re.sub('<.*>', '', name)
+    name = ''.join([ch if ch.isalnum() else '_' for ch in name])
+    name = re.sub('_$', '', re.sub('_+', '_', name))
+    return '__doc_' + name
+
+
+def process_comment(comment):
+    result = ''
+
+    # Remove C++ comment syntax
+    leading_spaces = float('inf')
+    for s in comment.expandtabs(tabsize=4).splitlines():
+        s = s.strip()
+        if s.startswith('/*'):
+            s = s[2:].lstrip('*')
+        elif s.endswith('*/'):
+            s = s[:-2].rstrip('*')
+        elif s.startswith('///'):
+            s = s[3:]
+        if s.startswith('*'):
+            s = s[1:]
+        if len(s) > 0:
+            leading_spaces = min(leading_spaces, len(s) - len(s.lstrip()))
+        result += s + '\n'
+
+    if leading_spaces != float('inf'):
+        result2 = ""
+        for s in result.splitlines():
+            result2 += s[leading_spaces:] + '\n'
+        result = result2
+
+    # Doxygen tags
+    cpp_group = '([\w:]+)'
+    param_group = '([\[\w:\]]+)'
+
+    s = result
+    s = re.sub(r'\\c\s+%s' % cpp_group, r'``\1``', s)
+    s = re.sub(r'\\a\s+%s' % cpp_group, r'*\1*', s)
+    s = re.sub(r'\\e\s+%s' % cpp_group, r'*\1*', s)
+    s = re.sub(r'\\em\s+%s' % cpp_group, r'*\1*', s)
+    s = re.sub(r'\\b\s+%s' % cpp_group, r'**\1**', s)
+    s = re.sub(r'\\ingroup\s+%s' % cpp_group, r'', s)
+    s = re.sub(r'\\param%s?\s+%s' % (param_group, cpp_group),
+               r'\n\n$Parameter ``\2``:\n\n', s)
+    s = re.sub(r'\\tparam%s?\s+%s' % (param_group, cpp_group),
+               r'\n\n$Template parameter ``\2``:\n\n', s)
+
+    for in_, out_ in {
+        'return': 'Returns',
+        'author': 'Author',
+        'authors': 'Authors',
+        'copyright': 'Copyright',
+        'date': 'Date',
+        'remark': 'Remark',
+        'sa': 'See also',
+        'see': 'See also',
+        'extends': 'Extends',
+        'throw': 'Throws',
+        'throws': 'Throws'
+    }.items():
+        s = re.sub(r'\\%s\s*' % in_, r'\n\n$%s:\n\n' % out_, s)
+
+    s = re.sub(r'\\details\s*', r'\n\n', s)
+    s = re.sub(r'\\brief\s*', r'', s)
+    s = re.sub(r'\\short\s*', r'', s)
+    s = re.sub(r'\\ref\s*', r'', s)
+
+    s = re.sub(r'\\code\s?(.*?)\s?\\endcode',
+               r"```\n\1\n```\n", s, flags=re.DOTALL)
+
+    # HTML/TeX tags
+    s = re.sub(r'<tt>(.*?)</tt>', r'``\1``', s, flags=re.DOTALL)
+    s = re.sub(r'<pre>(.*?)</pre>', r"```\n\1\n```\n", s, flags=re.DOTALL)
+    s = re.sub(r'<em>(.*?)</em>', r'*\1*', s, flags=re.DOTALL)
+    s = re.sub(r'<b>(.*?)</b>', r'**\1**', s, flags=re.DOTALL)
+    s = re.sub(r'\\f\$(.*?)\\f\$', r'$\1$', s, flags=re.DOTALL)
+    s = re.sub(r'<li>', r'\n\n* ', s)
+    s = re.sub(r'</?ul>', r'', s)
+    s = re.sub(r'</li>', r'\n\n', s)
+
+    s = s.replace('``true``', '``True``')
+    s = s.replace('``false``', '``False``')
+
+    # Re-flow text
+    wrapper = textwrap.TextWrapper()
+    wrapper.expand_tabs = True
+    wrapper.replace_whitespace = True
+    wrapper.drop_whitespace = True
+    wrapper.width = 70
+    wrapper.initial_indent = wrapper.subsequent_indent = ''
+
+    result = ''
+    in_code_segment = False
+    for x in re.split(r'(```)', s):
+        if x == '```':
+            if not in_code_segment:
+                result += '```\n'
+            else:
+                result += '\n```\n\n'
+            in_code_segment = not in_code_segment
+        elif in_code_segment:
+            result += x.strip()
+        else:
+            for y in re.split(r'(?: *\n *){2,}', x):
+                wrapped = wrapper.fill(re.sub(r'\s+', ' ', y).strip())
+                if len(wrapped) > 0 and wrapped[0] == '$':
+                    result += wrapped[1:] + '\n'
+                    wrapper.initial_indent = \
+                        wrapper.subsequent_indent = ' ' * 4
+                else:
+                    if len(wrapped) > 0:
+                        result += wrapped + '\n\n'
+                    wrapper.initial_indent = wrapper.subsequent_indent = ''
+    return result.rstrip().lstrip('\n')
+
+
+def extract(filename, node, prefix):
+    if not (node.location.file is None or
+            os.path.samefile(d(node.location.file.name), filename)):
+        return 0
+    if node.kind in RECURSE_LIST:
+        sub_prefix = prefix
+        if node.kind != CursorKind.TRANSLATION_UNIT:
+            if len(sub_prefix) > 0:
+                sub_prefix += '_'
+            sub_prefix += d(node.spelling)
+        for i in node.get_children():
+            extract(filename, i, sub_prefix)
+    if node.kind in PRINT_LIST:
+        comment = d(node.raw_comment) if node.raw_comment is not None else ''
+        comment = process_comment(comment)
+        sub_prefix = prefix
+        if len(sub_prefix) > 0:
+            sub_prefix += '_'
+        if len(node.spelling) > 0:
+            name = sanitize_name(sub_prefix + d(node.spelling))
+            global output
+            output.append((name, filename, comment))
+
+
+class ExtractionThread(Thread):
+    def __init__(self, filename, parameters):
+        Thread.__init__(self)
+        self.filename = filename
+        self.parameters = parameters
+        job_semaphore.acquire()
+
+    def run(self):
+        print('Processing "%s" ..' % self.filename, file=sys.stderr)
+        try:
+            index = cindex.Index(
+                cindex.conf.lib.clang_createIndex(False, True))
+            tu = index.parse(self.filename, self.parameters)
+            extract(self.filename, tu.cursor, '')
+        finally:
+            job_semaphore.release()
+
+if __name__ == '__main__':
+    parameters = ['-x', 'c++', '-std=c++11']
+    filenames = []
+
+    if platform.system() == 'Darwin':
+        dev_path = '/Applications/Xcode.app/Contents/Developer/'
+        lib_dir = dev_path + 'Toolchains/XcodeDefault.xctoolchain/usr/lib/'
+        sdk_dir = dev_path + 'Platforms/MacOSX.platform/Developer/SDKs'
+        libclang = lib_dir + 'libclang.dylib'
+
+        if os.path.exists(libclang):
+            cindex.Config.set_library_path(os.path.dirname(libclang))
+
+        if os.path.exists(sdk_dir):
+            sysroot_dir = os.path.join(sdk_dir, next(os.walk(sdk_dir))[1][0])
+            parameters.append('-isysroot')
+            parameters.append(sysroot_dir)
+
+    for item in sys.argv[1:]:
+        if item.startswith('-'):
+            parameters.append(item)
+        else:
+            filenames.append(item)
+
+    if len(filenames) == 0:
+        print('Syntax: %s [.. a list of header files ..]' % sys.argv[0])
+        exit(-1)
+
+    print('''/*
+  This file contains docstrings for the Python bindings.
+  Do not edit! These were automatically extracted by mkdoc.py
+ */
+
+#define __EXPAND(x)                                      x
+#define __COUNT(_1, _2, _3, _4, _5, _6, _7, COUNT, ...)  COUNT
+#define __VA_SIZE(...)                                   __EXPAND(__COUNT(__VA_ARGS__, 7, 6, 5, 4, 3, 2, 1))
+#define __CAT1(a, b)                                     a ## b
+#define __CAT2(a, b)                                     __CAT1(a, b)
+#define __DOC1(n1)                                       __doc_##n1
+#define __DOC2(n1, n2)                                   __doc_##n1##_##n2
+#define __DOC3(n1, n2, n3)                               __doc_##n1##_##n2##_##n3
+#define __DOC4(n1, n2, n3, n4)                           __doc_##n1##_##n2##_##n3##_##n4
+#define __DOC5(n1, n2, n3, n4, n5)                       __doc_##n1##_##n2##_##n3##_##n4##_##n5
+#define __DOC6(n1, n2, n3, n4, n5, n6)                   __doc_##n1##_##n2##_##n3##_##n4##_##n5##_##n6
+#define __DOC7(n1, n2, n3, n4, n5, n6, n7)               __doc_##n1##_##n2##_##n3##_##n4##_##n5##_##n6##_##n7
+#define DOC(...)                                         __EXPAND(__EXPAND(__CAT2(__DOC, __VA_SIZE(__VA_ARGS__)))(__VA_ARGS__))
+
+#if defined(__GNUG__)
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-variable"
+#endif
+''')
+
+    output.clear()
+    for filename in filenames:
+        thr = ExtractionThread(filename, parameters)
+        thr.start()
+
+    print('Waiting for jobs to finish ..', file=sys.stderr)
+    for i in range(job_count):
+        job_semaphore.acquire()
+
+    name_ctr = 1
+    name_prev = None
+    for name, _, comment in list(sorted(output, key=lambda x: (x[0], x[1]))):
+        if name == name_prev:
+            name_ctr += 1
+            name = name + "_%i" % name_ctr
+        else:
+            name_prev = name
+            name_ctr = 1
+        print('\nstatic const char *%s =%sR"doc(%s)doc";' %
+              (name, '\n' if '\n' in comment else ' ', comment))
+
+    print('''
+#if defined(__GNUG__)
+#pragma GCC diagnostic pop
+#endif
+''')
diff --git a/ml/dlib/dlib/external/pybind11/tools/pybind11Config.cmake.in b/ml/dlib/dlib/external/pybind11/tools/pybind11Config.cmake.in
new file mode 100644
index 000000000..3dd1b2c1a
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/pybind11Config.cmake.in
@@ -0,0 +1,100 @@
+# pybind11Config.cmake
+# --------------------
+#
+# PYBIND11 cmake module.
+# This module sets the following variables in your project::
+#
+#   pybind11_FOUND - true if pybind11 and all required components found on the system
+#   pybind11_VERSION - pybind11 version in format Major.Minor.Release
+#   pybind11_INCLUDE_DIRS - Directories where pybind11 and python headers are located.
+#   pybind11_INCLUDE_DIR - Directory where pybind11 headers are located.
+#   pybind11_DEFINITIONS - Definitions necessary to use pybind11, namely USING_pybind11.
+#   pybind11_LIBRARIES - compile flags and python libraries (as needed) to link against.
+#   pybind11_LIBRARY - empty.
+#   CMAKE_MODULE_PATH - appends location of accompanying FindPythonLibsNew.cmake and
+#                       pybind11Tools.cmake modules.
+#
+#
+# Available components: None
+#
+#
+# Exported targets::
+#
+# If pybind11 is found, this module defines the following :prop_tgt:`IMPORTED`
+# interface library targets::
+#
+#   pybind11::module - for extension modules
+#   pybind11::embed - for embedding the Python interpreter
+#
+# Python headers, libraries (as needed by platform), and the C++ standard
+# are attached to the target. Set PythonLibsNew variables to influence
+# python detection and PYBIND11_CPP_STANDARD (-std=c++11 or -std=c++14) to
+# influence standard setting. ::
+#
+#   find_package(pybind11 CONFIG REQUIRED)
+#   message(STATUS "Found pybind11 v${pybind11_VERSION}: ${pybind11_INCLUDE_DIRS}")
+#
+#   # Create an extension module
+#   add_library(mylib MODULE main.cpp)
+#   target_link_libraries(mylib pybind11::module)
+#
+#   # Or embed the Python interpreter into an executable
+#   add_executable(myexe main.cpp)
+#   target_link_libraries(myexe pybind11::embed)
+#
+# Suggested usage::
+#
+# find_package with version info is not recommended except for release versions. ::
+#
+#   find_package(pybind11 CONFIG)
+#   find_package(pybind11 2.0 EXACT CONFIG REQUIRED)
+#
+#
+# The following variables can be set to guide the search for this package::
+#
+#   pybind11_DIR - CMake variable, set to directory containing this Config file
+#   CMAKE_PREFIX_PATH - CMake variable, set to root directory of this package
+#   PATH - environment variable, set to bin directory of this package
+#   CMAKE_DISABLE_FIND_PACKAGE_pybind11 - CMake variable, disables
+#     find_package(pybind11) when not REQUIRED, perhaps to force internal build
+
+@PACKAGE_INIT@
+
+set(PN pybind11)
+
+# location of pybind11/pybind11.h
+set(${PN}_INCLUDE_DIR "${PACKAGE_PREFIX_DIR}/@CMAKE_INSTALL_INCLUDEDIR@")
+
+set(${PN}_LIBRARY "")
+set(${PN}_DEFINITIONS USING_${PN})
+
+check_required_components(${PN})
+
+# make detectable the FindPythonLibsNew.cmake module
+list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_LIST_DIR})
+
+include(pybind11Tools)
+
+if(NOT (CMAKE_VERSION VERSION_LESS 3.0))
+#-----------------------------------------------------------------------------
+# Don't include targets if this file is being picked up by another
+# project which has already built this as a subproject
+#-----------------------------------------------------------------------------
+if(NOT TARGET ${PN}::pybind11)
+    include("${CMAKE_CURRENT_LIST_DIR}/${PN}Targets.cmake")
+
+    find_package(PythonLibsNew ${PYBIND11_PYTHON_VERSION} MODULE REQUIRED)
+    set_property(TARGET ${PN}::pybind11 APPEND PROPERTY INTERFACE_INCLUDE_DIRECTORIES ${PYTHON_INCLUDE_DIRS})
+    set_property(TARGET ${PN}::embed APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${PYTHON_LIBRARIES})
+    if(WIN32 OR CYGWIN)
+      set_property(TARGET ${PN}::module APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${PYTHON_LIBRARIES})
+    endif()
+
+    set_property(TARGET ${PN}::pybind11 APPEND PROPERTY INTERFACE_COMPILE_OPTIONS "${PYBIND11_CPP_STANDARD}")
+
+    get_property(_iid TARGET ${PN}::pybind11 PROPERTY INTERFACE_INCLUDE_DIRECTORIES)
+    get_property(_ill TARGET ${PN}::module PROPERTY INTERFACE_LINK_LIBRARIES)
+    set(${PN}_INCLUDE_DIRS ${_iid})
+    set(${PN}_LIBRARIES ${_ico} ${_ill})
+endif()
+endif()
diff --git a/ml/dlib/dlib/external/pybind11/tools/pybind11Tools.cmake b/ml/dlib/dlib/external/pybind11/tools/pybind11Tools.cmake
new file mode 100644
index 000000000..a7c471a07
--- /dev/null
+++ b/ml/dlib/dlib/external/pybind11/tools/pybind11Tools.cmake
@@ -0,0 +1,202 @@
+# tools/pybind11Tools.cmake -- Build system for the pybind11 modules
+#
+# Copyright (c) 2015 Wenzel Jakob <wenzel@inf.ethz.ch>
+#
+# All rights reserved. Use of this source code is governed by a
+# BSD-style license that can be found in the LICENSE file.
+
+cmake_minimum_required(VERSION 2.8.12)
+
+# Add a CMake parameter for choosing a desired Python version
+if(NOT PYBIND11_PYTHON_VERSION)
+  set(PYBIND11_PYTHON_VERSION "" CACHE STRING "Python version to use for compiling modules")
+endif()
+
+set(Python_ADDITIONAL_VERSIONS 3.7 3.6 3.5 3.4)
+find_package(PythonLibsNew ${PYBIND11_PYTHON_VERSION} REQUIRED)
+
+include(CheckCXXCompilerFlag)
+include(CMakeParseArguments)
+
+if(NOT PYBIND11_CPP_STANDARD AND NOT CMAKE_CXX_STANDARD)
+  if(NOT MSVC)
+    check_cxx_compiler_flag("-std=c++14" HAS_CPP14_FLAG)
+
+    if (HAS_CPP14_FLAG)
+      set(PYBIND11_CPP_STANDARD -std=c++14)
+    else()
+      check_cxx_compiler_flag("-std=c++11" HAS_CPP11_FLAG)
+      if (HAS_CPP11_FLAG)
+        set(PYBIND11_CPP_STANDARD -std=c++11)
+      else()
+        message(FATAL_ERROR "Unsupported compiler -- pybind11 requires C++11 support!")
+      endif()
+    endif()
+  elseif(MSVC)
+    set(PYBIND11_CPP_STANDARD /std:c++14)
+  endif()
+
+  set(PYBIND11_CPP_STANDARD ${PYBIND11_CPP_STANDARD} CACHE STRING
+      "C++ standard flag, e.g. -std=c++11, -std=c++14, /std:c++14.  Defaults to C++14 mode." FORCE)
+endif()
+
+# Checks whether the given CXX/linker flags can compile and link a cxx file.  cxxflags and
+# linkerflags are lists of flags to use.  The result variable is a unique variable name for each set
+# of flags: the compilation result will be cached base on the result variable.  If the flags work,
+# sets them in cxxflags_out/linkerflags_out internal cache variables (in addition to ${result}).
+function(_pybind11_return_if_cxx_and_linker_flags_work result cxxflags linkerflags cxxflags_out linkerflags_out)
+  set(CMAKE_REQUIRED_LIBRARIES ${linkerflags})
+  check_cxx_compiler_flag("${cxxflags}" ${result})
+  if (${result})
+    set(${cxxflags_out} "${cxxflags}" CACHE INTERNAL "" FORCE)
+    set(${linkerflags_out} "${linkerflags}" CACHE INTERNAL "" FORCE)
+  endif()
+endfunction()
+
+# Internal: find the appropriate link time optimization flags for this compiler
+function(_pybind11_add_lto_flags target_name prefer_thin_lto)
+  if (NOT DEFINED PYBIND11_LTO_CXX_FLAGS)
+    set(PYBIND11_LTO_CXX_FLAGS "" CACHE INTERNAL "")
+    set(PYBIND11_LTO_LINKER_FLAGS "" CACHE INTERNAL "")
+
+    if(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+      set(cxx_append "")
+      set(linker_append "")
+      if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND NOT APPLE)
+        # Clang Gold plugin does not support -Os; append -O3 to MinSizeRel builds to override it
+        set(linker_append ";$<$<CONFIG:MinSizeRel>:-O3>")
+      elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU")
+        set(cxx_append ";-fno-fat-lto-objects")
+      endif()
+
+      if (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND prefer_thin_lto)
+        _pybind11_return_if_cxx_and_linker_flags_work(HAS_FLTO_THIN
+          "-flto=thin${cxx_append}" "-flto=thin${linker_append}"
+          PYBIND11_LTO_CXX_FLAGS PYBIND11_LTO_LINKER_FLAGS)
+      endif()
+
+      if (NOT HAS_FLTO_THIN)
+        _pybind11_return_if_cxx_and_linker_flags_work(HAS_FLTO
+          "-flto${cxx_append}" "-flto${linker_append}"
+          PYBIND11_LTO_CXX_FLAGS PYBIND11_LTO_LINKER_FLAGS)
+      endif()
+    elseif (CMAKE_CXX_COMPILER_ID MATCHES "Intel")
+      # Intel equivalent to LTO is called IPO
+      _pybind11_return_if_cxx_and_linker_flags_work(HAS_INTEL_IPO
+      "-ipo" "-ipo" PYBIND11_LTO_CXX_FLAGS PYBIND11_LTO_LINKER_FLAGS)
+    elseif(MSVC)
+      # cmake only interprets libraries as linker flags when they start with a - (otherwise it
+      # converts /LTCG to \LTCG as if it was a Windows path).  Luckily MSVC supports passing flags
+      # with - instead of /, even if it is a bit non-standard:
+      _pybind11_return_if_cxx_and_linker_flags_work(HAS_MSVC_GL_LTCG
+        "/GL" "-LTCG" PYBIND11_LTO_CXX_FLAGS PYBIND11_LTO_LINKER_FLAGS)
+    endif()
+
+    if (PYBIND11_LTO_CXX_FLAGS)
+      message(STATUS "LTO enabled")
+    else()
+      message(STATUS "LTO disabled (not supported by the compiler and/or linker)")
+    endif()
+  endif()
+
+  # Enable LTO flags if found, except for Debug builds
+  if (PYBIND11_LTO_CXX_FLAGS)
+    target_compile_options(${target_name} PRIVATE "$<$<NOT:$<CONFIG:Debug>>:${PYBIND11_LTO_CXX_FLAGS}>")
+  endif()
+  if (PYBIND11_LTO_LINKER_FLAGS)
+    target_link_libraries(${target_name} PRIVATE "$<$<NOT:$<CONFIG:Debug>>:${PYBIND11_LTO_LINKER_FLAGS}>")
+  endif()
+endfunction()
+
+# Build a Python extension module:
+# pybind11_add_module(<name> [MODULE | SHARED] [EXCLUDE_FROM_ALL]
+#                     [NO_EXTRAS] [THIN_LTO] source1 [source2 ...])
+#
+function(pybind11_add_module target_name)
+  set(options MODULE SHARED EXCLUDE_FROM_ALL NO_EXTRAS THIN_LTO)
+  cmake_parse_arguments(ARG "${options}" "" "" ${ARGN})
+
+  if(ARG_MODULE AND ARG_SHARED)
+    message(FATAL_ERROR "Can't be both MODULE and SHARED")
+  elseif(ARG_SHARED)
+    set(lib_type SHARED)
+  else()
+    set(lib_type MODULE)
+  endif()
+
+  if(ARG_EXCLUDE_FROM_ALL)
+    set(exclude_from_all EXCLUDE_FROM_ALL)
+  endif()
+
+  add_library(${target_name} ${lib_type} ${exclude_from_all} ${ARG_UNPARSED_ARGUMENTS})
+
+  target_include_directories(${target_name}
+    PRIVATE ${PYBIND11_INCLUDE_DIR}  # from project CMakeLists.txt
+    PRIVATE ${pybind11_INCLUDE_DIR}  # from pybind11Config
+    PRIVATE ${PYTHON_INCLUDE_DIRS})
+
+  # The prefix and extension are provided by FindPythonLibsNew.cmake
+  set_target_properties(${target_name} PROPERTIES PREFIX "${PYTHON_MODULE_PREFIX}")
+  set_target_properties(${target_name} PROPERTIES SUFFIX "${PYTHON_MODULE_EXTENSION}")
+
+  # -fvisibility=hidden is required to allow multiple modules compiled against
+  # different pybind versions to work properly, and for some features (e.g.
+  # py::module_local).  We force it on everything inside the `pybind11`
+  # namespace; also turning it on for a pybind module compilation here avoids
+  # potential warnings or issues from having mixed hidden/non-hidden types.
+  set_target_properties(${target_name} PROPERTIES CXX_VISIBILITY_PRESET "hidden")
+
+  if(WIN32 OR CYGWIN)
+    # Link against the Python shared library on Windows
+    target_link_libraries(${target_name} PRIVATE ${PYTHON_LIBRARIES})
+  elseif(APPLE)
+    # It's quite common to have multiple copies of the same Python version
+    # installed on one's system. E.g.: one copy from the OS and another copy
+    # that's statically linked into an application like Blender or Maya.
+    # If we link our plugin library against the OS Python here and import it
+    # into Blender or Maya later on, this will cause segfaults when multiple
+    # conflicting Python instances are active at the same time (even when they
+    # are of the same version).
+
+    # Windows is not affected by this issue since it handles DLL imports
+    # differently. The solution for Linux and Mac OS is simple: we just don't
+    # link against the Python library. The resulting shared library will have
+    # missing symbols, but that's perfectly fine -- they will be resolved at
+    # import time.
+
+    target_link_libraries(${target_name} PRIVATE "-undefined dynamic_lookup")
+
+    if(ARG_SHARED)
+      # Suppress CMake >= 3.0 warning for shared libraries
+      set_target_properties(${target_name} PROPERTIES MACOSX_RPATH ON)
+    endif()
+  endif()
+
+  # Make sure C++11/14 are enabled
+  target_compile_options(${target_name} PUBLIC ${PYBIND11_CPP_STANDARD})
+
+  if(ARG_NO_EXTRAS)
+    return()
+  endif()
+
+  _pybind11_add_lto_flags(${target_name} ${ARG_THIN_LTO})
+
+  if (NOT MSVC AND NOT ${CMAKE_BUILD_TYPE} MATCHES Debug)
+    # Strip unnecessary sections of the binary on Linux/Mac OS
+    if(CMAKE_STRIP)
+      if(APPLE)
+        add_custom_command(TARGET ${target_name} POST_BUILD
+                           COMMAND ${CMAKE_STRIP} -x $<TARGET_FILE:${target_name}>)
+      else()
+        add_custom_command(TARGET ${target_name} POST_BUILD
+                           COMMAND ${CMAKE_STRIP} $<TARGET_FILE:${target_name}>)
+      endif()
+    endif()
+  endif()
+
+  if(MSVC)
+    # /MP enables multithreaded builds (relevant when there are many files), /bigobj is
+    # needed for bigger binding projects due to the limit to 64k addressable sections
+    target_compile_options(${target_name} PRIVATE /MP /bigobj)
+  endif()
+endfunction()
diff --git a/ml/dlib/dlib/external/zlib/README b/ml/dlib/dlib/external/zlib/README
new file mode 100644
index 000000000..5ca9d127e
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/README
@@ -0,0 +1,115 @@
+ZLIB DATA COMPRESSION LIBRARY
+
+zlib 1.2.8 is a general purpose data compression library.  All the code is
+thread safe.  The data format used by the zlib library is described by RFCs
+(Request for Comments) 1950 to 1952 in the files
+http://tools.ietf.org/html/rfc1950 (zlib format), rfc1951 (deflate format) and
+rfc1952 (gzip format).
+
+All functions of the compression library are documented in the file zlib.h
+(volunteer to write man pages welcome, contact zlib@gzip.org).  A usage example
+of the library is given in the file test/example.c which also tests that
+the library is working correctly.  Another example is given in the file
+test/minigzip.c.  The compression library itself is composed of all source
+files in the root directory.
+
+To compile all files and run the test program, follow the instructions given at
+the top of Makefile.in.  In short "./configure; make test", and if that goes
+well, "make install" should work for most flavors of Unix.  For Windows, use
+one of the special makefiles in win32/ or contrib/vstudio/ .  For VMS, use
+make_vms.com.
+
+Questions about zlib should be sent to <zlib@gzip.org>, or to Gilles Vollant
+<info@winimage.com> for the Windows DLL version.  The zlib home page is
+http://zlib.net/ .  Before reporting a problem, please check this site to
+verify that you have the latest version of zlib; otherwise get the latest
+version and check whether the problem still exists or not.
+
+PLEASE read the zlib FAQ http://zlib.net/zlib_faq.html before asking for help.
+
+Mark Nelson <markn@ieee.org> wrote an article about zlib for the Jan.  1997
+issue of Dr.  Dobb's Journal; a copy of the article is available at
+http://marknelson.us/1997/01/01/zlib-engine/ .
+
+The changes made in version 1.2.8 are documented in the file ChangeLog.
+
+Unsupported third party contributions are provided in directory contrib/ .
+
+zlib is available in Java using the java.util.zip package, documented at
+http://java.sun.com/developer/technicalArticles/Programming/compression/ .
+
+A Perl interface to zlib written by Paul Marquess <pmqs@cpan.org> is available
+at CPAN (Comprehensive Perl Archive Network) sites, including
+http://search.cpan.org/~pmqs/IO-Compress-Zlib/ .
+
+A Python interface to zlib written by A.M. Kuchling <amk@amk.ca> is
+available in Python 1.5 and later versions, see
+http://docs.python.org/library/zlib.html .
+
+zlib is built into tcl: http://wiki.tcl.tk/4610 .
+
+An experimental package to read and write files in .zip format, written on top
+of zlib by Gilles Vollant <info@winimage.com>, is available in the
+contrib/minizip directory of zlib.
+
+
+Notes for some targets:
+
+- For Windows DLL versions, please see win32/DLL_FAQ.txt
+
+- For 64-bit Irix, deflate.c must be compiled without any optimization. With
+  -O, one libpng test fails. The test works in 32 bit mode (with the -n32
+  compiler flag). The compiler bug has been reported to SGI.
+
+- zlib doesn't work with gcc 2.6.3 on a DEC 3000/300LX under OSF/1 2.1 it works
+  when compiled with cc.
+
+- On Digital Unix 4.0D (formely OSF/1) on AlphaServer, the cc option -std1 is
+  necessary to get gzprintf working correctly. This is done by configure.
+
+- zlib doesn't work on HP-UX 9.05 with some versions of /bin/cc. It works with
+  other compilers. Use "make test" to check your compiler.
+
+- gzdopen is not supported on RISCOS or BEOS.
+
+- For PalmOs, see http://palmzlib.sourceforge.net/
+
+
+Acknowledgments:
+
+  The deflate format used by zlib was defined by Phil Katz.  The deflate and
+  zlib specifications were written by L.  Peter Deutsch.  Thanks to all the
+  people who reported problems and suggested various improvements in zlib; they
+  are too numerous to cite here.
+
+Copyright notice:
+
+ (C) 1995-2013 Jean-loup Gailly and Mark Adler
+
+  This software is provided 'as-is', without any express or implied
+  warranty.  In no event will the authors be held liable for any damages
+  arising from the use of this software.
+
+  Permission is granted to anyone to use this software for any purpose,
+  including commercial applications, and to alter it and redistribute it
+  freely, subject to the following restrictions:
+
+  1. The origin of this software must not be misrepresented; you must not
+     claim that you wrote the original software. If you use this software
+     in a product, an acknowledgment in the product documentation would be
+     appreciated but is not required.
+  2. Altered source versions must be plainly marked as such, and must not be
+     misrepresented as being the original software.
+  3. This notice may not be removed or altered from any source distribution.
+
+  Jean-loup Gailly        Mark Adler
+  jloup@gzip.org          madler@alumni.caltech.edu
+
+If you use the zlib library in a product, we would appreciate *not* receiving
+lengthy legal documents to sign.  The sources are provided for free but without
+warranty of any kind.  The library has been entirely written by Jean-loup
+Gailly and Mark Adler; it does not include third-party code.
+
+If you redistribute modified sources, we would appreciate that you include in
+the file ChangeLog history information documenting your changes.  Please read
+the FAQ for more information on the distribution of modified source versions.
diff --git a/ml/dlib/dlib/external/zlib/adler32.c b/ml/dlib/dlib/external/zlib/adler32.c
new file mode 100644
index 000000000..a868f073d
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/adler32.c
@@ -0,0 +1,179 @@
+/* adler32.c -- compute the Adler-32 checksum of a data stream
+ * Copyright (C) 1995-2011 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* @(#) $Id$ */
+
+#include "zutil.h"
+
+#define local static
+
+local uLong adler32_combine_ OF((uLong adler1, uLong adler2, z_off64_t len2));
+
+#define BASE 65521      /* largest prime smaller than 65536 */
+#define NMAX 5552
+/* NMAX is the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1 */
+
+#define DO1(buf,i)  {adler += (buf)[i]; sum2 += adler;}
+#define DO2(buf,i)  DO1(buf,i); DO1(buf,i+1);
+#define DO4(buf,i)  DO2(buf,i); DO2(buf,i+2);
+#define DO8(buf,i)  DO4(buf,i); DO4(buf,i+4);
+#define DO16(buf)   DO8(buf,0); DO8(buf,8);
+
+/* use NO_DIVIDE if your processor does not do division in hardware --
+   try it both ways to see which is faster */
+#ifdef NO_DIVIDE
+/* note that this assumes BASE is 65521, where 65536 % 65521 == 15
+   (thank you to John Reiser for pointing this out) */
+#  define CHOP(a) \
+    do { \
+        unsigned long tmp = a >> 16; \
+        a &= 0xffffUL; \
+        a += (tmp << 4) - tmp; \
+    } while (0)
+#  define MOD28(a) \
+    do { \
+        CHOP(a); \
+        if (a >= BASE) a -= BASE; \
+    } while (0)
+#  define MOD(a) \
+    do { \
+        CHOP(a); \
+        MOD28(a); \
+    } while (0)
+#  define MOD63(a) \
+    do { /* this assumes a is not negative */ \
+        z_off64_t tmp = a >> 32; \
+        a &= 0xffffffffL; \
+        a += (tmp << 8) - (tmp << 5) + tmp; \
+        tmp = a >> 16; \
+        a &= 0xffffL; \
+        a += (tmp << 4) - tmp; \
+        tmp = a >> 16; \
+        a &= 0xffffL; \
+        a += (tmp << 4) - tmp; \
+        if (a >= BASE) a -= BASE; \
+    } while (0)
+#else
+#  define MOD(a) a %= BASE
+#  define MOD28(a) a %= BASE
+#  define MOD63(a) a %= BASE
+#endif
+
+/* ========================================================================= */
+uLong ZEXPORT adler32(adler, buf, len)
+    uLong adler;
+    const Bytef *buf;
+    uInt len;
+{
+    unsigned long sum2;
+    unsigned n;
+
+    /* split Adler-32 into component sums */
+    sum2 = (adler >> 16) & 0xffff;
+    adler &= 0xffff;
+
+    /* in case user likes doing a byte at a time, keep it fast */
+    if (len == 1) {
+        adler += buf[0];
+        if (adler >= BASE)
+            adler -= BASE;
+        sum2 += adler;
+        if (sum2 >= BASE)
+            sum2 -= BASE;
+        return adler | (sum2 << 16);
+    }
+
+    /* initial Adler-32 value (deferred check for len == 1 speed) */
+    if (buf == Z_NULL)
+        return 1L;
+
+    /* in case short lengths are provided, keep it somewhat fast */
+    if (len < 16) {
+        while (len--) {
+            adler += *buf++;
+            sum2 += adler;
+        }
+        if (adler >= BASE)
+            adler -= BASE;
+        MOD28(sum2);            /* only added so many BASE's */
+        return adler | (sum2 << 16);
+    }
+
+    /* do length NMAX blocks -- requires just one modulo operation */
+    while (len >= NMAX) {
+        len -= NMAX;
+        n = NMAX / 16;          /* NMAX is divisible by 16 */
+        do {
+            DO16(buf);          /* 16 sums unrolled */
+            buf += 16;
+        } while (--n);
+        MOD(adler);
+        MOD(sum2);
+    }
+
+    /* do remaining bytes (less than NMAX, still just one modulo) */
+    if (len) {                  /* avoid modulos if none remaining */
+        while (len >= 16) {
+            len -= 16;
+            DO16(buf);
+            buf += 16;
+        }
+        while (len--) {
+            adler += *buf++;
+            sum2 += adler;
+        }
+        MOD(adler);
+        MOD(sum2);
+    }
+
+    /* return recombined sums */
+    return adler | (sum2 << 16);
+}
+
+/* ========================================================================= */
+local uLong adler32_combine_(adler1, adler2, len2)
+    uLong adler1;
+    uLong adler2;
+    z_off64_t len2;
+{
+    unsigned long sum1;
+    unsigned long sum2;
+    unsigned rem;
+
+    /* for negative len, return invalid adler32 as a clue for debugging */
+    if (len2 < 0)
+        return 0xffffffffUL;
+
+    /* the derivation of this formula is left as an exercise for the reader */
+    MOD63(len2);                /* assumes len2 >= 0 */
+    rem = (unsigned)len2;
+    sum1 = adler1 & 0xffff;
+    sum2 = rem * sum1;
+    MOD(sum2);
+    sum1 += (adler2 & 0xffff) + BASE - 1;
+    sum2 += ((adler1 >> 16) & 0xffff) + ((adler2 >> 16) & 0xffff) + BASE - rem;
+    if (sum1 >= BASE) sum1 -= BASE;
+    if (sum1 >= BASE) sum1 -= BASE;
+    if (sum2 >= (BASE << 1)) sum2 -= (BASE << 1);
+    if (sum2 >= BASE) sum2 -= BASE;
+    return sum1 | (sum2 << 16);
+}
+
+/* ========================================================================= */
+uLong ZEXPORT adler32_combine(adler1, adler2, len2)
+    uLong adler1;
+    uLong adler2;
+    z_off_t len2;
+{
+    return adler32_combine_(adler1, adler2, len2);
+}
+
+uLong ZEXPORT adler32_combine64(adler1, adler2, len2)
+    uLong adler1;
+    uLong adler2;
+    z_off64_t len2;
+{
+    return adler32_combine_(adler1, adler2, len2);
+}
diff --git a/ml/dlib/dlib/external/zlib/compress.c b/ml/dlib/dlib/external/zlib/compress.c
new file mode 100644
index 000000000..6e9762676
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/compress.c
@@ -0,0 +1,80 @@
+/* compress.c -- compress a memory buffer
+ * Copyright (C) 1995-2005 Jean-loup Gailly.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* @(#) $Id$ */
+
+#define ZLIB_INTERNAL
+#include "zlib.h"
+
+/* ===========================================================================
+     Compresses the source buffer into the destination buffer. The level
+   parameter has the same meaning as in deflateInit.  sourceLen is the byte
+   length of the source buffer. Upon entry, destLen is the total size of the
+   destination buffer, which must be at least 0.1% larger than sourceLen plus
+   12 bytes. Upon exit, destLen is the actual size of the compressed buffer.
+
+     compress2 returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_BUF_ERROR if there was not enough room in the output buffer,
+   Z_STREAM_ERROR if the level parameter is invalid.
+*/
+int ZEXPORT compress2 (dest, destLen, source, sourceLen, level)
+    Bytef *dest;
+    uLongf *destLen;
+    const Bytef *source;
+    uLong sourceLen;
+    int level;
+{
+    z_stream stream;
+    int err;
+
+    stream.next_in = (z_const Bytef *)source;
+    stream.avail_in = (uInt)sourceLen;
+#ifdef MAXSEG_64K
+    /* Check for source > 64K on 16-bit machine: */
+    if ((uLong)stream.avail_in != sourceLen) return Z_BUF_ERROR;
+#endif
+    stream.next_out = dest;
+    stream.avail_out = (uInt)*destLen;
+    if ((uLong)stream.avail_out != *destLen) return Z_BUF_ERROR;
+
+    stream.zalloc = (alloc_func)0;
+    stream.zfree = (free_func)0;
+    stream.opaque = (voidpf)0;
+
+    err = deflateInit(&stream, level);
+    if (err != Z_OK) return err;
+
+    err = deflate(&stream, Z_FINISH);
+    if (err != Z_STREAM_END) {
+        deflateEnd(&stream);
+        return err == Z_OK ? Z_BUF_ERROR : err;
+    }
+    *destLen = stream.total_out;
+
+    err = deflateEnd(&stream);
+    return err;
+}
+
+/* ===========================================================================
+ */
+int ZEXPORT compress (dest, destLen, source, sourceLen)
+    Bytef *dest;
+    uLongf *destLen;
+    const Bytef *source;
+    uLong sourceLen;
+{
+    return compress2(dest, destLen, source, sourceLen, Z_DEFAULT_COMPRESSION);
+}
+
+/* ===========================================================================
+     If the default memLevel or windowBits for deflateInit() is changed, then
+   this function needs to be updated.
+ */
+uLong ZEXPORT compressBound (sourceLen)
+    uLong sourceLen;
+{
+    return sourceLen + (sourceLen >> 12) + (sourceLen >> 14) +
+           (sourceLen >> 25) + 13;
+}
diff --git a/ml/dlib/dlib/external/zlib/crc32.c b/ml/dlib/dlib/external/zlib/crc32.c
new file mode 100644
index 000000000..979a7190a
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/crc32.c
@@ -0,0 +1,425 @@
+/* crc32.c -- compute the CRC-32 of a data stream
+ * Copyright (C) 1995-2006, 2010, 2011, 2012 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ *
+ * Thanks to Rodney Brown <rbrown64@csc.com.au> for his contribution of faster
+ * CRC methods: exclusive-oring 32 bits of data at a time, and pre-computing
+ * tables for updating the shift register in one step with three exclusive-ors
+ * instead of four steps with four exclusive-ors.  This results in about a
+ * factor of two increase in speed on a Power PC G4 (PPC7455) using gcc -O3.
+ */
+
+/* @(#) $Id$ */
+
+/*
+  Note on the use of DYNAMIC_CRC_TABLE: there is no mutex or semaphore
+  protection on the static variables used to control the first-use generation
+  of the crc tables.  Therefore, if you #define DYNAMIC_CRC_TABLE, you should
+  first call get_crc_table() to initialize the tables before allowing more than
+  one thread to use crc32().
+
+  DYNAMIC_CRC_TABLE and MAKECRCH can be #defined to write out crc32.h.
+ */
+
+#ifdef MAKECRCH
+#  include <stdio.h>
+#  ifndef DYNAMIC_CRC_TABLE
+#    define DYNAMIC_CRC_TABLE
+#  endif /* !DYNAMIC_CRC_TABLE */
+#endif /* MAKECRCH */
+
+#include "zutil.h"      /* for STDC and FAR definitions */
+
+#define local static
+
+/* Definitions for doing the crc four data bytes at a time. */
+#if !defined(NOBYFOUR) && defined(Z_U4)
+#  define BYFOUR
+#endif
+#ifdef BYFOUR
+   local unsigned long crc32_little OF((unsigned long,
+                        const unsigned char FAR *, unsigned));
+   local unsigned long crc32_big OF((unsigned long,
+                        const unsigned char FAR *, unsigned));
+#  define TBLS 8
+#else
+#  define TBLS 1
+#endif /* BYFOUR */
+
+/* Local functions for crc concatenation */
+local unsigned long gf2_matrix_times OF((unsigned long *mat,
+                                         unsigned long vec));
+local void gf2_matrix_square OF((unsigned long *square, unsigned long *mat));
+local uLong crc32_combine_ OF((uLong crc1, uLong crc2, z_off64_t len2));
+
+
+#ifdef DYNAMIC_CRC_TABLE
+
+local volatile int crc_table_empty = 1;
+local z_crc_t FAR crc_table[TBLS][256];
+local void make_crc_table OF((void));
+#ifdef MAKECRCH
+   local void write_table OF((FILE *, const z_crc_t FAR *));
+#endif /* MAKECRCH */
+/*
+  Generate tables for a byte-wise 32-bit CRC calculation on the polynomial:
+  x^32+x^26+x^23+x^22+x^16+x^12+x^11+x^10+x^8+x^7+x^5+x^4+x^2+x+1.
+
+  Polynomials over GF(2) are represented in binary, one bit per coefficient,
+  with the lowest powers in the most significant bit.  Then adding polynomials
+  is just exclusive-or, and multiplying a polynomial by x is a right shift by
+  one.  If we call the above polynomial p, and represent a byte as the
+  polynomial q, also with the lowest power in the most significant bit (so the
+  byte 0xb1 is the polynomial x^7+x^3+x+1), then the CRC is (q*x^32) mod p,
+  where a mod b means the remainder after dividing a by b.
+
+  This calculation is done using the shift-register method of multiplying and
+  taking the remainder.  The register is initialized to zero, and for each
+  incoming bit, x^32 is added mod p to the register if the bit is a one (where
+  x^32 mod p is p+x^32 = x^26+...+1), and the register is multiplied mod p by
+  x (which is shifting right by one and adding x^32 mod p if the bit shifted
+  out is a one).  We start with the highest power (least significant bit) of
+  q and repeat for all eight bits of q.
+
+  The first table is simply the CRC of all possible eight bit values.  This is
+  all the information needed to generate CRCs on data a byte at a time for all
+  combinations of CRC register values and incoming bytes.  The remaining tables
+  allow for word-at-a-time CRC calculation for both big-endian and little-
+  endian machines, where a word is four bytes.
+*/
+local void make_crc_table()
+{
+    z_crc_t c;
+    int n, k;
+    z_crc_t poly;                       /* polynomial exclusive-or pattern */
+    /* terms of polynomial defining this crc (except x^32): */
+    static volatile int first = 1;      /* flag to limit concurrent making */
+    static const unsigned char p[] = {0,1,2,4,5,7,8,10,11,12,16,22,23,26};
+
+    /* See if another task is already doing this (not thread-safe, but better
+       than nothing -- significantly reduces duration of vulnerability in
+       case the advice about DYNAMIC_CRC_TABLE is ignored) */
+    if (first) {
+        first = 0;
+
+        /* make exclusive-or pattern from polynomial (0xedb88320UL) */
+        poly = 0;
+        for (n = 0; n < (int)(sizeof(p)/sizeof(unsigned char)); n++)
+            poly |= (z_crc_t)1 << (31 - p[n]);
+
+        /* generate a crc for every 8-bit value */
+        for (n = 0; n < 256; n++) {
+            c = (z_crc_t)n;
+            for (k = 0; k < 8; k++)
+                c = c & 1 ? poly ^ (c >> 1) : c >> 1;
+            crc_table[0][n] = c;
+        }
+
+#ifdef BYFOUR
+        /* generate crc for each value followed by one, two, and three zeros,
+           and then the byte reversal of those as well as the first table */
+        for (n = 0; n < 256; n++) {
+            c = crc_table[0][n];
+            crc_table[4][n] = ZSWAP32(c);
+            for (k = 1; k < 4; k++) {
+                c = crc_table[0][c & 0xff] ^ (c >> 8);
+                crc_table[k][n] = c;
+                crc_table[k + 4][n] = ZSWAP32(c);
+            }
+        }
+#endif /* BYFOUR */
+
+        crc_table_empty = 0;
+    }
+    else {      /* not first */
+        /* wait for the other guy to finish (not efficient, but rare) */
+        while (crc_table_empty)
+            ;
+    }
+
+#ifdef MAKECRCH
+    /* write out CRC tables to crc32.h */
+    {
+        FILE *out;
+
+        out = fopen("crc32.h", "w");
+        if (out == NULL) return;
+        fprintf(out, "/* crc32.h -- tables for rapid CRC calculation\n");
+        fprintf(out, " * Generated automatically by crc32.c\n */\n\n");
+        fprintf(out, "local const z_crc_t FAR ");
+        fprintf(out, "crc_table[TBLS][256] =\n{\n  {\n");
+        write_table(out, crc_table[0]);
+#  ifdef BYFOUR
+        fprintf(out, "#ifdef BYFOUR\n");
+        for (k = 1; k < 8; k++) {
+            fprintf(out, "  },\n  {\n");
+            write_table(out, crc_table[k]);
+        }
+        fprintf(out, "#endif\n");
+#  endif /* BYFOUR */
+        fprintf(out, "  }\n};\n");
+        fclose(out);
+    }
+#endif /* MAKECRCH */
+}
+
+#ifdef MAKECRCH
+local void write_table(out, table)
+    FILE *out;
+    const z_crc_t FAR *table;
+{
+    int n;
+
+    for (n = 0; n < 256; n++)
+        fprintf(out, "%s0x%08lxUL%s", n % 5 ? "" : "    ",
+                (unsigned long)(table[n]),
+                n == 255 ? "\n" : (n % 5 == 4 ? ",\n" : ", "));
+}
+#endif /* MAKECRCH */
+
+#else /* !DYNAMIC_CRC_TABLE */
+/* ========================================================================
+ * Tables of CRC-32s of all single-byte values, made by make_crc_table().
+ */
+#include "crc32.h"
+#endif /* DYNAMIC_CRC_TABLE */
+
+/* =========================================================================
+ * This function can be used by asm versions of crc32()
+ */
+const z_crc_t FAR * ZEXPORT get_crc_table()
+{
+#ifdef DYNAMIC_CRC_TABLE
+    if (crc_table_empty)
+        make_crc_table();
+#endif /* DYNAMIC_CRC_TABLE */
+    return (const z_crc_t FAR *)crc_table;
+}
+
+/* ========================================================================= */
+#define DO1 crc = crc_table[0][((int)crc ^ (*buf++)) & 0xff] ^ (crc >> 8)
+#define DO8 DO1; DO1; DO1; DO1; DO1; DO1; DO1; DO1
+
+/* ========================================================================= */
+unsigned long ZEXPORT crc32(crc, buf, len)
+    unsigned long crc;
+    const unsigned char FAR *buf;
+    uInt len;
+{
+    if (buf == Z_NULL) return 0UL;
+
+#ifdef DYNAMIC_CRC_TABLE
+    if (crc_table_empty)
+        make_crc_table();
+#endif /* DYNAMIC_CRC_TABLE */
+
+#ifdef BYFOUR
+    if (sizeof(void *) == sizeof(ptrdiff_t)) {
+        z_crc_t endian;
+
+        endian = 1;
+        if (*((unsigned char *)(&endian)))
+            return crc32_little(crc, buf, len);
+        else
+            return crc32_big(crc, buf, len);
+    }
+#endif /* BYFOUR */
+    crc = crc ^ 0xffffffffUL;
+    while (len >= 8) {
+        DO8;
+        len -= 8;
+    }
+    if (len) do {
+        DO1;
+    } while (--len);
+    return crc ^ 0xffffffffUL;
+}
+
+#ifdef BYFOUR
+
+/* ========================================================================= */
+#define DOLIT4 c ^= *buf4++; \
+        c = crc_table[3][c & 0xff] ^ crc_table[2][(c >> 8) & 0xff] ^ \
+            crc_table[1][(c >> 16) & 0xff] ^ crc_table[0][c >> 24]
+#define DOLIT32 DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4
+
+/* ========================================================================= */
+local unsigned long crc32_little(crc, buf, len)
+    unsigned long crc;
+    const unsigned char FAR *buf;
+    unsigned len;
+{
+    register z_crc_t c;
+    register const z_crc_t FAR *buf4;
+
+    c = (z_crc_t)crc;
+    c = ~c;
+    while (len && ((ptrdiff_t)buf & 3)) {
+        c = crc_table[0][(c ^ *buf++) & 0xff] ^ (c >> 8);
+        len--;
+    }
+
+    buf4 = (const z_crc_t FAR *)(const void FAR *)buf;
+    while (len >= 32) {
+        DOLIT32;
+        len -= 32;
+    }
+    while (len >= 4) {
+        DOLIT4;
+        len -= 4;
+    }
+    buf = (const unsigned char FAR *)buf4;
+
+    if (len) do {
+        c = crc_table[0][(c ^ *buf++) & 0xff] ^ (c >> 8);
+    } while (--len);
+    c = ~c;
+    return (unsigned long)c;
+}
+
+/* ========================================================================= */
+#define DOBIG4 c ^= *++buf4; \
+        c = crc_table[4][c & 0xff] ^ crc_table[5][(c >> 8) & 0xff] ^ \
+            crc_table[6][(c >> 16) & 0xff] ^ crc_table[7][c >> 24]
+#define DOBIG32 DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4
+
+/* ========================================================================= */
+local unsigned long crc32_big(crc, buf, len)
+    unsigned long crc;
+    const unsigned char FAR *buf;
+    unsigned len;
+{
+    register z_crc_t c;
+    register const z_crc_t FAR *buf4;
+
+    c = ZSWAP32((z_crc_t)crc);
+    c = ~c;
+    while (len && ((ptrdiff_t)buf & 3)) {
+        c = crc_table[4][(c >> 24) ^ *buf++] ^ (c << 8);
+        len--;
+    }
+
+    buf4 = (const z_crc_t FAR *)(const void FAR *)buf;
+    buf4--;
+    while (len >= 32) {
+        DOBIG32;
+        len -= 32;
+    }
+    while (len >= 4) {
+        DOBIG4;
+        len -= 4;
+    }
+    buf4++;
+    buf = (const unsigned char FAR *)buf4;
+
+    if (len) do {
+        c = crc_table[4][(c >> 24) ^ *buf++] ^ (c << 8);
+    } while (--len);
+    c = ~c;
+    return (unsigned long)(ZSWAP32(c));
+}
+
+#endif /* BYFOUR */
+
+#define GF2_DIM 32      /* dimension of GF(2) vectors (length of CRC) */
+
+/* ========================================================================= */
+local unsigned long gf2_matrix_times(mat, vec)
+    unsigned long *mat;
+    unsigned long vec;
+{
+    unsigned long sum;
+
+    sum = 0;
+    while (vec) {
+        if (vec & 1)
+            sum ^= *mat;
+        vec >>= 1;
+        mat++;
+    }
+    return sum;
+}
+
+/* ========================================================================= */
+local void gf2_matrix_square(square, mat)
+    unsigned long *square;
+    unsigned long *mat;
+{
+    int n;
+
+    for (n = 0; n < GF2_DIM; n++)
+        square[n] = gf2_matrix_times(mat, mat[n]);
+}
+
+/* ========================================================================= */
+local uLong crc32_combine_(crc1, crc2, len2)
+    uLong crc1;
+    uLong crc2;
+    z_off64_t len2;
+{
+    int n;
+    unsigned long row;
+    unsigned long even[GF2_DIM];    /* even-power-of-two zeros operator */
+    unsigned long odd[GF2_DIM];     /* odd-power-of-two zeros operator */
+
+    /* degenerate case (also disallow negative lengths) */
+    if (len2 <= 0)
+        return crc1;
+
+    /* put operator for one zero bit in odd */
+    odd[0] = 0xedb88320UL;          /* CRC-32 polynomial */
+    row = 1;
+    for (n = 1; n < GF2_DIM; n++) {
+        odd[n] = row;
+        row <<= 1;
+    }
+
+    /* put operator for two zero bits in even */
+    gf2_matrix_square(even, odd);
+
+    /* put operator for four zero bits in odd */
+    gf2_matrix_square(odd, even);
+
+    /* apply len2 zeros to crc1 (first square will put the operator for one
+       zero byte, eight zero bits, in even) */
+    do {
+        /* apply zeros operator for this bit of len2 */
+        gf2_matrix_square(even, odd);
+        if (len2 & 1)
+            crc1 = gf2_matrix_times(even, crc1);
+        len2 >>= 1;
+
+        /* if no more bits set, then done */
+        if (len2 == 0)
+            break;
+
+        /* another iteration of the loop with odd and even swapped */
+        gf2_matrix_square(odd, even);
+        if (len2 & 1)
+            crc1 = gf2_matrix_times(odd, crc1);
+        len2 >>= 1;
+
+        /* if no more bits set, then done */
+    } while (len2 != 0);
+
+    /* return combined crc */
+    crc1 ^= crc2;
+    return crc1;
+}
+
+/* ========================================================================= */
+uLong ZEXPORT crc32_combine(crc1, crc2, len2)
+    uLong crc1;
+    uLong crc2;
+    z_off_t len2;
+{
+    return crc32_combine_(crc1, crc2, len2);
+}
+
+uLong ZEXPORT crc32_combine64(crc1, crc2, len2)
+    uLong crc1;
+    uLong crc2;
+    z_off64_t len2;
+{
+    return crc32_combine_(crc1, crc2, len2);
+}
diff --git a/ml/dlib/dlib/external/zlib/crc32.h b/ml/dlib/dlib/external/zlib/crc32.h
new file mode 100644
index 000000000..9e0c77810
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/crc32.h
@@ -0,0 +1,441 @@
+/* crc32.h -- tables for rapid CRC calculation
+ * Generated automatically by crc32.c
+ */
+
+local const z_crc_t FAR crc_table[TBLS][256] =
+{
+  {
+    0x00000000UL, 0x77073096UL, 0xee0e612cUL, 0x990951baUL, 0x076dc419UL,
+    0x706af48fUL, 0xe963a535UL, 0x9e6495a3UL, 0x0edb8832UL, 0x79dcb8a4UL,
+    0xe0d5e91eUL, 0x97d2d988UL, 0x09b64c2bUL, 0x7eb17cbdUL, 0xe7b82d07UL,
+    0x90bf1d91UL, 0x1db71064UL, 0x6ab020f2UL, 0xf3b97148UL, 0x84be41deUL,
+    0x1adad47dUL, 0x6ddde4ebUL, 0xf4d4b551UL, 0x83d385c7UL, 0x136c9856UL,
+    0x646ba8c0UL, 0xfd62f97aUL, 0x8a65c9ecUL, 0x14015c4fUL, 0x63066cd9UL,
+    0xfa0f3d63UL, 0x8d080df5UL, 0x3b6e20c8UL, 0x4c69105eUL, 0xd56041e4UL,
+    0xa2677172UL, 0x3c03e4d1UL, 0x4b04d447UL, 0xd20d85fdUL, 0xa50ab56bUL,
+    0x35b5a8faUL, 0x42b2986cUL, 0xdbbbc9d6UL, 0xacbcf940UL, 0x32d86ce3UL,
+    0x45df5c75UL, 0xdcd60dcfUL, 0xabd13d59UL, 0x26d930acUL, 0x51de003aUL,
+    0xc8d75180UL, 0xbfd06116UL, 0x21b4f4b5UL, 0x56b3c423UL, 0xcfba9599UL,
+    0xb8bda50fUL, 0x2802b89eUL, 0x5f058808UL, 0xc60cd9b2UL, 0xb10be924UL,
+    0x2f6f7c87UL, 0x58684c11UL, 0xc1611dabUL, 0xb6662d3dUL, 0x76dc4190UL,
+    0x01db7106UL, 0x98d220bcUL, 0xefd5102aUL, 0x71b18589UL, 0x06b6b51fUL,
+    0x9fbfe4a5UL, 0xe8b8d433UL, 0x7807c9a2UL, 0x0f00f934UL, 0x9609a88eUL,
+    0xe10e9818UL, 0x7f6a0dbbUL, 0x086d3d2dUL, 0x91646c97UL, 0xe6635c01UL,
+    0x6b6b51f4UL, 0x1c6c6162UL, 0x856530d8UL, 0xf262004eUL, 0x6c0695edUL,
+    0x1b01a57bUL, 0x8208f4c1UL, 0xf50fc457UL, 0x65b0d9c6UL, 0x12b7e950UL,
+    0x8bbeb8eaUL, 0xfcb9887cUL, 0x62dd1ddfUL, 0x15da2d49UL, 0x8cd37cf3UL,
+    0xfbd44c65UL, 0x4db26158UL, 0x3ab551ceUL, 0xa3bc0074UL, 0xd4bb30e2UL,
+    0x4adfa541UL, 0x3dd895d7UL, 0xa4d1c46dUL, 0xd3d6f4fbUL, 0x4369e96aUL,
+    0x346ed9fcUL, 0xad678846UL, 0xda60b8d0UL, 0x44042d73UL, 0x33031de5UL,
+    0xaa0a4c5fUL, 0xdd0d7cc9UL, 0x5005713cUL, 0x270241aaUL, 0xbe0b1010UL,
+    0xc90c2086UL, 0x5768b525UL, 0x206f85b3UL, 0xb966d409UL, 0xce61e49fUL,
+    0x5edef90eUL, 0x29d9c998UL, 0xb0d09822UL, 0xc7d7a8b4UL, 0x59b33d17UL,
+    0x2eb40d81UL, 0xb7bd5c3bUL, 0xc0ba6cadUL, 0xedb88320UL, 0x9abfb3b6UL,
+    0x03b6e20cUL, 0x74b1d29aUL, 0xead54739UL, 0x9dd277afUL, 0x04db2615UL,
+    0x73dc1683UL, 0xe3630b12UL, 0x94643b84UL, 0x0d6d6a3eUL, 0x7a6a5aa8UL,
+    0xe40ecf0bUL, 0x9309ff9dUL, 0x0a00ae27UL, 0x7d079eb1UL, 0xf00f9344UL,
+    0x8708a3d2UL, 0x1e01f268UL, 0x6906c2feUL, 0xf762575dUL, 0x806567cbUL,
+    0x196c3671UL, 0x6e6b06e7UL, 0xfed41b76UL, 0x89d32be0UL, 0x10da7a5aUL,
+    0x67dd4accUL, 0xf9b9df6fUL, 0x8ebeeff9UL, 0x17b7be43UL, 0x60b08ed5UL,
+    0xd6d6a3e8UL, 0xa1d1937eUL, 0x38d8c2c4UL, 0x4fdff252UL, 0xd1bb67f1UL,
+    0xa6bc5767UL, 0x3fb506ddUL, 0x48b2364bUL, 0xd80d2bdaUL, 0xaf0a1b4cUL,
+    0x36034af6UL, 0x41047a60UL, 0xdf60efc3UL, 0xa867df55UL, 0x316e8eefUL,
+    0x4669be79UL, 0xcb61b38cUL, 0xbc66831aUL, 0x256fd2a0UL, 0x5268e236UL,
+    0xcc0c7795UL, 0xbb0b4703UL, 0x220216b9UL, 0x5505262fUL, 0xc5ba3bbeUL,
+    0xb2bd0b28UL, 0x2bb45a92UL, 0x5cb36a04UL, 0xc2d7ffa7UL, 0xb5d0cf31UL,
+    0x2cd99e8bUL, 0x5bdeae1dUL, 0x9b64c2b0UL, 0xec63f226UL, 0x756aa39cUL,
+    0x026d930aUL, 0x9c0906a9UL, 0xeb0e363fUL, 0x72076785UL, 0x05005713UL,
+    0x95bf4a82UL, 0xe2b87a14UL, 0x7bb12baeUL, 0x0cb61b38UL, 0x92d28e9bUL,
+    0xe5d5be0dUL, 0x7cdcefb7UL, 0x0bdbdf21UL, 0x86d3d2d4UL, 0xf1d4e242UL,
+    0x68ddb3f8UL, 0x1fda836eUL, 0x81be16cdUL, 0xf6b9265bUL, 0x6fb077e1UL,
+    0x18b74777UL, 0x88085ae6UL, 0xff0f6a70UL, 0x66063bcaUL, 0x11010b5cUL,
+    0x8f659effUL, 0xf862ae69UL, 0x616bffd3UL, 0x166ccf45UL, 0xa00ae278UL,
+    0xd70dd2eeUL, 0x4e048354UL, 0x3903b3c2UL, 0xa7672661UL, 0xd06016f7UL,
+    0x4969474dUL, 0x3e6e77dbUL, 0xaed16a4aUL, 0xd9d65adcUL, 0x40df0b66UL,
+    0x37d83bf0UL, 0xa9bcae53UL, 0xdebb9ec5UL, 0x47b2cf7fUL, 0x30b5ffe9UL,
+    0xbdbdf21cUL, 0xcabac28aUL, 0x53b39330UL, 0x24b4a3a6UL, 0xbad03605UL,
+    0xcdd70693UL, 0x54de5729UL, 0x23d967bfUL, 0xb3667a2eUL, 0xc4614ab8UL,
+    0x5d681b02UL, 0x2a6f2b94UL, 0xb40bbe37UL, 0xc30c8ea1UL, 0x5a05df1bUL,
+    0x2d02ef8dUL
+#ifdef BYFOUR
+  },
+  {
+    0x00000000UL, 0x191b3141UL, 0x32366282UL, 0x2b2d53c3UL, 0x646cc504UL,
+    0x7d77f445UL, 0x565aa786UL, 0x4f4196c7UL, 0xc8d98a08UL, 0xd1c2bb49UL,
+    0xfaefe88aUL, 0xe3f4d9cbUL, 0xacb54f0cUL, 0xb5ae7e4dUL, 0x9e832d8eUL,
+    0x87981ccfUL, 0x4ac21251UL, 0x53d92310UL, 0x78f470d3UL, 0x61ef4192UL,
+    0x2eaed755UL, 0x37b5e614UL, 0x1c98b5d7UL, 0x05838496UL, 0x821b9859UL,
+    0x9b00a918UL, 0xb02dfadbUL, 0xa936cb9aUL, 0xe6775d5dUL, 0xff6c6c1cUL,
+    0xd4413fdfUL, 0xcd5a0e9eUL, 0x958424a2UL, 0x8c9f15e3UL, 0xa7b24620UL,
+    0xbea97761UL, 0xf1e8e1a6UL, 0xe8f3d0e7UL, 0xc3de8324UL, 0xdac5b265UL,
+    0x5d5daeaaUL, 0x44469febUL, 0x6f6bcc28UL, 0x7670fd69UL, 0x39316baeUL,
+    0x202a5aefUL, 0x0b07092cUL, 0x121c386dUL, 0xdf4636f3UL, 0xc65d07b2UL,
+    0xed705471UL, 0xf46b6530UL, 0xbb2af3f7UL, 0xa231c2b6UL, 0x891c9175UL,
+    0x9007a034UL, 0x179fbcfbUL, 0x0e848dbaUL, 0x25a9de79UL, 0x3cb2ef38UL,
+    0x73f379ffUL, 0x6ae848beUL, 0x41c51b7dUL, 0x58de2a3cUL, 0xf0794f05UL,
+    0xe9627e44UL, 0xc24f2d87UL, 0xdb541cc6UL, 0x94158a01UL, 0x8d0ebb40UL,
+    0xa623e883UL, 0xbf38d9c2UL, 0x38a0c50dUL, 0x21bbf44cUL, 0x0a96a78fUL,
+    0x138d96ceUL, 0x5ccc0009UL, 0x45d73148UL, 0x6efa628bUL, 0x77e153caUL,
+    0xbabb5d54UL, 0xa3a06c15UL, 0x888d3fd6UL, 0x91960e97UL, 0xded79850UL,
+    0xc7cca911UL, 0xece1fad2UL, 0xf5facb93UL, 0x7262d75cUL, 0x6b79e61dUL,
+    0x4054b5deUL, 0x594f849fUL, 0x160e1258UL, 0x0f152319UL, 0x243870daUL,
+    0x3d23419bUL, 0x65fd6ba7UL, 0x7ce65ae6UL, 0x57cb0925UL, 0x4ed03864UL,
+    0x0191aea3UL, 0x188a9fe2UL, 0x33a7cc21UL, 0x2abcfd60UL, 0xad24e1afUL,
+    0xb43fd0eeUL, 0x9f12832dUL, 0x8609b26cUL, 0xc94824abUL, 0xd05315eaUL,
+    0xfb7e4629UL, 0xe2657768UL, 0x2f3f79f6UL, 0x362448b7UL, 0x1d091b74UL,
+    0x04122a35UL, 0x4b53bcf2UL, 0x52488db3UL, 0x7965de70UL, 0x607eef31UL,
+    0xe7e6f3feUL, 0xfefdc2bfUL, 0xd5d0917cUL, 0xcccba03dUL, 0x838a36faUL,
+    0x9a9107bbUL, 0xb1bc5478UL, 0xa8a76539UL, 0x3b83984bUL, 0x2298a90aUL,
+    0x09b5fac9UL, 0x10aecb88UL, 0x5fef5d4fUL, 0x46f46c0eUL, 0x6dd93fcdUL,
+    0x74c20e8cUL, 0xf35a1243UL, 0xea412302UL, 0xc16c70c1UL, 0xd8774180UL,
+    0x9736d747UL, 0x8e2de606UL, 0xa500b5c5UL, 0xbc1b8484UL, 0x71418a1aUL,
+    0x685abb5bUL, 0x4377e898UL, 0x5a6cd9d9UL, 0x152d4f1eUL, 0x0c367e5fUL,
+    0x271b2d9cUL, 0x3e001cddUL, 0xb9980012UL, 0xa0833153UL, 0x8bae6290UL,
+    0x92b553d1UL, 0xddf4c516UL, 0xc4eff457UL, 0xefc2a794UL, 0xf6d996d5UL,
+    0xae07bce9UL, 0xb71c8da8UL, 0x9c31de6bUL, 0x852aef2aUL, 0xca6b79edUL,
+    0xd37048acUL, 0xf85d1b6fUL, 0xe1462a2eUL, 0x66de36e1UL, 0x7fc507a0UL,
+    0x54e85463UL, 0x4df36522UL, 0x02b2f3e5UL, 0x1ba9c2a4UL, 0x30849167UL,
+    0x299fa026UL, 0xe4c5aeb8UL, 0xfdde9ff9UL, 0xd6f3cc3aUL, 0xcfe8fd7bUL,
+    0x80a96bbcUL, 0x99b25afdUL, 0xb29f093eUL, 0xab84387fUL, 0x2c1c24b0UL,
+    0x350715f1UL, 0x1e2a4632UL, 0x07317773UL, 0x4870e1b4UL, 0x516bd0f5UL,
+    0x7a468336UL, 0x635db277UL, 0xcbfad74eUL, 0xd2e1e60fUL, 0xf9ccb5ccUL,
+    0xe0d7848dUL, 0xaf96124aUL, 0xb68d230bUL, 0x9da070c8UL, 0x84bb4189UL,
+    0x03235d46UL, 0x1a386c07UL, 0x31153fc4UL, 0x280e0e85UL, 0x674f9842UL,
+    0x7e54a903UL, 0x5579fac0UL, 0x4c62cb81UL, 0x8138c51fUL, 0x9823f45eUL,
+    0xb30ea79dUL, 0xaa1596dcUL, 0xe554001bUL, 0xfc4f315aUL, 0xd7626299UL,
+    0xce7953d8UL, 0x49e14f17UL, 0x50fa7e56UL, 0x7bd72d95UL, 0x62cc1cd4UL,
+    0x2d8d8a13UL, 0x3496bb52UL, 0x1fbbe891UL, 0x06a0d9d0UL, 0x5e7ef3ecUL,
+    0x4765c2adUL, 0x6c48916eUL, 0x7553a02fUL, 0x3a1236e8UL, 0x230907a9UL,
+    0x0824546aUL, 0x113f652bUL, 0x96a779e4UL, 0x8fbc48a5UL, 0xa4911b66UL,
+    0xbd8a2a27UL, 0xf2cbbce0UL, 0xebd08da1UL, 0xc0fdde62UL, 0xd9e6ef23UL,
+    0x14bce1bdUL, 0x0da7d0fcUL, 0x268a833fUL, 0x3f91b27eUL, 0x70d024b9UL,
+    0x69cb15f8UL, 0x42e6463bUL, 0x5bfd777aUL, 0xdc656bb5UL, 0xc57e5af4UL,
+    0xee530937UL, 0xf7483876UL, 0xb809aeb1UL, 0xa1129ff0UL, 0x8a3fcc33UL,
+    0x9324fd72UL
+  },
+  {
+    0x00000000UL, 0x01c26a37UL, 0x0384d46eUL, 0x0246be59UL, 0x0709a8dcUL,
+    0x06cbc2ebUL, 0x048d7cb2UL, 0x054f1685UL, 0x0e1351b8UL, 0x0fd13b8fUL,
+    0x0d9785d6UL, 0x0c55efe1UL, 0x091af964UL, 0x08d89353UL, 0x0a9e2d0aUL,
+    0x0b5c473dUL, 0x1c26a370UL, 0x1de4c947UL, 0x1fa2771eUL, 0x1e601d29UL,
+    0x1b2f0bacUL, 0x1aed619bUL, 0x18abdfc2UL, 0x1969b5f5UL, 0x1235f2c8UL,
+    0x13f798ffUL, 0x11b126a6UL, 0x10734c91UL, 0x153c5a14UL, 0x14fe3023UL,
+    0x16b88e7aUL, 0x177ae44dUL, 0x384d46e0UL, 0x398f2cd7UL, 0x3bc9928eUL,
+    0x3a0bf8b9UL, 0x3f44ee3cUL, 0x3e86840bUL, 0x3cc03a52UL, 0x3d025065UL,
+    0x365e1758UL, 0x379c7d6fUL, 0x35dac336UL, 0x3418a901UL, 0x3157bf84UL,
+    0x3095d5b3UL, 0x32d36beaUL, 0x331101ddUL, 0x246be590UL, 0x25a98fa7UL,
+    0x27ef31feUL, 0x262d5bc9UL, 0x23624d4cUL, 0x22a0277bUL, 0x20e69922UL,
+    0x2124f315UL, 0x2a78b428UL, 0x2bbade1fUL, 0x29fc6046UL, 0x283e0a71UL,
+    0x2d711cf4UL, 0x2cb376c3UL, 0x2ef5c89aUL, 0x2f37a2adUL, 0x709a8dc0UL,
+    0x7158e7f7UL, 0x731e59aeUL, 0x72dc3399UL, 0x7793251cUL, 0x76514f2bUL,
+    0x7417f172UL, 0x75d59b45UL, 0x7e89dc78UL, 0x7f4bb64fUL, 0x7d0d0816UL,
+    0x7ccf6221UL, 0x798074a4UL, 0x78421e93UL, 0x7a04a0caUL, 0x7bc6cafdUL,
+    0x6cbc2eb0UL, 0x6d7e4487UL, 0x6f38fadeUL, 0x6efa90e9UL, 0x6bb5866cUL,
+    0x6a77ec5bUL, 0x68315202UL, 0x69f33835UL, 0x62af7f08UL, 0x636d153fUL,
+    0x612bab66UL, 0x60e9c151UL, 0x65a6d7d4UL, 0x6464bde3UL, 0x662203baUL,
+    0x67e0698dUL, 0x48d7cb20UL, 0x4915a117UL, 0x4b531f4eUL, 0x4a917579UL,
+    0x4fde63fcUL, 0x4e1c09cbUL, 0x4c5ab792UL, 0x4d98dda5UL, 0x46c49a98UL,
+    0x4706f0afUL, 0x45404ef6UL, 0x448224c1UL, 0x41cd3244UL, 0x400f5873UL,
+    0x4249e62aUL, 0x438b8c1dUL, 0x54f16850UL, 0x55330267UL, 0x5775bc3eUL,
+    0x56b7d609UL, 0x53f8c08cUL, 0x523aaabbUL, 0x507c14e2UL, 0x51be7ed5UL,
+    0x5ae239e8UL, 0x5b2053dfUL, 0x5966ed86UL, 0x58a487b1UL, 0x5deb9134UL,
+    0x5c29fb03UL, 0x5e6f455aUL, 0x5fad2f6dUL, 0xe1351b80UL, 0xe0f771b7UL,
+    0xe2b1cfeeUL, 0xe373a5d9UL, 0xe63cb35cUL, 0xe7fed96bUL, 0xe5b86732UL,
+    0xe47a0d05UL, 0xef264a38UL, 0xeee4200fUL, 0xeca29e56UL, 0xed60f461UL,
+    0xe82fe2e4UL, 0xe9ed88d3UL, 0xebab368aUL, 0xea695cbdUL, 0xfd13b8f0UL,
+    0xfcd1d2c7UL, 0xfe976c9eUL, 0xff5506a9UL, 0xfa1a102cUL, 0xfbd87a1bUL,
+    0xf99ec442UL, 0xf85cae75UL, 0xf300e948UL, 0xf2c2837fUL, 0xf0843d26UL,
+    0xf1465711UL, 0xf4094194UL, 0xf5cb2ba3UL, 0xf78d95faUL, 0xf64fffcdUL,
+    0xd9785d60UL, 0xd8ba3757UL, 0xdafc890eUL, 0xdb3ee339UL, 0xde71f5bcUL,
+    0xdfb39f8bUL, 0xddf521d2UL, 0xdc374be5UL, 0xd76b0cd8UL, 0xd6a966efUL,
+    0xd4efd8b6UL, 0xd52db281UL, 0xd062a404UL, 0xd1a0ce33UL, 0xd3e6706aUL,
+    0xd2241a5dUL, 0xc55efe10UL, 0xc49c9427UL, 0xc6da2a7eUL, 0xc7184049UL,
+    0xc25756ccUL, 0xc3953cfbUL, 0xc1d382a2UL, 0xc011e895UL, 0xcb4dafa8UL,
+    0xca8fc59fUL, 0xc8c97bc6UL, 0xc90b11f1UL, 0xcc440774UL, 0xcd866d43UL,
+    0xcfc0d31aUL, 0xce02b92dUL, 0x91af9640UL, 0x906dfc77UL, 0x922b422eUL,
+    0x93e92819UL, 0x96a63e9cUL, 0x976454abUL, 0x9522eaf2UL, 0x94e080c5UL,
+    0x9fbcc7f8UL, 0x9e7eadcfUL, 0x9c381396UL, 0x9dfa79a1UL, 0x98b56f24UL,
+    0x99770513UL, 0x9b31bb4aUL, 0x9af3d17dUL, 0x8d893530UL, 0x8c4b5f07UL,
+    0x8e0de15eUL, 0x8fcf8b69UL, 0x8a809decUL, 0x8b42f7dbUL, 0x89044982UL,
+    0x88c623b5UL, 0x839a6488UL, 0x82580ebfUL, 0x801eb0e6UL, 0x81dcdad1UL,
+    0x8493cc54UL, 0x8551a663UL, 0x8717183aUL, 0x86d5720dUL, 0xa9e2d0a0UL,
+    0xa820ba97UL, 0xaa6604ceUL, 0xaba46ef9UL, 0xaeeb787cUL, 0xaf29124bUL,
+    0xad6fac12UL, 0xacadc625UL, 0xa7f18118UL, 0xa633eb2fUL, 0xa4755576UL,
+    0xa5b73f41UL, 0xa0f829c4UL, 0xa13a43f3UL, 0xa37cfdaaUL, 0xa2be979dUL,
+    0xb5c473d0UL, 0xb40619e7UL, 0xb640a7beUL, 0xb782cd89UL, 0xb2cddb0cUL,
+    0xb30fb13bUL, 0xb1490f62UL, 0xb08b6555UL, 0xbbd72268UL, 0xba15485fUL,
+    0xb853f606UL, 0xb9919c31UL, 0xbcde8ab4UL, 0xbd1ce083UL, 0xbf5a5edaUL,
+    0xbe9834edUL
+  },
+  {
+    0x00000000UL, 0xb8bc6765UL, 0xaa09c88bUL, 0x12b5afeeUL, 0x8f629757UL,
+    0x37def032UL, 0x256b5fdcUL, 0x9dd738b9UL, 0xc5b428efUL, 0x7d084f8aUL,
+    0x6fbde064UL, 0xd7018701UL, 0x4ad6bfb8UL, 0xf26ad8ddUL, 0xe0df7733UL,
+    0x58631056UL, 0x5019579fUL, 0xe8a530faUL, 0xfa109f14UL, 0x42acf871UL,
+    0xdf7bc0c8UL, 0x67c7a7adUL, 0x75720843UL, 0xcdce6f26UL, 0x95ad7f70UL,
+    0x2d111815UL, 0x3fa4b7fbUL, 0x8718d09eUL, 0x1acfe827UL, 0xa2738f42UL,
+    0xb0c620acUL, 0x087a47c9UL, 0xa032af3eUL, 0x188ec85bUL, 0x0a3b67b5UL,
+    0xb28700d0UL, 0x2f503869UL, 0x97ec5f0cUL, 0x8559f0e2UL, 0x3de59787UL,
+    0x658687d1UL, 0xdd3ae0b4UL, 0xcf8f4f5aUL, 0x7733283fUL, 0xeae41086UL,
+    0x525877e3UL, 0x40edd80dUL, 0xf851bf68UL, 0xf02bf8a1UL, 0x48979fc4UL,
+    0x5a22302aUL, 0xe29e574fUL, 0x7f496ff6UL, 0xc7f50893UL, 0xd540a77dUL,
+    0x6dfcc018UL, 0x359fd04eUL, 0x8d23b72bUL, 0x9f9618c5UL, 0x272a7fa0UL,
+    0xbafd4719UL, 0x0241207cUL, 0x10f48f92UL, 0xa848e8f7UL, 0x9b14583dUL,
+    0x23a83f58UL, 0x311d90b6UL, 0x89a1f7d3UL, 0x1476cf6aUL, 0xaccaa80fUL,
+    0xbe7f07e1UL, 0x06c36084UL, 0x5ea070d2UL, 0xe61c17b7UL, 0xf4a9b859UL,
+    0x4c15df3cUL, 0xd1c2e785UL, 0x697e80e0UL, 0x7bcb2f0eUL, 0xc377486bUL,
+    0xcb0d0fa2UL, 0x73b168c7UL, 0x6104c729UL, 0xd9b8a04cUL, 0x446f98f5UL,
+    0xfcd3ff90UL, 0xee66507eUL, 0x56da371bUL, 0x0eb9274dUL, 0xb6054028UL,
+    0xa4b0efc6UL, 0x1c0c88a3UL, 0x81dbb01aUL, 0x3967d77fUL, 0x2bd27891UL,
+    0x936e1ff4UL, 0x3b26f703UL, 0x839a9066UL, 0x912f3f88UL, 0x299358edUL,
+    0xb4446054UL, 0x0cf80731UL, 0x1e4da8dfUL, 0xa6f1cfbaUL, 0xfe92dfecUL,
+    0x462eb889UL, 0x549b1767UL, 0xec277002UL, 0x71f048bbUL, 0xc94c2fdeUL,
+    0xdbf98030UL, 0x6345e755UL, 0x6b3fa09cUL, 0xd383c7f9UL, 0xc1366817UL,
+    0x798a0f72UL, 0xe45d37cbUL, 0x5ce150aeUL, 0x4e54ff40UL, 0xf6e89825UL,
+    0xae8b8873UL, 0x1637ef16UL, 0x048240f8UL, 0xbc3e279dUL, 0x21e91f24UL,
+    0x99557841UL, 0x8be0d7afUL, 0x335cb0caUL, 0xed59b63bUL, 0x55e5d15eUL,
+    0x47507eb0UL, 0xffec19d5UL, 0x623b216cUL, 0xda874609UL, 0xc832e9e7UL,
+    0x708e8e82UL, 0x28ed9ed4UL, 0x9051f9b1UL, 0x82e4565fUL, 0x3a58313aUL,
+    0xa78f0983UL, 0x1f336ee6UL, 0x0d86c108UL, 0xb53aa66dUL, 0xbd40e1a4UL,
+    0x05fc86c1UL, 0x1749292fUL, 0xaff54e4aUL, 0x322276f3UL, 0x8a9e1196UL,
+    0x982bbe78UL, 0x2097d91dUL, 0x78f4c94bUL, 0xc048ae2eUL, 0xd2fd01c0UL,
+    0x6a4166a5UL, 0xf7965e1cUL, 0x4f2a3979UL, 0x5d9f9697UL, 0xe523f1f2UL,
+    0x4d6b1905UL, 0xf5d77e60UL, 0xe762d18eUL, 0x5fdeb6ebUL, 0xc2098e52UL,
+    0x7ab5e937UL, 0x680046d9UL, 0xd0bc21bcUL, 0x88df31eaUL, 0x3063568fUL,
+    0x22d6f961UL, 0x9a6a9e04UL, 0x07bda6bdUL, 0xbf01c1d8UL, 0xadb46e36UL,
+    0x15080953UL, 0x1d724e9aUL, 0xa5ce29ffUL, 0xb77b8611UL, 0x0fc7e174UL,
+    0x9210d9cdUL, 0x2aacbea8UL, 0x38191146UL, 0x80a57623UL, 0xd8c66675UL,
+    0x607a0110UL, 0x72cfaefeUL, 0xca73c99bUL, 0x57a4f122UL, 0xef189647UL,
+    0xfdad39a9UL, 0x45115eccUL, 0x764dee06UL, 0xcef18963UL, 0xdc44268dUL,
+    0x64f841e8UL, 0xf92f7951UL, 0x41931e34UL, 0x5326b1daUL, 0xeb9ad6bfUL,
+    0xb3f9c6e9UL, 0x0b45a18cUL, 0x19f00e62UL, 0xa14c6907UL, 0x3c9b51beUL,
+    0x842736dbUL, 0x96929935UL, 0x2e2efe50UL, 0x2654b999UL, 0x9ee8defcUL,
+    0x8c5d7112UL, 0x34e11677UL, 0xa9362eceUL, 0x118a49abUL, 0x033fe645UL,
+    0xbb838120UL, 0xe3e09176UL, 0x5b5cf613UL, 0x49e959fdUL, 0xf1553e98UL,
+    0x6c820621UL, 0xd43e6144UL, 0xc68bceaaUL, 0x7e37a9cfUL, 0xd67f4138UL,
+    0x6ec3265dUL, 0x7c7689b3UL, 0xc4caeed6UL, 0x591dd66fUL, 0xe1a1b10aUL,
+    0xf3141ee4UL, 0x4ba87981UL, 0x13cb69d7UL, 0xab770eb2UL, 0xb9c2a15cUL,
+    0x017ec639UL, 0x9ca9fe80UL, 0x241599e5UL, 0x36a0360bUL, 0x8e1c516eUL,
+    0x866616a7UL, 0x3eda71c2UL, 0x2c6fde2cUL, 0x94d3b949UL, 0x090481f0UL,
+    0xb1b8e695UL, 0xa30d497bUL, 0x1bb12e1eUL, 0x43d23e48UL, 0xfb6e592dUL,
+    0xe9dbf6c3UL, 0x516791a6UL, 0xccb0a91fUL, 0x740cce7aUL, 0x66b96194UL,
+    0xde0506f1UL
+  },
+  {
+    0x00000000UL, 0x96300777UL, 0x2c610eeeUL, 0xba510999UL, 0x19c46d07UL,
+    0x8ff46a70UL, 0x35a563e9UL, 0xa395649eUL, 0x3288db0eUL, 0xa4b8dc79UL,
+    0x1ee9d5e0UL, 0x88d9d297UL, 0x2b4cb609UL, 0xbd7cb17eUL, 0x072db8e7UL,
+    0x911dbf90UL, 0x6410b71dUL, 0xf220b06aUL, 0x4871b9f3UL, 0xde41be84UL,
+    0x7dd4da1aUL, 0xebe4dd6dUL, 0x51b5d4f4UL, 0xc785d383UL, 0x56986c13UL,
+    0xc0a86b64UL, 0x7af962fdUL, 0xecc9658aUL, 0x4f5c0114UL, 0xd96c0663UL,
+    0x633d0ffaUL, 0xf50d088dUL, 0xc8206e3bUL, 0x5e10694cUL, 0xe44160d5UL,
+    0x727167a2UL, 0xd1e4033cUL, 0x47d4044bUL, 0xfd850dd2UL, 0x6bb50aa5UL,
+    0xfaa8b535UL, 0x6c98b242UL, 0xd6c9bbdbUL, 0x40f9bcacUL, 0xe36cd832UL,
+    0x755cdf45UL, 0xcf0dd6dcUL, 0x593dd1abUL, 0xac30d926UL, 0x3a00de51UL,
+    0x8051d7c8UL, 0x1661d0bfUL, 0xb5f4b421UL, 0x23c4b356UL, 0x9995bacfUL,
+    0x0fa5bdb8UL, 0x9eb80228UL, 0x0888055fUL, 0xb2d90cc6UL, 0x24e90bb1UL,
+    0x877c6f2fUL, 0x114c6858UL, 0xab1d61c1UL, 0x3d2d66b6UL, 0x9041dc76UL,
+    0x0671db01UL, 0xbc20d298UL, 0x2a10d5efUL, 0x8985b171UL, 0x1fb5b606UL,
+    0xa5e4bf9fUL, 0x33d4b8e8UL, 0xa2c90778UL, 0x34f9000fUL, 0x8ea80996UL,
+    0x18980ee1UL, 0xbb0d6a7fUL, 0x2d3d6d08UL, 0x976c6491UL, 0x015c63e6UL,
+    0xf4516b6bUL, 0x62616c1cUL, 0xd8306585UL, 0x4e0062f2UL, 0xed95066cUL,
+    0x7ba5011bUL, 0xc1f40882UL, 0x57c40ff5UL, 0xc6d9b065UL, 0x50e9b712UL,
+    0xeab8be8bUL, 0x7c88b9fcUL, 0xdf1ddd62UL, 0x492dda15UL, 0xf37cd38cUL,
+    0x654cd4fbUL, 0x5861b24dUL, 0xce51b53aUL, 0x7400bca3UL, 0xe230bbd4UL,
+    0x41a5df4aUL, 0xd795d83dUL, 0x6dc4d1a4UL, 0xfbf4d6d3UL, 0x6ae96943UL,
+    0xfcd96e34UL, 0x468867adUL, 0xd0b860daUL, 0x732d0444UL, 0xe51d0333UL,
+    0x5f4c0aaaUL, 0xc97c0dddUL, 0x3c710550UL, 0xaa410227UL, 0x10100bbeUL,
+    0x86200cc9UL, 0x25b56857UL, 0xb3856f20UL, 0x09d466b9UL, 0x9fe461ceUL,
+    0x0ef9de5eUL, 0x98c9d929UL, 0x2298d0b0UL, 0xb4a8d7c7UL, 0x173db359UL,
+    0x810db42eUL, 0x3b5cbdb7UL, 0xad6cbac0UL, 0x2083b8edUL, 0xb6b3bf9aUL,
+    0x0ce2b603UL, 0x9ad2b174UL, 0x3947d5eaUL, 0xaf77d29dUL, 0x1526db04UL,
+    0x8316dc73UL, 0x120b63e3UL, 0x843b6494UL, 0x3e6a6d0dUL, 0xa85a6a7aUL,
+    0x0bcf0ee4UL, 0x9dff0993UL, 0x27ae000aUL, 0xb19e077dUL, 0x44930ff0UL,
+    0xd2a30887UL, 0x68f2011eUL, 0xfec20669UL, 0x5d5762f7UL, 0xcb676580UL,
+    0x71366c19UL, 0xe7066b6eUL, 0x761bd4feUL, 0xe02bd389UL, 0x5a7ada10UL,
+    0xcc4add67UL, 0x6fdfb9f9UL, 0xf9efbe8eUL, 0x43beb717UL, 0xd58eb060UL,
+    0xe8a3d6d6UL, 0x7e93d1a1UL, 0xc4c2d838UL, 0x52f2df4fUL, 0xf167bbd1UL,
+    0x6757bca6UL, 0xdd06b53fUL, 0x4b36b248UL, 0xda2b0dd8UL, 0x4c1b0aafUL,
+    0xf64a0336UL, 0x607a0441UL, 0xc3ef60dfUL, 0x55df67a8UL, 0xef8e6e31UL,
+    0x79be6946UL, 0x8cb361cbUL, 0x1a8366bcUL, 0xa0d26f25UL, 0x36e26852UL,
+    0x95770cccUL, 0x03470bbbUL, 0xb9160222UL, 0x2f260555UL, 0xbe3bbac5UL,
+    0x280bbdb2UL, 0x925ab42bUL, 0x046ab35cUL, 0xa7ffd7c2UL, 0x31cfd0b5UL,
+    0x8b9ed92cUL, 0x1daede5bUL, 0xb0c2649bUL, 0x26f263ecUL, 0x9ca36a75UL,
+    0x0a936d02UL, 0xa906099cUL, 0x3f360eebUL, 0x85670772UL, 0x13570005UL,
+    0x824abf95UL, 0x147ab8e2UL, 0xae2bb17bUL, 0x381bb60cUL, 0x9b8ed292UL,
+    0x0dbed5e5UL, 0xb7efdc7cUL, 0x21dfdb0bUL, 0xd4d2d386UL, 0x42e2d4f1UL,
+    0xf8b3dd68UL, 0x6e83da1fUL, 0xcd16be81UL, 0x5b26b9f6UL, 0xe177b06fUL,
+    0x7747b718UL, 0xe65a0888UL, 0x706a0fffUL, 0xca3b0666UL, 0x5c0b0111UL,
+    0xff9e658fUL, 0x69ae62f8UL, 0xd3ff6b61UL, 0x45cf6c16UL, 0x78e20aa0UL,
+    0xeed20dd7UL, 0x5483044eUL, 0xc2b30339UL, 0x612667a7UL, 0xf71660d0UL,
+    0x4d476949UL, 0xdb776e3eUL, 0x4a6ad1aeUL, 0xdc5ad6d9UL, 0x660bdf40UL,
+    0xf03bd837UL, 0x53aebca9UL, 0xc59ebbdeUL, 0x7fcfb247UL, 0xe9ffb530UL,
+    0x1cf2bdbdUL, 0x8ac2bacaUL, 0x3093b353UL, 0xa6a3b424UL, 0x0536d0baUL,
+    0x9306d7cdUL, 0x2957de54UL, 0xbf67d923UL, 0x2e7a66b3UL, 0xb84a61c4UL,
+    0x021b685dUL, 0x942b6f2aUL, 0x37be0bb4UL, 0xa18e0cc3UL, 0x1bdf055aUL,
+    0x8def022dUL
+  },
+  {
+    0x00000000UL, 0x41311b19UL, 0x82623632UL, 0xc3532d2bUL, 0x04c56c64UL,
+    0x45f4777dUL, 0x86a75a56UL, 0xc796414fUL, 0x088ad9c8UL, 0x49bbc2d1UL,
+    0x8ae8effaUL, 0xcbd9f4e3UL, 0x0c4fb5acUL, 0x4d7eaeb5UL, 0x8e2d839eUL,
+    0xcf1c9887UL, 0x5112c24aUL, 0x1023d953UL, 0xd370f478UL, 0x9241ef61UL,
+    0x55d7ae2eUL, 0x14e6b537UL, 0xd7b5981cUL, 0x96848305UL, 0x59981b82UL,
+    0x18a9009bUL, 0xdbfa2db0UL, 0x9acb36a9UL, 0x5d5d77e6UL, 0x1c6c6cffUL,
+    0xdf3f41d4UL, 0x9e0e5acdUL, 0xa2248495UL, 0xe3159f8cUL, 0x2046b2a7UL,
+    0x6177a9beUL, 0xa6e1e8f1UL, 0xe7d0f3e8UL, 0x2483dec3UL, 0x65b2c5daUL,
+    0xaaae5d5dUL, 0xeb9f4644UL, 0x28cc6b6fUL, 0x69fd7076UL, 0xae6b3139UL,
+    0xef5a2a20UL, 0x2c09070bUL, 0x6d381c12UL, 0xf33646dfUL, 0xb2075dc6UL,
+    0x715470edUL, 0x30656bf4UL, 0xf7f32abbUL, 0xb6c231a2UL, 0x75911c89UL,
+    0x34a00790UL, 0xfbbc9f17UL, 0xba8d840eUL, 0x79dea925UL, 0x38efb23cUL,
+    0xff79f373UL, 0xbe48e86aUL, 0x7d1bc541UL, 0x3c2ade58UL, 0x054f79f0UL,
+    0x447e62e9UL, 0x872d4fc2UL, 0xc61c54dbUL, 0x018a1594UL, 0x40bb0e8dUL,
+    0x83e823a6UL, 0xc2d938bfUL, 0x0dc5a038UL, 0x4cf4bb21UL, 0x8fa7960aUL,
+    0xce968d13UL, 0x0900cc5cUL, 0x4831d745UL, 0x8b62fa6eUL, 0xca53e177UL,
+    0x545dbbbaUL, 0x156ca0a3UL, 0xd63f8d88UL, 0x970e9691UL, 0x5098d7deUL,
+    0x11a9ccc7UL, 0xd2fae1ecUL, 0x93cbfaf5UL, 0x5cd76272UL, 0x1de6796bUL,
+    0xdeb55440UL, 0x9f844f59UL, 0x58120e16UL, 0x1923150fUL, 0xda703824UL,
+    0x9b41233dUL, 0xa76bfd65UL, 0xe65ae67cUL, 0x2509cb57UL, 0x6438d04eUL,
+    0xa3ae9101UL, 0xe29f8a18UL, 0x21cca733UL, 0x60fdbc2aUL, 0xafe124adUL,
+    0xeed03fb4UL, 0x2d83129fUL, 0x6cb20986UL, 0xab2448c9UL, 0xea1553d0UL,
+    0x29467efbUL, 0x687765e2UL, 0xf6793f2fUL, 0xb7482436UL, 0x741b091dUL,
+    0x352a1204UL, 0xf2bc534bUL, 0xb38d4852UL, 0x70de6579UL, 0x31ef7e60UL,
+    0xfef3e6e7UL, 0xbfc2fdfeUL, 0x7c91d0d5UL, 0x3da0cbccUL, 0xfa368a83UL,
+    0xbb07919aUL, 0x7854bcb1UL, 0x3965a7a8UL, 0x4b98833bUL, 0x0aa99822UL,
+    0xc9fab509UL, 0x88cbae10UL, 0x4f5def5fUL, 0x0e6cf446UL, 0xcd3fd96dUL,
+    0x8c0ec274UL, 0x43125af3UL, 0x022341eaUL, 0xc1706cc1UL, 0x804177d8UL,
+    0x47d73697UL, 0x06e62d8eUL, 0xc5b500a5UL, 0x84841bbcUL, 0x1a8a4171UL,
+    0x5bbb5a68UL, 0x98e87743UL, 0xd9d96c5aUL, 0x1e4f2d15UL, 0x5f7e360cUL,
+    0x9c2d1b27UL, 0xdd1c003eUL, 0x120098b9UL, 0x533183a0UL, 0x9062ae8bUL,
+    0xd153b592UL, 0x16c5f4ddUL, 0x57f4efc4UL, 0x94a7c2efUL, 0xd596d9f6UL,
+    0xe9bc07aeUL, 0xa88d1cb7UL, 0x6bde319cUL, 0x2aef2a85UL, 0xed796bcaUL,
+    0xac4870d3UL, 0x6f1b5df8UL, 0x2e2a46e1UL, 0xe136de66UL, 0xa007c57fUL,
+    0x6354e854UL, 0x2265f34dUL, 0xe5f3b202UL, 0xa4c2a91bUL, 0x67918430UL,
+    0x26a09f29UL, 0xb8aec5e4UL, 0xf99fdefdUL, 0x3accf3d6UL, 0x7bfde8cfUL,
+    0xbc6ba980UL, 0xfd5ab299UL, 0x3e099fb2UL, 0x7f3884abUL, 0xb0241c2cUL,
+    0xf1150735UL, 0x32462a1eUL, 0x73773107UL, 0xb4e17048UL, 0xf5d06b51UL,
+    0x3683467aUL, 0x77b25d63UL, 0x4ed7facbUL, 0x0fe6e1d2UL, 0xccb5ccf9UL,
+    0x8d84d7e0UL, 0x4a1296afUL, 0x0b238db6UL, 0xc870a09dUL, 0x8941bb84UL,
+    0x465d2303UL, 0x076c381aUL, 0xc43f1531UL, 0x850e0e28UL, 0x42984f67UL,
+    0x03a9547eUL, 0xc0fa7955UL, 0x81cb624cUL, 0x1fc53881UL, 0x5ef42398UL,
+    0x9da70eb3UL, 0xdc9615aaUL, 0x1b0054e5UL, 0x5a314ffcUL, 0x996262d7UL,
+    0xd85379ceUL, 0x174fe149UL, 0x567efa50UL, 0x952dd77bUL, 0xd41ccc62UL,
+    0x138a8d2dUL, 0x52bb9634UL, 0x91e8bb1fUL, 0xd0d9a006UL, 0xecf37e5eUL,
+    0xadc26547UL, 0x6e91486cUL, 0x2fa05375UL, 0xe836123aUL, 0xa9070923UL,
+    0x6a542408UL, 0x2b653f11UL, 0xe479a796UL, 0xa548bc8fUL, 0x661b91a4UL,
+    0x272a8abdUL, 0xe0bccbf2UL, 0xa18dd0ebUL, 0x62defdc0UL, 0x23efe6d9UL,
+    0xbde1bc14UL, 0xfcd0a70dUL, 0x3f838a26UL, 0x7eb2913fUL, 0xb924d070UL,
+    0xf815cb69UL, 0x3b46e642UL, 0x7a77fd5bUL, 0xb56b65dcUL, 0xf45a7ec5UL,
+    0x370953eeUL, 0x763848f7UL, 0xb1ae09b8UL, 0xf09f12a1UL, 0x33cc3f8aUL,
+    0x72fd2493UL
+  },
+  {
+    0x00000000UL, 0x376ac201UL, 0x6ed48403UL, 0x59be4602UL, 0xdca80907UL,
+    0xebc2cb06UL, 0xb27c8d04UL, 0x85164f05UL, 0xb851130eUL, 0x8f3bd10fUL,
+    0xd685970dUL, 0xe1ef550cUL, 0x64f91a09UL, 0x5393d808UL, 0x0a2d9e0aUL,
+    0x3d475c0bUL, 0x70a3261cUL, 0x47c9e41dUL, 0x1e77a21fUL, 0x291d601eUL,
+    0xac0b2f1bUL, 0x9b61ed1aUL, 0xc2dfab18UL, 0xf5b56919UL, 0xc8f23512UL,
+    0xff98f713UL, 0xa626b111UL, 0x914c7310UL, 0x145a3c15UL, 0x2330fe14UL,
+    0x7a8eb816UL, 0x4de47a17UL, 0xe0464d38UL, 0xd72c8f39UL, 0x8e92c93bUL,
+    0xb9f80b3aUL, 0x3cee443fUL, 0x0b84863eUL, 0x523ac03cUL, 0x6550023dUL,
+    0x58175e36UL, 0x6f7d9c37UL, 0x36c3da35UL, 0x01a91834UL, 0x84bf5731UL,
+    0xb3d59530UL, 0xea6bd332UL, 0xdd011133UL, 0x90e56b24UL, 0xa78fa925UL,
+    0xfe31ef27UL, 0xc95b2d26UL, 0x4c4d6223UL, 0x7b27a022UL, 0x2299e620UL,
+    0x15f32421UL, 0x28b4782aUL, 0x1fdeba2bUL, 0x4660fc29UL, 0x710a3e28UL,
+    0xf41c712dUL, 0xc376b32cUL, 0x9ac8f52eUL, 0xada2372fUL, 0xc08d9a70UL,
+    0xf7e75871UL, 0xae591e73UL, 0x9933dc72UL, 0x1c259377UL, 0x2b4f5176UL,
+    0x72f11774UL, 0x459bd575UL, 0x78dc897eUL, 0x4fb64b7fUL, 0x16080d7dUL,
+    0x2162cf7cUL, 0xa4748079UL, 0x931e4278UL, 0xcaa0047aUL, 0xfdcac67bUL,
+    0xb02ebc6cUL, 0x87447e6dUL, 0xdefa386fUL, 0xe990fa6eUL, 0x6c86b56bUL,
+    0x5bec776aUL, 0x02523168UL, 0x3538f369UL, 0x087faf62UL, 0x3f156d63UL,
+    0x66ab2b61UL, 0x51c1e960UL, 0xd4d7a665UL, 0xe3bd6464UL, 0xba032266UL,
+    0x8d69e067UL, 0x20cbd748UL, 0x17a11549UL, 0x4e1f534bUL, 0x7975914aUL,
+    0xfc63de4fUL, 0xcb091c4eUL, 0x92b75a4cUL, 0xa5dd984dUL, 0x989ac446UL,
+    0xaff00647UL, 0xf64e4045UL, 0xc1248244UL, 0x4432cd41UL, 0x73580f40UL,
+    0x2ae64942UL, 0x1d8c8b43UL, 0x5068f154UL, 0x67023355UL, 0x3ebc7557UL,
+    0x09d6b756UL, 0x8cc0f853UL, 0xbbaa3a52UL, 0xe2147c50UL, 0xd57ebe51UL,
+    0xe839e25aUL, 0xdf53205bUL, 0x86ed6659UL, 0xb187a458UL, 0x3491eb5dUL,
+    0x03fb295cUL, 0x5a456f5eUL, 0x6d2fad5fUL, 0x801b35e1UL, 0xb771f7e0UL,
+    0xeecfb1e2UL, 0xd9a573e3UL, 0x5cb33ce6UL, 0x6bd9fee7UL, 0x3267b8e5UL,
+    0x050d7ae4UL, 0x384a26efUL, 0x0f20e4eeUL, 0x569ea2ecUL, 0x61f460edUL,
+    0xe4e22fe8UL, 0xd388ede9UL, 0x8a36abebUL, 0xbd5c69eaUL, 0xf0b813fdUL,
+    0xc7d2d1fcUL, 0x9e6c97feUL, 0xa90655ffUL, 0x2c101afaUL, 0x1b7ad8fbUL,
+    0x42c49ef9UL, 0x75ae5cf8UL, 0x48e900f3UL, 0x7f83c2f2UL, 0x263d84f0UL,
+    0x115746f1UL, 0x944109f4UL, 0xa32bcbf5UL, 0xfa958df7UL, 0xcdff4ff6UL,
+    0x605d78d9UL, 0x5737bad8UL, 0x0e89fcdaUL, 0x39e33edbUL, 0xbcf571deUL,
+    0x8b9fb3dfUL, 0xd221f5ddUL, 0xe54b37dcUL, 0xd80c6bd7UL, 0xef66a9d6UL,
+    0xb6d8efd4UL, 0x81b22dd5UL, 0x04a462d0UL, 0x33cea0d1UL, 0x6a70e6d3UL,
+    0x5d1a24d2UL, 0x10fe5ec5UL, 0x27949cc4UL, 0x7e2adac6UL, 0x494018c7UL,
+    0xcc5657c2UL, 0xfb3c95c3UL, 0xa282d3c1UL, 0x95e811c0UL, 0xa8af4dcbUL,
+    0x9fc58fcaUL, 0xc67bc9c8UL, 0xf1110bc9UL, 0x740744ccUL, 0x436d86cdUL,
+    0x1ad3c0cfUL, 0x2db902ceUL, 0x4096af91UL, 0x77fc6d90UL, 0x2e422b92UL,
+    0x1928e993UL, 0x9c3ea696UL, 0xab546497UL, 0xf2ea2295UL, 0xc580e094UL,
+    0xf8c7bc9fUL, 0xcfad7e9eUL, 0x9613389cUL, 0xa179fa9dUL, 0x246fb598UL,
+    0x13057799UL, 0x4abb319bUL, 0x7dd1f39aUL, 0x3035898dUL, 0x075f4b8cUL,
+    0x5ee10d8eUL, 0x698bcf8fUL, 0xec9d808aUL, 0xdbf7428bUL, 0x82490489UL,
+    0xb523c688UL, 0x88649a83UL, 0xbf0e5882UL, 0xe6b01e80UL, 0xd1dadc81UL,
+    0x54cc9384UL, 0x63a65185UL, 0x3a181787UL, 0x0d72d586UL, 0xa0d0e2a9UL,
+    0x97ba20a8UL, 0xce0466aaUL, 0xf96ea4abUL, 0x7c78ebaeUL, 0x4b1229afUL,
+    0x12ac6fadUL, 0x25c6adacUL, 0x1881f1a7UL, 0x2feb33a6UL, 0x765575a4UL,
+    0x413fb7a5UL, 0xc429f8a0UL, 0xf3433aa1UL, 0xaafd7ca3UL, 0x9d97bea2UL,
+    0xd073c4b5UL, 0xe71906b4UL, 0xbea740b6UL, 0x89cd82b7UL, 0x0cdbcdb2UL,
+    0x3bb10fb3UL, 0x620f49b1UL, 0x55658bb0UL, 0x6822d7bbUL, 0x5f4815baUL,
+    0x06f653b8UL, 0x319c91b9UL, 0xb48adebcUL, 0x83e01cbdUL, 0xda5e5abfUL,
+    0xed3498beUL
+  },
+  {
+    0x00000000UL, 0x6567bcb8UL, 0x8bc809aaUL, 0xeeafb512UL, 0x5797628fUL,
+    0x32f0de37UL, 0xdc5f6b25UL, 0xb938d79dUL, 0xef28b4c5UL, 0x8a4f087dUL,
+    0x64e0bd6fUL, 0x018701d7UL, 0xb8bfd64aUL, 0xddd86af2UL, 0x3377dfe0UL,
+    0x56106358UL, 0x9f571950UL, 0xfa30a5e8UL, 0x149f10faUL, 0x71f8ac42UL,
+    0xc8c07bdfUL, 0xada7c767UL, 0x43087275UL, 0x266fcecdUL, 0x707fad95UL,
+    0x1518112dUL, 0xfbb7a43fUL, 0x9ed01887UL, 0x27e8cf1aUL, 0x428f73a2UL,
+    0xac20c6b0UL, 0xc9477a08UL, 0x3eaf32a0UL, 0x5bc88e18UL, 0xb5673b0aUL,
+    0xd00087b2UL, 0x6938502fUL, 0x0c5fec97UL, 0xe2f05985UL, 0x8797e53dUL,
+    0xd1878665UL, 0xb4e03addUL, 0x5a4f8fcfUL, 0x3f283377UL, 0x8610e4eaUL,
+    0xe3775852UL, 0x0dd8ed40UL, 0x68bf51f8UL, 0xa1f82bf0UL, 0xc49f9748UL,
+    0x2a30225aUL, 0x4f579ee2UL, 0xf66f497fUL, 0x9308f5c7UL, 0x7da740d5UL,
+    0x18c0fc6dUL, 0x4ed09f35UL, 0x2bb7238dUL, 0xc518969fUL, 0xa07f2a27UL,
+    0x1947fdbaUL, 0x7c204102UL, 0x928ff410UL, 0xf7e848a8UL, 0x3d58149bUL,
+    0x583fa823UL, 0xb6901d31UL, 0xd3f7a189UL, 0x6acf7614UL, 0x0fa8caacUL,
+    0xe1077fbeUL, 0x8460c306UL, 0xd270a05eUL, 0xb7171ce6UL, 0x59b8a9f4UL,
+    0x3cdf154cUL, 0x85e7c2d1UL, 0xe0807e69UL, 0x0e2fcb7bUL, 0x6b4877c3UL,
+    0xa20f0dcbUL, 0xc768b173UL, 0x29c70461UL, 0x4ca0b8d9UL, 0xf5986f44UL,
+    0x90ffd3fcUL, 0x7e5066eeUL, 0x1b37da56UL, 0x4d27b90eUL, 0x284005b6UL,
+    0xc6efb0a4UL, 0xa3880c1cUL, 0x1ab0db81UL, 0x7fd76739UL, 0x9178d22bUL,
+    0xf41f6e93UL, 0x03f7263bUL, 0x66909a83UL, 0x883f2f91UL, 0xed589329UL,
+    0x546044b4UL, 0x3107f80cUL, 0xdfa84d1eUL, 0xbacff1a6UL, 0xecdf92feUL,
+    0x89b82e46UL, 0x67179b54UL, 0x027027ecUL, 0xbb48f071UL, 0xde2f4cc9UL,
+    0x3080f9dbUL, 0x55e74563UL, 0x9ca03f6bUL, 0xf9c783d3UL, 0x176836c1UL,
+    0x720f8a79UL, 0xcb375de4UL, 0xae50e15cUL, 0x40ff544eUL, 0x2598e8f6UL,
+    0x73888baeUL, 0x16ef3716UL, 0xf8408204UL, 0x9d273ebcUL, 0x241fe921UL,
+    0x41785599UL, 0xafd7e08bUL, 0xcab05c33UL, 0x3bb659edUL, 0x5ed1e555UL,
+    0xb07e5047UL, 0xd519ecffUL, 0x6c213b62UL, 0x094687daUL, 0xe7e932c8UL,
+    0x828e8e70UL, 0xd49eed28UL, 0xb1f95190UL, 0x5f56e482UL, 0x3a31583aUL,
+    0x83098fa7UL, 0xe66e331fUL, 0x08c1860dUL, 0x6da63ab5UL, 0xa4e140bdUL,
+    0xc186fc05UL, 0x2f294917UL, 0x4a4ef5afUL, 0xf3762232UL, 0x96119e8aUL,
+    0x78be2b98UL, 0x1dd99720UL, 0x4bc9f478UL, 0x2eae48c0UL, 0xc001fdd2UL,
+    0xa566416aUL, 0x1c5e96f7UL, 0x79392a4fUL, 0x97969f5dUL, 0xf2f123e5UL,
+    0x05196b4dUL, 0x607ed7f5UL, 0x8ed162e7UL, 0xebb6de5fUL, 0x528e09c2UL,
+    0x37e9b57aUL, 0xd9460068UL, 0xbc21bcd0UL, 0xea31df88UL, 0x8f566330UL,
+    0x61f9d622UL, 0x049e6a9aUL, 0xbda6bd07UL, 0xd8c101bfUL, 0x366eb4adUL,
+    0x53090815UL, 0x9a4e721dUL, 0xff29cea5UL, 0x11867bb7UL, 0x74e1c70fUL,
+    0xcdd91092UL, 0xa8beac2aUL, 0x46111938UL, 0x2376a580UL, 0x7566c6d8UL,
+    0x10017a60UL, 0xfeaecf72UL, 0x9bc973caUL, 0x22f1a457UL, 0x479618efUL,
+    0xa939adfdUL, 0xcc5e1145UL, 0x06ee4d76UL, 0x6389f1ceUL, 0x8d2644dcUL,
+    0xe841f864UL, 0x51792ff9UL, 0x341e9341UL, 0xdab12653UL, 0xbfd69aebUL,
+    0xe9c6f9b3UL, 0x8ca1450bUL, 0x620ef019UL, 0x07694ca1UL, 0xbe519b3cUL,
+    0xdb362784UL, 0x35999296UL, 0x50fe2e2eUL, 0x99b95426UL, 0xfcdee89eUL,
+    0x12715d8cUL, 0x7716e134UL, 0xce2e36a9UL, 0xab498a11UL, 0x45e63f03UL,
+    0x208183bbUL, 0x7691e0e3UL, 0x13f65c5bUL, 0xfd59e949UL, 0x983e55f1UL,
+    0x2106826cUL, 0x44613ed4UL, 0xaace8bc6UL, 0xcfa9377eUL, 0x38417fd6UL,
+    0x5d26c36eUL, 0xb389767cUL, 0xd6eecac4UL, 0x6fd61d59UL, 0x0ab1a1e1UL,
+    0xe41e14f3UL, 0x8179a84bUL, 0xd769cb13UL, 0xb20e77abUL, 0x5ca1c2b9UL,
+    0x39c67e01UL, 0x80fea99cUL, 0xe5991524UL, 0x0b36a036UL, 0x6e511c8eUL,
+    0xa7166686UL, 0xc271da3eUL, 0x2cde6f2cUL, 0x49b9d394UL, 0xf0810409UL,
+    0x95e6b8b1UL, 0x7b490da3UL, 0x1e2eb11bUL, 0x483ed243UL, 0x2d596efbUL,
+    0xc3f6dbe9UL, 0xa6916751UL, 0x1fa9b0ccUL, 0x7ace0c74UL, 0x9461b966UL,
+    0xf10605deUL
+#endif
+  }
+};
diff --git a/ml/dlib/dlib/external/zlib/deflate.c b/ml/dlib/dlib/external/zlib/deflate.c
new file mode 100644
index 000000000..696957705
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/deflate.c
@@ -0,0 +1,1967 @@
+/* deflate.c -- compress data using the deflation algorithm
+ * Copyright (C) 1995-2013 Jean-loup Gailly and Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/*
+ *  ALGORITHM
+ *
+ *      The "deflation" process depends on being able to identify portions
+ *      of the input text which are identical to earlier input (within a
+ *      sliding window trailing behind the input currently being processed).
+ *
+ *      The most straightforward technique turns out to be the fastest for
+ *      most input files: try all possible matches and select the longest.
+ *      The key feature of this algorithm is that insertions into the string
+ *      dictionary are very simple and thus fast, and deletions are avoided
+ *      completely. Insertions are performed at each input character, whereas
+ *      string matches are performed only when the previous match ends. So it
+ *      is preferable to spend more time in matches to allow very fast string
+ *      insertions and avoid deletions. The matching algorithm for small
+ *      strings is inspired from that of Rabin & Karp. A brute force approach
+ *      is used to find longer strings when a small match has been found.
+ *      A similar algorithm is used in comic (by Jan-Mark Wams) and freeze
+ *      (by Leonid Broukhis).
+ *         A previous version of this file used a more sophisticated algorithm
+ *      (by Fiala and Greene) which is guaranteed to run in linear amortized
+ *      time, but has a larger average cost, uses more memory and is patented.
+ *      However the F&G algorithm may be faster for some highly redundant
+ *      files if the parameter max_chain_length (described below) is too large.
+ *
+ *  ACKNOWLEDGEMENTS
+ *
+ *      The idea of lazy evaluation of matches is due to Jan-Mark Wams, and
+ *      I found it in 'freeze' written by Leonid Broukhis.
+ *      Thanks to many people for bug reports and testing.
+ *
+ *  REFERENCES
+ *
+ *      Deutsch, L.P.,"DEFLATE Compressed Data Format Specification".
+ *      Available in http://tools.ietf.org/html/rfc1951
+ *
+ *      A description of the Rabin and Karp algorithm is given in the book
+ *         "Algorithms" by R. Sedgewick, Addison-Wesley, p252.
+ *
+ *      Fiala,E.R., and Greene,D.H.
+ *         Data Compression with Finite Windows, Comm.ACM, 32,4 (1989) 490-595
+ *
+ */
+
+/* @(#) $Id$ */
+
+#include "deflate.h"
+
+const char deflate_copyright[] =
+   " deflate 1.2.8 Copyright 1995-2013 Jean-loup Gailly and Mark Adler ";
+/*
+  If you use the zlib library in a product, an acknowledgment is welcome
+  in the documentation of your product. If for some reason you cannot
+  include such an acknowledgment, I would appreciate that you keep this
+  copyright string in the executable of your product.
+ */
+
+/* ===========================================================================
+ *  Function prototypes.
+ */
+typedef enum {
+    need_more,      /* block not completed, need more input or more output */
+    block_done,     /* block flush performed */
+    finish_started, /* finish started, need only more output at next deflate */
+    finish_done     /* finish done, accept no more input or output */
+} block_state;
+
+typedef block_state (*compress_func) OF((deflate_state *s, int flush));
+/* Compression function. Returns the block state after the call. */
+
+local void fill_window    OF((deflate_state *s));
+local block_state deflate_stored OF((deflate_state *s, int flush));
+local block_state deflate_fast   OF((deflate_state *s, int flush));
+#ifndef FASTEST
+local block_state deflate_slow   OF((deflate_state *s, int flush));
+#endif
+local block_state deflate_rle    OF((deflate_state *s, int flush));
+local block_state deflate_huff   OF((deflate_state *s, int flush));
+local void lm_init        OF((deflate_state *s));
+local void putShortMSB    OF((deflate_state *s, uInt b));
+local void flush_pending  OF((z_streamp strm));
+local int read_buf        OF((z_streamp strm, Bytef *buf, unsigned size));
+#ifdef ASMV
+      void match_init OF((void)); /* asm code initialization */
+      uInt longest_match  OF((deflate_state *s, IPos cur_match));
+#else
+local uInt longest_match  OF((deflate_state *s, IPos cur_match));
+#endif
+
+#ifdef DEBUG
+local  void check_match OF((deflate_state *s, IPos start, IPos match,
+                            int length));
+#endif
+
+/* ===========================================================================
+ * Local data
+ */
+
+#define NIL 0
+/* Tail of hash chains */
+
+#ifndef TOO_FAR
+#  define TOO_FAR 4096
+#endif
+/* Matches of length 3 are discarded if their distance exceeds TOO_FAR */
+
+/* Values for max_lazy_match, good_match and max_chain_length, depending on
+ * the desired pack level (0..9). The values given below have been tuned to
+ * exclude worst case performance for pathological files. Better values may be
+ * found for specific files.
+ */
+typedef struct config_s {
+   ush good_length; /* reduce lazy search above this match length */
+   ush max_lazy;    /* do not perform lazy search above this match length */
+   ush nice_length; /* quit search above this match length */
+   ush max_chain;
+   compress_func func;
+} config;
+
+#ifdef FASTEST
+local const config configuration_table[2] = {
+/*      good lazy nice chain */
+/* 0 */ {0,    0,  0,    0, deflate_stored},  /* store only */
+/* 1 */ {4,    4,  8,    4, deflate_fast}}; /* max speed, no lazy matches */
+#else
+local const config configuration_table[10] = {
+/*      good lazy nice chain */
+/* 0 */ {0,    0,  0,    0, deflate_stored},  /* store only */
+/* 1 */ {4,    4,  8,    4, deflate_fast}, /* max speed, no lazy matches */
+/* 2 */ {4,    5, 16,    8, deflate_fast},
+/* 3 */ {4,    6, 32,   32, deflate_fast},
+
+/* 4 */ {4,    4, 16,   16, deflate_slow},  /* lazy matches */
+/* 5 */ {8,   16, 32,   32, deflate_slow},
+/* 6 */ {8,   16, 128, 128, deflate_slow},
+/* 7 */ {8,   32, 128, 256, deflate_slow},
+/* 8 */ {32, 128, 258, 1024, deflate_slow},
+/* 9 */ {32, 258, 258, 4096, deflate_slow}}; /* max compression */
+#endif
+
+/* Note: the deflate() code requires max_lazy >= MIN_MATCH and max_chain >= 4
+ * For deflate_fast() (levels <= 3) good is ignored and lazy has a different
+ * meaning.
+ */
+
+#define EQUAL 0
+/* result of memcmp for equal strings */
+
+#ifndef NO_DUMMY_DECL
+struct static_tree_desc_s {int dummy;}; /* for buggy compilers */
+#endif
+
+/* rank Z_BLOCK between Z_NO_FLUSH and Z_PARTIAL_FLUSH */
+#define RANK(f) (((f) << 1) - ((f) > 4 ? 9 : 0))
+
+/* ===========================================================================
+ * Update a hash value with the given input byte
+ * IN  assertion: all calls to to UPDATE_HASH are made with consecutive
+ *    input characters, so that a running hash key can be computed from the
+ *    previous key instead of complete recalculation each time.
+ */
+#define UPDATE_HASH(s,h,c) (h = (((h)<<s->hash_shift) ^ (c)) & s->hash_mask)
+
+
+/* ===========================================================================
+ * Insert string str in the dictionary and set match_head to the previous head
+ * of the hash chain (the most recent string with same hash key). Return
+ * the previous length of the hash chain.
+ * If this file is compiled with -DFASTEST, the compression level is forced
+ * to 1, and no hash chains are maintained.
+ * IN  assertion: all calls to to INSERT_STRING are made with consecutive
+ *    input characters and the first MIN_MATCH bytes of str are valid
+ *    (except for the last MIN_MATCH-1 bytes of the input file).
+ */
+#ifdef FASTEST
+#define INSERT_STRING(s, str, match_head) \
+   (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
+    match_head = s->head[s->ins_h], \
+    s->head[s->ins_h] = (Pos)(str))
+#else
+#define INSERT_STRING(s, str, match_head) \
+   (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
+    match_head = s->prev[(str) & s->w_mask] = s->head[s->ins_h], \
+    s->head[s->ins_h] = (Pos)(str))
+#endif
+
+/* ===========================================================================
+ * Initialize the hash table (avoiding 64K overflow for 16 bit systems).
+ * prev[] will be initialized on the fly.
+ */
+#define CLEAR_HASH(s) \
+    s->head[s->hash_size-1] = NIL; \
+    zmemzero((Bytef *)s->head, (unsigned)(s->hash_size-1)*sizeof(*s->head));
+
+/* ========================================================================= */
+int ZEXPORT deflateInit_(strm, level, version, stream_size)
+    z_streamp strm;
+    int level;
+    const char *version;
+    int stream_size;
+{
+    return deflateInit2_(strm, level, Z_DEFLATED, MAX_WBITS, DEF_MEM_LEVEL,
+                         Z_DEFAULT_STRATEGY, version, stream_size);
+    /* To do: ignore strm->next_in if we use it as window */
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateInit2_(strm, level, method, windowBits, memLevel, strategy,
+                  version, stream_size)
+    z_streamp strm;
+    int  level;
+    int  method;
+    int  windowBits;
+    int  memLevel;
+    int  strategy;
+    const char *version;
+    int stream_size;
+{
+    deflate_state *s;
+    int wrap = 1;
+    static const char my_version[] = ZLIB_VERSION;
+
+    ushf *overlay;
+    /* We overlay pending_buf and d_buf+l_buf. This works since the average
+     * output size for (length,distance) codes is <= 24 bits.
+     */
+
+    if (version == Z_NULL || version[0] != my_version[0] ||
+        stream_size != sizeof(z_stream)) {
+        return Z_VERSION_ERROR;
+    }
+    if (strm == Z_NULL) return Z_STREAM_ERROR;
+
+    strm->msg = Z_NULL;
+    if (strm->zalloc == (alloc_func)0) {
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+        strm->zalloc = zcalloc;
+        strm->opaque = (voidpf)0;
+#endif
+    }
+    if (strm->zfree == (free_func)0)
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+        strm->zfree = zcfree;
+#endif
+
+#ifdef FASTEST
+    if (level != 0) level = 1;
+#else
+    if (level == Z_DEFAULT_COMPRESSION) level = 6;
+#endif
+
+    if (windowBits < 0) { /* suppress zlib wrapper */
+        wrap = 0;
+        windowBits = -windowBits;
+    }
+#ifdef GZIP
+    else if (windowBits > 15) {
+        wrap = 2;       /* write gzip wrapper instead */
+        windowBits -= 16;
+    }
+#endif
+    if (memLevel < 1 || memLevel > MAX_MEM_LEVEL || method != Z_DEFLATED ||
+        windowBits < 8 || windowBits > 15 || level < 0 || level > 9 ||
+        strategy < 0 || strategy > Z_FIXED) {
+        return Z_STREAM_ERROR;
+    }
+    if (windowBits == 8) windowBits = 9;  /* until 256-byte window bug fixed */
+    s = (deflate_state *) ZALLOC(strm, 1, sizeof(deflate_state));
+    if (s == Z_NULL) return Z_MEM_ERROR;
+    strm->state = (struct internal_state FAR *)s;
+    s->strm = strm;
+
+    s->wrap = wrap;
+    s->gzhead = Z_NULL;
+    s->w_bits = windowBits;
+    s->w_size = 1 << s->w_bits;
+    s->w_mask = s->w_size - 1;
+
+    s->hash_bits = memLevel + 7;
+    s->hash_size = 1 << s->hash_bits;
+    s->hash_mask = s->hash_size - 1;
+    s->hash_shift =  ((s->hash_bits+MIN_MATCH-1)/MIN_MATCH);
+
+    s->window = (Bytef *) ZALLOC(strm, s->w_size, 2*sizeof(Byte));
+    s->prev   = (Posf *)  ZALLOC(strm, s->w_size, sizeof(Pos));
+    s->head   = (Posf *)  ZALLOC(strm, s->hash_size, sizeof(Pos));
+
+    s->high_water = 0;      /* nothing written to s->window yet */
+
+    s->lit_bufsize = 1 << (memLevel + 6); /* 16K elements by default */
+
+    overlay = (ushf *) ZALLOC(strm, s->lit_bufsize, sizeof(ush)+2);
+    s->pending_buf = (uchf *) overlay;
+    s->pending_buf_size = (ulg)s->lit_bufsize * (sizeof(ush)+2L);
+
+    if (s->window == Z_NULL || s->prev == Z_NULL || s->head == Z_NULL ||
+        s->pending_buf == Z_NULL) {
+        s->status = FINISH_STATE;
+        strm->msg = ERR_MSG(Z_MEM_ERROR);
+        deflateEnd (strm);
+        return Z_MEM_ERROR;
+    }
+    s->d_buf = overlay + s->lit_bufsize/sizeof(ush);
+    s->l_buf = s->pending_buf + (1+sizeof(ush))*s->lit_bufsize;
+
+    s->level = level;
+    s->strategy = strategy;
+    s->method = (Byte)method;
+
+    return deflateReset(strm);
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateSetDictionary (strm, dictionary, dictLength)
+    z_streamp strm;
+    const Bytef *dictionary;
+    uInt  dictLength;
+{
+    deflate_state *s;
+    uInt str, n;
+    int wrap;
+    unsigned avail;
+    z_const unsigned char *next;
+
+    if (strm == Z_NULL || strm->state == Z_NULL || dictionary == Z_NULL)
+        return Z_STREAM_ERROR;
+    s = strm->state;
+    wrap = s->wrap;
+    if (wrap == 2 || (wrap == 1 && s->status != INIT_STATE) || s->lookahead)
+        return Z_STREAM_ERROR;
+
+    /* when using zlib wrappers, compute Adler-32 for provided dictionary */
+    if (wrap == 1)
+        strm->adler = adler32(strm->adler, dictionary, dictLength);
+    s->wrap = 0;                    /* avoid computing Adler-32 in read_buf */
+
+    /* if dictionary would fill window, just replace the history */
+    if (dictLength >= s->w_size) {
+        if (wrap == 0) {            /* already empty otherwise */
+            CLEAR_HASH(s);
+            s->strstart = 0;
+            s->block_start = 0L;
+            s->insert = 0;
+        }
+        dictionary += dictLength - s->w_size;  /* use the tail */
+        dictLength = s->w_size;
+    }
+
+    /* insert dictionary into window and hash */
+    avail = strm->avail_in;
+    next = strm->next_in;
+    strm->avail_in = dictLength;
+    strm->next_in = (z_const Bytef *)dictionary;
+    fill_window(s);
+    while (s->lookahead >= MIN_MATCH) {
+        str = s->strstart;
+        n = s->lookahead - (MIN_MATCH-1);
+        do {
+            UPDATE_HASH(s, s->ins_h, s->window[str + MIN_MATCH-1]);
+#ifndef FASTEST
+            s->prev[str & s->w_mask] = s->head[s->ins_h];
+#endif
+            s->head[s->ins_h] = (Pos)str;
+            str++;
+        } while (--n);
+        s->strstart = str;
+        s->lookahead = MIN_MATCH-1;
+        fill_window(s);
+    }
+    s->strstart += s->lookahead;
+    s->block_start = (long)s->strstart;
+    s->insert = s->lookahead;
+    s->lookahead = 0;
+    s->match_length = s->prev_length = MIN_MATCH-1;
+    s->match_available = 0;
+    strm->next_in = next;
+    strm->avail_in = avail;
+    s->wrap = wrap;
+    return Z_OK;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateResetKeep (strm)
+    z_streamp strm;
+{
+    deflate_state *s;
+
+    if (strm == Z_NULL || strm->state == Z_NULL ||
+        strm->zalloc == (alloc_func)0 || strm->zfree == (free_func)0) {
+        return Z_STREAM_ERROR;
+    }
+
+    strm->total_in = strm->total_out = 0;
+    strm->msg = Z_NULL; /* use zfree if we ever allocate msg dynamically */
+    strm->data_type = Z_UNKNOWN;
+
+    s = (deflate_state *)strm->state;
+    s->pending = 0;
+    s->pending_out = s->pending_buf;
+
+    if (s->wrap < 0) {
+        s->wrap = -s->wrap; /* was made negative by deflate(..., Z_FINISH); */
+    }
+    s->status = s->wrap ? INIT_STATE : BUSY_STATE;
+    strm->adler =
+#ifdef GZIP
+        s->wrap == 2 ? crc32(0L, Z_NULL, 0) :
+#endif
+        adler32(0L, Z_NULL, 0);
+    s->last_flush = Z_NO_FLUSH;
+
+    _tr_init(s);
+
+    return Z_OK;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateReset (strm)
+    z_streamp strm;
+{
+    int ret;
+
+    ret = deflateResetKeep(strm);
+    if (ret == Z_OK)
+        lm_init(strm->state);
+    return ret;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateSetHeader (strm, head)
+    z_streamp strm;
+    gz_headerp head;
+{
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    if (strm->state->wrap != 2) return Z_STREAM_ERROR;
+    strm->state->gzhead = head;
+    return Z_OK;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflatePending (strm, pending, bits)
+    unsigned *pending;
+    int *bits;
+    z_streamp strm;
+{
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    if (pending != Z_NULL)
+        *pending = strm->state->pending;
+    if (bits != Z_NULL)
+        *bits = strm->state->bi_valid;
+    return Z_OK;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflatePrime (strm, bits, value)
+    z_streamp strm;
+    int bits;
+    int value;
+{
+    deflate_state *s;
+    int put;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    s = strm->state;
+    if ((Bytef *)(s->d_buf) < s->pending_out + ((Buf_size + 7) >> 3))
+        return Z_BUF_ERROR;
+    do {
+        put = Buf_size - s->bi_valid;
+        if (put > bits)
+            put = bits;
+        s->bi_buf |= (ush)((value & ((1 << put) - 1)) << s->bi_valid);
+        s->bi_valid += put;
+        _tr_flush_bits(s);
+        value >>= put;
+        bits -= put;
+    } while (bits);
+    return Z_OK;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateParams(strm, level, strategy)
+    z_streamp strm;
+    int level;
+    int strategy;
+{
+    deflate_state *s;
+    compress_func func;
+    int err = Z_OK;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    s = strm->state;
+
+#ifdef FASTEST
+    if (level != 0) level = 1;
+#else
+    if (level == Z_DEFAULT_COMPRESSION) level = 6;
+#endif
+    if (level < 0 || level > 9 || strategy < 0 || strategy > Z_FIXED) {
+        return Z_STREAM_ERROR;
+    }
+    func = configuration_table[s->level].func;
+
+    if ((strategy != s->strategy || func != configuration_table[level].func) &&
+        strm->total_in != 0) {
+        /* Flush the last buffer: */
+        err = deflate(strm, Z_BLOCK);
+        if (err == Z_BUF_ERROR && s->pending == 0)
+            err = Z_OK;
+    }
+    if (s->level != level) {
+        s->level = level;
+        s->max_lazy_match   = configuration_table[level].max_lazy;
+        s->good_match       = configuration_table[level].good_length;
+        s->nice_match       = configuration_table[level].nice_length;
+        s->max_chain_length = configuration_table[level].max_chain;
+    }
+    s->strategy = strategy;
+    return err;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateTune(strm, good_length, max_lazy, nice_length, max_chain)
+    z_streamp strm;
+    int good_length;
+    int max_lazy;
+    int nice_length;
+    int max_chain;
+{
+    deflate_state *s;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    s = strm->state;
+    s->good_match = good_length;
+    s->max_lazy_match = max_lazy;
+    s->nice_match = nice_length;
+    s->max_chain_length = max_chain;
+    return Z_OK;
+}
+
+/* =========================================================================
+ * For the default windowBits of 15 and memLevel of 8, this function returns
+ * a close to exact, as well as small, upper bound on the compressed size.
+ * They are coded as constants here for a reason--if the #define's are
+ * changed, then this function needs to be changed as well.  The return
+ * value for 15 and 8 only works for those exact settings.
+ *
+ * For any setting other than those defaults for windowBits and memLevel,
+ * the value returned is a conservative worst case for the maximum expansion
+ * resulting from using fixed blocks instead of stored blocks, which deflate
+ * can emit on compressed data for some combinations of the parameters.
+ *
+ * This function could be more sophisticated to provide closer upper bounds for
+ * every combination of windowBits and memLevel.  But even the conservative
+ * upper bound of about 14% expansion does not seem onerous for output buffer
+ * allocation.
+ */
+uLong ZEXPORT deflateBound(strm, sourceLen)
+    z_streamp strm;
+    uLong sourceLen;
+{
+    deflate_state *s;
+    uLong complen, wraplen;
+    Bytef *str;
+
+    /* conservative upper bound for compressed data */
+    complen = sourceLen +
+              ((sourceLen + 7) >> 3) + ((sourceLen + 63) >> 6) + 5;
+
+    /* if can't get parameters, return conservative bound plus zlib wrapper */
+    if (strm == Z_NULL || strm->state == Z_NULL)
+        return complen + 6;
+
+    /* compute wrapper length */
+    s = strm->state;
+    switch (s->wrap) {
+    case 0:                                 /* raw deflate */
+        wraplen = 0;
+        break;
+    case 1:                                 /* zlib wrapper */
+        wraplen = 6 + (s->strstart ? 4 : 0);
+        break;
+    case 2:                                 /* gzip wrapper */
+        wraplen = 18;
+        if (s->gzhead != Z_NULL) {          /* user-supplied gzip header */
+            if (s->gzhead->extra != Z_NULL)
+                wraplen += 2 + s->gzhead->extra_len;
+            str = s->gzhead->name;
+            if (str != Z_NULL)
+                do {
+                    wraplen++;
+                } while (*str++);
+            str = s->gzhead->comment;
+            if (str != Z_NULL)
+                do {
+                    wraplen++;
+                } while (*str++);
+            if (s->gzhead->hcrc)
+                wraplen += 2;
+        }
+        break;
+    default:                                /* for compiler happiness */
+        wraplen = 6;
+    }
+
+    /* if not default parameters, return conservative bound */
+    if (s->w_bits != 15 || s->hash_bits != 8 + 7)
+        return complen + wraplen;
+
+    /* default settings: return tight bound for that case */
+    return sourceLen + (sourceLen >> 12) + (sourceLen >> 14) +
+           (sourceLen >> 25) + 13 - 6 + wraplen;
+}
+
+/* =========================================================================
+ * Put a short in the pending buffer. The 16-bit value is put in MSB order.
+ * IN assertion: the stream state is correct and there is enough room in
+ * pending_buf.
+ */
+local void putShortMSB (s, b)
+    deflate_state *s;
+    uInt b;
+{
+    put_byte(s, (Byte)(b >> 8));
+    put_byte(s, (Byte)(b & 0xff));
+}
+
+/* =========================================================================
+ * Flush as much pending output as possible. All deflate() output goes
+ * through this function so some applications may wish to modify it
+ * to avoid allocating a large strm->next_out buffer and copying into it.
+ * (See also read_buf()).
+ */
+local void flush_pending(strm)
+    z_streamp strm;
+{
+    unsigned len;
+    deflate_state *s = strm->state;
+
+    _tr_flush_bits(s);
+    len = s->pending;
+    if (len > strm->avail_out) len = strm->avail_out;
+    if (len == 0) return;
+
+    zmemcpy(strm->next_out, s->pending_out, len);
+    strm->next_out  += len;
+    s->pending_out  += len;
+    strm->total_out += len;
+    strm->avail_out  -= len;
+    s->pending -= len;
+    if (s->pending == 0) {
+        s->pending_out = s->pending_buf;
+    }
+}
+
+/* ========================================================================= */
+int ZEXPORT deflate (strm, flush)
+    z_streamp strm;
+    int flush;
+{
+    int old_flush; /* value of flush param for previous deflate call */
+    deflate_state *s;
+
+    if (strm == Z_NULL || strm->state == Z_NULL ||
+        flush > Z_BLOCK || flush < 0) {
+        return Z_STREAM_ERROR;
+    }
+    s = strm->state;
+
+    if (strm->next_out == Z_NULL ||
+        (strm->next_in == Z_NULL && strm->avail_in != 0) ||
+        (s->status == FINISH_STATE && flush != Z_FINISH)) {
+        ERR_RETURN(strm, Z_STREAM_ERROR);
+    }
+    if (strm->avail_out == 0) ERR_RETURN(strm, Z_BUF_ERROR);
+
+    s->strm = strm; /* just in case */
+    old_flush = s->last_flush;
+    s->last_flush = flush;
+
+    /* Write the header */
+    if (s->status == INIT_STATE) {
+#ifdef GZIP
+        if (s->wrap == 2) {
+            strm->adler = crc32(0L, Z_NULL, 0);
+            put_byte(s, 31);
+            put_byte(s, 139);
+            put_byte(s, 8);
+            if (s->gzhead == Z_NULL) {
+                put_byte(s, 0);
+                put_byte(s, 0);
+                put_byte(s, 0);
+                put_byte(s, 0);
+                put_byte(s, 0);
+                put_byte(s, s->level == 9 ? 2 :
+                            (s->strategy >= Z_HUFFMAN_ONLY || s->level < 2 ?
+                             4 : 0));
+                put_byte(s, OS_CODE);
+                s->status = BUSY_STATE;
+            }
+            else {
+                put_byte(s, (s->gzhead->text ? 1 : 0) +
+                            (s->gzhead->hcrc ? 2 : 0) +
+                            (s->gzhead->extra == Z_NULL ? 0 : 4) +
+                            (s->gzhead->name == Z_NULL ? 0 : 8) +
+                            (s->gzhead->comment == Z_NULL ? 0 : 16)
+                        );
+                put_byte(s, (Byte)(s->gzhead->time & 0xff));
+                put_byte(s, (Byte)((s->gzhead->time >> 8) & 0xff));
+                put_byte(s, (Byte)((s->gzhead->time >> 16) & 0xff));
+                put_byte(s, (Byte)((s->gzhead->time >> 24) & 0xff));
+                put_byte(s, s->level == 9 ? 2 :
+                            (s->strategy >= Z_HUFFMAN_ONLY || s->level < 2 ?
+                             4 : 0));
+                put_byte(s, s->gzhead->os & 0xff);
+                if (s->gzhead->extra != Z_NULL) {
+                    put_byte(s, s->gzhead->extra_len & 0xff);
+                    put_byte(s, (s->gzhead->extra_len >> 8) & 0xff);
+                }
+                if (s->gzhead->hcrc)
+                    strm->adler = crc32(strm->adler, s->pending_buf,
+                                        s->pending);
+                s->gzindex = 0;
+                s->status = EXTRA_STATE;
+            }
+        }
+        else
+#endif
+        {
+            uInt header = (Z_DEFLATED + ((s->w_bits-8)<<4)) << 8;
+            uInt level_flags;
+
+            if (s->strategy >= Z_HUFFMAN_ONLY || s->level < 2)
+                level_flags = 0;
+            else if (s->level < 6)
+                level_flags = 1;
+            else if (s->level == 6)
+                level_flags = 2;
+            else
+                level_flags = 3;
+            header |= (level_flags << 6);
+            if (s->strstart != 0) header |= PRESET_DICT;
+            header += 31 - (header % 31);
+
+            s->status = BUSY_STATE;
+            putShortMSB(s, header);
+
+            /* Save the adler32 of the preset dictionary: */
+            if (s->strstart != 0) {
+                putShortMSB(s, (uInt)(strm->adler >> 16));
+                putShortMSB(s, (uInt)(strm->adler & 0xffff));
+            }
+            strm->adler = adler32(0L, Z_NULL, 0);
+        }
+    }
+#ifdef GZIP
+    if (s->status == EXTRA_STATE) {
+        if (s->gzhead->extra != Z_NULL) {
+            uInt beg = s->pending;  /* start of bytes to update crc */
+
+            while (s->gzindex < (s->gzhead->extra_len & 0xffff)) {
+                if (s->pending == s->pending_buf_size) {
+                    if (s->gzhead->hcrc && s->pending > beg)
+                        strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                            s->pending - beg);
+                    flush_pending(strm);
+                    beg = s->pending;
+                    if (s->pending == s->pending_buf_size)
+                        break;
+                }
+                put_byte(s, s->gzhead->extra[s->gzindex]);
+                s->gzindex++;
+            }
+            if (s->gzhead->hcrc && s->pending > beg)
+                strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                    s->pending - beg);
+            if (s->gzindex == s->gzhead->extra_len) {
+                s->gzindex = 0;
+                s->status = NAME_STATE;
+            }
+        }
+        else
+            s->status = NAME_STATE;
+    }
+    if (s->status == NAME_STATE) {
+        if (s->gzhead->name != Z_NULL) {
+            uInt beg = s->pending;  /* start of bytes to update crc */
+            int val;
+
+            do {
+                if (s->pending == s->pending_buf_size) {
+                    if (s->gzhead->hcrc && s->pending > beg)
+                        strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                            s->pending - beg);
+                    flush_pending(strm);
+                    beg = s->pending;
+                    if (s->pending == s->pending_buf_size) {
+                        val = 1;
+                        break;
+                    }
+                }
+                val = s->gzhead->name[s->gzindex++];
+                put_byte(s, val);
+            } while (val != 0);
+            if (s->gzhead->hcrc && s->pending > beg)
+                strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                    s->pending - beg);
+            if (val == 0) {
+                s->gzindex = 0;
+                s->status = COMMENT_STATE;
+            }
+        }
+        else
+            s->status = COMMENT_STATE;
+    }
+    if (s->status == COMMENT_STATE) {
+        if (s->gzhead->comment != Z_NULL) {
+            uInt beg = s->pending;  /* start of bytes to update crc */
+            int val;
+
+            do {
+                if (s->pending == s->pending_buf_size) {
+                    if (s->gzhead->hcrc && s->pending > beg)
+                        strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                            s->pending - beg);
+                    flush_pending(strm);
+                    beg = s->pending;
+                    if (s->pending == s->pending_buf_size) {
+                        val = 1;
+                        break;
+                    }
+                }
+                val = s->gzhead->comment[s->gzindex++];
+                put_byte(s, val);
+            } while (val != 0);
+            if (s->gzhead->hcrc && s->pending > beg)
+                strm->adler = crc32(strm->adler, s->pending_buf + beg,
+                                    s->pending - beg);
+            if (val == 0)
+                s->status = HCRC_STATE;
+        }
+        else
+            s->status = HCRC_STATE;
+    }
+    if (s->status == HCRC_STATE) {
+        if (s->gzhead->hcrc) {
+            if (s->pending + 2 > s->pending_buf_size)
+                flush_pending(strm);
+            if (s->pending + 2 <= s->pending_buf_size) {
+                put_byte(s, (Byte)(strm->adler & 0xff));
+                put_byte(s, (Byte)((strm->adler >> 8) & 0xff));
+                strm->adler = crc32(0L, Z_NULL, 0);
+                s->status = BUSY_STATE;
+            }
+        }
+        else
+            s->status = BUSY_STATE;
+    }
+#endif
+
+    /* Flush as much pending output as possible */
+    if (s->pending != 0) {
+        flush_pending(strm);
+        if (strm->avail_out == 0) {
+            /* Since avail_out is 0, deflate will be called again with
+             * more output space, but possibly with both pending and
+             * avail_in equal to zero. There won't be anything to do,
+             * but this is not an error situation so make sure we
+             * return OK instead of BUF_ERROR at next call of deflate:
+             */
+            s->last_flush = -1;
+            return Z_OK;
+        }
+
+    /* Make sure there is something to do and avoid duplicate consecutive
+     * flushes. For repeated and useless calls with Z_FINISH, we keep
+     * returning Z_STREAM_END instead of Z_BUF_ERROR.
+     */
+    } else if (strm->avail_in == 0 && RANK(flush) <= RANK(old_flush) &&
+               flush != Z_FINISH) {
+        ERR_RETURN(strm, Z_BUF_ERROR);
+    }
+
+    /* User must not provide more input after the first FINISH: */
+    if (s->status == FINISH_STATE && strm->avail_in != 0) {
+        ERR_RETURN(strm, Z_BUF_ERROR);
+    }
+
+    /* Start a new block or continue the current one.
+     */
+    if (strm->avail_in != 0 || s->lookahead != 0 ||
+        (flush != Z_NO_FLUSH && s->status != FINISH_STATE)) {
+        block_state bstate;
+
+        bstate = s->strategy == Z_HUFFMAN_ONLY ? deflate_huff(s, flush) :
+                    (s->strategy == Z_RLE ? deflate_rle(s, flush) :
+                        (*(configuration_table[s->level].func))(s, flush));
+
+        if (bstate == finish_started || bstate == finish_done) {
+            s->status = FINISH_STATE;
+        }
+        if (bstate == need_more || bstate == finish_started) {
+            if (strm->avail_out == 0) {
+                s->last_flush = -1; /* avoid BUF_ERROR next call, see above */
+            }
+            return Z_OK;
+            /* If flush != Z_NO_FLUSH && avail_out == 0, the next call
+             * of deflate should use the same flush parameter to make sure
+             * that the flush is complete. So we don't have to output an
+             * empty block here, this will be done at next call. This also
+             * ensures that for a very small output buffer, we emit at most
+             * one empty block.
+             */
+        }
+        if (bstate == block_done) {
+            if (flush == Z_PARTIAL_FLUSH) {
+                _tr_align(s);
+            } else if (flush != Z_BLOCK) { /* FULL_FLUSH or SYNC_FLUSH */
+                _tr_stored_block(s, (char*)0, 0L, 0);
+                /* For a full flush, this empty block will be recognized
+                 * as a special marker by inflate_sync().
+                 */
+                if (flush == Z_FULL_FLUSH) {
+                    CLEAR_HASH(s);             /* forget history */
+                    if (s->lookahead == 0) {
+                        s->strstart = 0;
+                        s->block_start = 0L;
+                        s->insert = 0;
+                    }
+                }
+            }
+            flush_pending(strm);
+            if (strm->avail_out == 0) {
+              s->last_flush = -1; /* avoid BUF_ERROR at next call, see above */
+              return Z_OK;
+            }
+        }
+    }
+    Assert(strm->avail_out > 0, "bug2");
+
+    if (flush != Z_FINISH) return Z_OK;
+    if (s->wrap <= 0) return Z_STREAM_END;
+
+    /* Write the trailer */
+#ifdef GZIP
+    if (s->wrap == 2) {
+        put_byte(s, (Byte)(strm->adler & 0xff));
+        put_byte(s, (Byte)((strm->adler >> 8) & 0xff));
+        put_byte(s, (Byte)((strm->adler >> 16) & 0xff));
+        put_byte(s, (Byte)((strm->adler >> 24) & 0xff));
+        put_byte(s, (Byte)(strm->total_in & 0xff));
+        put_byte(s, (Byte)((strm->total_in >> 8) & 0xff));
+        put_byte(s, (Byte)((strm->total_in >> 16) & 0xff));
+        put_byte(s, (Byte)((strm->total_in >> 24) & 0xff));
+    }
+    else
+#endif
+    {
+        putShortMSB(s, (uInt)(strm->adler >> 16));
+        putShortMSB(s, (uInt)(strm->adler & 0xffff));
+    }
+    flush_pending(strm);
+    /* If avail_out is zero, the application will call deflate again
+     * to flush the rest.
+     */
+    if (s->wrap > 0) s->wrap = -s->wrap; /* write the trailer only once! */
+    return s->pending != 0 ? Z_OK : Z_STREAM_END;
+}
+
+/* ========================================================================= */
+int ZEXPORT deflateEnd (strm)
+    z_streamp strm;
+{
+    int status;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+
+    status = strm->state->status;
+    if (status != INIT_STATE &&
+        status != EXTRA_STATE &&
+        status != NAME_STATE &&
+        status != COMMENT_STATE &&
+        status != HCRC_STATE &&
+        status != BUSY_STATE &&
+        status != FINISH_STATE) {
+      return Z_STREAM_ERROR;
+    }
+
+    /* Deallocate in reverse order of allocations: */
+    TRY_FREE(strm, strm->state->pending_buf);
+    TRY_FREE(strm, strm->state->head);
+    TRY_FREE(strm, strm->state->prev);
+    TRY_FREE(strm, strm->state->window);
+
+    ZFREE(strm, strm->state);
+    strm->state = Z_NULL;
+
+    return status == BUSY_STATE ? Z_DATA_ERROR : Z_OK;
+}
+
+/* =========================================================================
+ * Copy the source state to the destination state.
+ * To simplify the source, this is not supported for 16-bit MSDOS (which
+ * doesn't have enough memory anyway to duplicate compression states).
+ */
+int ZEXPORT deflateCopy (dest, source)
+    z_streamp dest;
+    z_streamp source;
+{
+#ifdef MAXSEG_64K
+    return Z_STREAM_ERROR;
+#else
+    deflate_state *ds;
+    deflate_state *ss;
+    ushf *overlay;
+
+
+    if (source == Z_NULL || dest == Z_NULL || source->state == Z_NULL) {
+        return Z_STREAM_ERROR;
+    }
+
+    ss = source->state;
+
+    zmemcpy((voidpf)dest, (voidpf)source, sizeof(z_stream));
+
+    ds = (deflate_state *) ZALLOC(dest, 1, sizeof(deflate_state));
+    if (ds == Z_NULL) return Z_MEM_ERROR;
+    dest->state = (struct internal_state FAR *) ds;
+    zmemcpy((voidpf)ds, (voidpf)ss, sizeof(deflate_state));
+    ds->strm = dest;
+
+    ds->window = (Bytef *) ZALLOC(dest, ds->w_size, 2*sizeof(Byte));
+    ds->prev   = (Posf *)  ZALLOC(dest, ds->w_size, sizeof(Pos));
+    ds->head   = (Posf *)  ZALLOC(dest, ds->hash_size, sizeof(Pos));
+    overlay = (ushf *) ZALLOC(dest, ds->lit_bufsize, sizeof(ush)+2);
+    ds->pending_buf = (uchf *) overlay;
+
+    if (ds->window == Z_NULL || ds->prev == Z_NULL || ds->head == Z_NULL ||
+        ds->pending_buf == Z_NULL) {
+        deflateEnd (dest);
+        return Z_MEM_ERROR;
+    }
+    /* following zmemcpy do not work for 16-bit MSDOS */
+    zmemcpy(ds->window, ss->window, ds->w_size * 2 * sizeof(Byte));
+    zmemcpy((voidpf)ds->prev, (voidpf)ss->prev, ds->w_size * sizeof(Pos));
+    zmemcpy((voidpf)ds->head, (voidpf)ss->head, ds->hash_size * sizeof(Pos));
+    zmemcpy(ds->pending_buf, ss->pending_buf, (uInt)ds->pending_buf_size);
+
+    ds->pending_out = ds->pending_buf + (ss->pending_out - ss->pending_buf);
+    ds->d_buf = overlay + ds->lit_bufsize/sizeof(ush);
+    ds->l_buf = ds->pending_buf + (1+sizeof(ush))*ds->lit_bufsize;
+
+    ds->l_desc.dyn_tree = ds->dyn_ltree;
+    ds->d_desc.dyn_tree = ds->dyn_dtree;
+    ds->bl_desc.dyn_tree = ds->bl_tree;
+
+    return Z_OK;
+#endif /* MAXSEG_64K */
+}
+
+/* ===========================================================================
+ * Read a new buffer from the current input stream, update the adler32
+ * and total number of bytes read.  All deflate() input goes through
+ * this function so some applications may wish to modify it to avoid
+ * allocating a large strm->next_in buffer and copying from it.
+ * (See also flush_pending()).
+ */
+local int read_buf(strm, buf, size)
+    z_streamp strm;
+    Bytef *buf;
+    unsigned size;
+{
+    unsigned len = strm->avail_in;
+
+    if (len > size) len = size;
+    if (len == 0) return 0;
+
+    strm->avail_in  -= len;
+
+    zmemcpy(buf, strm->next_in, len);
+    if (strm->state->wrap == 1) {
+        strm->adler = adler32(strm->adler, buf, len);
+    }
+#ifdef GZIP
+    else if (strm->state->wrap == 2) {
+        strm->adler = crc32(strm->adler, buf, len);
+    }
+#endif
+    strm->next_in  += len;
+    strm->total_in += len;
+
+    return (int)len;
+}
+
+/* ===========================================================================
+ * Initialize the "longest match" routines for a new zlib stream
+ */
+local void lm_init (s)
+    deflate_state *s;
+{
+    s->window_size = (ulg)2L*s->w_size;
+
+    CLEAR_HASH(s);
+
+    /* Set the default configuration parameters:
+     */
+    s->max_lazy_match   = configuration_table[s->level].max_lazy;
+    s->good_match       = configuration_table[s->level].good_length;
+    s->nice_match       = configuration_table[s->level].nice_length;
+    s->max_chain_length = configuration_table[s->level].max_chain;
+
+    s->strstart = 0;
+    s->block_start = 0L;
+    s->lookahead = 0;
+    s->insert = 0;
+    s->match_length = s->prev_length = MIN_MATCH-1;
+    s->match_available = 0;
+    s->ins_h = 0;
+#ifndef FASTEST
+#ifdef ASMV
+    match_init(); /* initialize the asm code */
+#endif
+#endif
+}
+
+#ifndef FASTEST
+/* ===========================================================================
+ * Set match_start to the longest match starting at the given string and
+ * return its length. Matches shorter or equal to prev_length are discarded,
+ * in which case the result is equal to prev_length and match_start is
+ * garbage.
+ * IN assertions: cur_match is the head of the hash chain for the current
+ *   string (strstart) and its distance is <= MAX_DIST, and prev_length >= 1
+ * OUT assertion: the match length is not greater than s->lookahead.
+ */
+#ifndef ASMV
+/* For 80x86 and 680x0, an optimized version will be provided in match.asm or
+ * match.S. The code will be functionally equivalent.
+ */
+local uInt longest_match(s, cur_match)
+    deflate_state *s;
+    IPos cur_match;                             /* current match */
+{
+    unsigned chain_length = s->max_chain_length;/* max hash chain length */
+    register Bytef *scan = s->window + s->strstart; /* current string */
+    register Bytef *match;                       /* matched string */
+    register int len;                           /* length of current match */
+    int best_len = s->prev_length;              /* best match length so far */
+    int nice_match = s->nice_match;             /* stop if match long enough */
+    IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
+        s->strstart - (IPos)MAX_DIST(s) : NIL;
+    /* Stop when cur_match becomes <= limit. To simplify the code,
+     * we prevent matches with the string of window index 0.
+     */
+    Posf *prev = s->prev;
+    uInt wmask = s->w_mask;
+
+#ifdef UNALIGNED_OK
+    /* Compare two bytes at a time. Note: this is not always beneficial.
+     * Try with and without -DUNALIGNED_OK to check.
+     */
+    register Bytef *strend = s->window + s->strstart + MAX_MATCH - 1;
+    register ush scan_start = *(ushf*)scan;
+    register ush scan_end   = *(ushf*)(scan+best_len-1);
+#else
+    register Bytef *strend = s->window + s->strstart + MAX_MATCH;
+    register Byte scan_end1  = scan[best_len-1];
+    register Byte scan_end   = scan[best_len];
+#endif
+
+    /* The code is optimized for HASH_BITS >= 8 and MAX_MATCH-2 multiple of 16.
+     * It is easy to get rid of this optimization if necessary.
+     */
+    Assert(s->hash_bits >= 8 && MAX_MATCH == 258, "Code too clever");
+
+    /* Do not waste too much time if we already have a good match: */
+    if (s->prev_length >= s->good_match) {
+        chain_length >>= 2;
+    }
+    /* Do not look for matches beyond the end of the input. This is necessary
+     * to make deflate deterministic.
+     */
+    if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;
+
+    Assert((ulg)s->strstart <= s->window_size-MIN_LOOKAHEAD, "need lookahead");
+
+    do {
+        Assert(cur_match < s->strstart, "no future");
+        match = s->window + cur_match;
+
+        /* Skip to next match if the match length cannot increase
+         * or if the match length is less than 2.  Note that the checks below
+         * for insufficient lookahead only occur occasionally for performance
+         * reasons.  Therefore uninitialized memory will be accessed, and
+         * conditional jumps will be made that depend on those values.
+         * However the length of the match is limited to the lookahead, so
+         * the output of deflate is not affected by the uninitialized values.
+         */
+#if (defined(UNALIGNED_OK) && MAX_MATCH == 258)
+        /* This code assumes sizeof(unsigned short) == 2. Do not use
+         * UNALIGNED_OK if your compiler uses a different size.
+         */
+        if (*(ushf*)(match+best_len-1) != scan_end ||
+            *(ushf*)match != scan_start) continue;
+
+        /* It is not necessary to compare scan[2] and match[2] since they are
+         * always equal when the other bytes match, given that the hash keys
+         * are equal and that HASH_BITS >= 8. Compare 2 bytes at a time at
+         * strstart+3, +5, ... up to strstart+257. We check for insufficient
+         * lookahead only every 4th comparison; the 128th check will be made
+         * at strstart+257. If MAX_MATCH-2 is not a multiple of 8, it is
+         * necessary to put more guard bytes at the end of the window, or
+         * to check more often for insufficient lookahead.
+         */
+        Assert(scan[2] == match[2], "scan[2]?");
+        scan++, match++;
+        do {
+        } while (*(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
+                 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
+                 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
+                 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
+                 scan < strend);
+        /* The funny "do {}" generates better code on most compilers */
+
+        /* Here, scan <= window+strstart+257 */
+        Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
+        if (*scan == *match) scan++;
+
+        len = (MAX_MATCH - 1) - (int)(strend-scan);
+        scan = strend - (MAX_MATCH-1);
+
+#else /* UNALIGNED_OK */
+
+        if (match[best_len]   != scan_end  ||
+            match[best_len-1] != scan_end1 ||
+            *match            != *scan     ||
+            *++match          != scan[1])      continue;
+
+        /* The check at best_len-1 can be removed because it will be made
+         * again later. (This heuristic is not always a win.)
+         * It is not necessary to compare scan[2] and match[2] since they
+         * are always equal when the other bytes match, given that
+         * the hash keys are equal and that HASH_BITS >= 8.
+         */
+        scan += 2, match++;
+        Assert(*scan == *match, "match[2]?");
+
+        /* We check for insufficient lookahead only every 8th comparison;
+         * the 256th check will be made at strstart+258.
+         */
+        do {
+        } while (*++scan == *++match && *++scan == *++match &&
+                 *++scan == *++match && *++scan == *++match &&
+                 *++scan == *++match && *++scan == *++match &&
+                 *++scan == *++match && *++scan == *++match &&
+                 scan < strend);
+
+        Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
+
+        len = MAX_MATCH - (int)(strend - scan);
+        scan = strend - MAX_MATCH;
+
+#endif /* UNALIGNED_OK */
+
+        if (len > best_len) {
+            s->match_start = cur_match;
+            best_len = len;
+            if (len >= nice_match) break;
+#ifdef UNALIGNED_OK
+            scan_end = *(ushf*)(scan+best_len-1);
+#else
+            scan_end1  = scan[best_len-1];
+            scan_end   = scan[best_len];
+#endif
+        }
+    } while ((cur_match = prev[cur_match & wmask]) > limit
+             && --chain_length != 0);
+
+    if ((uInt)best_len <= s->lookahead) return (uInt)best_len;
+    return s->lookahead;
+}
+#endif /* ASMV */
+
+#else /* FASTEST */
+
+/* ---------------------------------------------------------------------------
+ * Optimized version for FASTEST only
+ */
+local uInt longest_match(s, cur_match)
+    deflate_state *s;
+    IPos cur_match;                             /* current match */
+{
+    register Bytef *scan = s->window + s->strstart; /* current string */
+    register Bytef *match;                       /* matched string */
+    register int len;                           /* length of current match */
+    register Bytef *strend = s->window + s->strstart + MAX_MATCH;
+
+    /* The code is optimized for HASH_BITS >= 8 and MAX_MATCH-2 multiple of 16.
+     * It is easy to get rid of this optimization if necessary.
+     */
+    Assert(s->hash_bits >= 8 && MAX_MATCH == 258, "Code too clever");
+
+    Assert((ulg)s->strstart <= s->window_size-MIN_LOOKAHEAD, "need lookahead");
+
+    Assert(cur_match < s->strstart, "no future");
+
+    match = s->window + cur_match;
+
+    /* Return failure if the match length is less than 2:
+     */
+    if (match[0] != scan[0] || match[1] != scan[1]) return MIN_MATCH-1;
+
+    /* The check at best_len-1 can be removed because it will be made
+     * again later. (This heuristic is not always a win.)
+     * It is not necessary to compare scan[2] and match[2] since they
+     * are always equal when the other bytes match, given that
+     * the hash keys are equal and that HASH_BITS >= 8.
+     */
+    scan += 2, match += 2;
+    Assert(*scan == *match, "match[2]?");
+
+    /* We check for insufficient lookahead only every 8th comparison;
+     * the 256th check will be made at strstart+258.
+     */
+    do {
+    } while (*++scan == *++match && *++scan == *++match &&
+             *++scan == *++match && *++scan == *++match &&
+             *++scan == *++match && *++scan == *++match &&
+             *++scan == *++match && *++scan == *++match &&
+             scan < strend);
+
+    Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
+
+    len = MAX_MATCH - (int)(strend - scan);
+
+    if (len < MIN_MATCH) return MIN_MATCH - 1;
+
+    s->match_start = cur_match;
+    return (uInt)len <= s->lookahead ? (uInt)len : s->lookahead;
+}
+
+#endif /* FASTEST */
+
+#ifdef DEBUG
+/* ===========================================================================
+ * Check that the match at match_start is indeed a match.
+ */
+local void check_match(s, start, match, length)
+    deflate_state *s;
+    IPos start, match;
+    int length;
+{
+    /* check that the match is indeed a match */
+    if (zmemcmp(s->window + match,
+                s->window + start, length) != EQUAL) {
+        fprintf(stderr, " start %u, match %u, length %d\n",
+                start, match, length);
+        do {
+            fprintf(stderr, "%c%c", s->window[match++], s->window[start++]);
+        } while (--length != 0);
+        z_error("invalid match");
+    }
+    if (z_verbose > 1) {
+        fprintf(stderr,"\\[%d,%d]", start-match, length);
+        do { putc(s->window[start++], stderr); } while (--length != 0);
+    }
+}
+#else
+#  define check_match(s, start, match, length)
+#endif /* DEBUG */
+
+/* ===========================================================================
+ * Fill the window when the lookahead becomes insufficient.
+ * Updates strstart and lookahead.
+ *
+ * IN assertion: lookahead < MIN_LOOKAHEAD
+ * OUT assertions: strstart <= window_size-MIN_LOOKAHEAD
+ *    At least one byte has been read, or avail_in == 0; reads are
+ *    performed for at least two bytes (required for the zip translate_eol
+ *    option -- not supported here).
+ */
+local void fill_window(s)
+    deflate_state *s;
+{
+    register unsigned n, m;
+    register Posf *p;
+    unsigned more;    /* Amount of free space at the end of the window. */
+    uInt wsize = s->w_size;
+
+    Assert(s->lookahead < MIN_LOOKAHEAD, "already enough lookahead");
+
+    do {
+        more = (unsigned)(s->window_size -(ulg)s->lookahead -(ulg)s->strstart);
+
+        /* Deal with !@#$% 64K limit: */
+        if (sizeof(int) <= 2) {
+            if (more == 0 && s->strstart == 0 && s->lookahead == 0) {
+                more = wsize;
+
+            } else if (more == (unsigned)(-1)) {
+                /* Very unlikely, but possible on 16 bit machine if
+                 * strstart == 0 && lookahead == 1 (input done a byte at time)
+                 */
+                more--;
+            }
+        }
+
+        /* If the window is almost full and there is insufficient lookahead,
+         * move the upper half to the lower one to make room in the upper half.
+         */
+        if (s->strstart >= wsize+MAX_DIST(s)) {
+
+            zmemcpy(s->window, s->window+wsize, (unsigned)wsize);
+            s->match_start -= wsize;
+            s->strstart    -= wsize; /* we now have strstart >= MAX_DIST */
+            s->block_start -= (long) wsize;
+
+            /* Slide the hash table (could be avoided with 32 bit values
+               at the expense of memory usage). We slide even when level == 0
+               to keep the hash table consistent if we switch back to level > 0
+               later. (Using level 0 permanently is not an optimal usage of
+               zlib, so we don't care about this pathological case.)
+             */
+            n = s->hash_size;
+            p = &s->head[n];
+            do {
+                m = *--p;
+                *p = (Pos)(m >= wsize ? m-wsize : NIL);
+            } while (--n);
+
+            n = wsize;
+#ifndef FASTEST
+            p = &s->prev[n];
+            do {
+                m = *--p;
+                *p = (Pos)(m >= wsize ? m-wsize : NIL);
+                /* If n is not on any hash chain, prev[n] is garbage but
+                 * its value will never be used.
+                 */
+            } while (--n);
+#endif
+            more += wsize;
+        }
+        if (s->strm->avail_in == 0) break;
+
+        /* If there was no sliding:
+         *    strstart <= WSIZE+MAX_DIST-1 && lookahead <= MIN_LOOKAHEAD - 1 &&
+         *    more == window_size - lookahead - strstart
+         * => more >= window_size - (MIN_LOOKAHEAD-1 + WSIZE + MAX_DIST-1)
+         * => more >= window_size - 2*WSIZE + 2
+         * In the BIG_MEM or MMAP case (not yet supported),
+         *   window_size == input_size + MIN_LOOKAHEAD  &&
+         *   strstart + s->lookahead <= input_size => more >= MIN_LOOKAHEAD.
+         * Otherwise, window_size == 2*WSIZE so more >= 2.
+         * If there was sliding, more >= WSIZE. So in all cases, more >= 2.
+         */
+        Assert(more >= 2, "more < 2");
+
+        n = read_buf(s->strm, s->window + s->strstart + s->lookahead, more);
+        s->lookahead += n;
+
+        /* Initialize the hash value now that we have some input: */
+        if (s->lookahead + s->insert >= MIN_MATCH) {
+            uInt str = s->strstart - s->insert;
+            s->ins_h = s->window[str];
+            UPDATE_HASH(s, s->ins_h, s->window[str + 1]);
+#if MIN_MATCH != 3
+            Call UPDATE_HASH() MIN_MATCH-3 more times
+#endif
+            while (s->insert) {
+                UPDATE_HASH(s, s->ins_h, s->window[str + MIN_MATCH-1]);
+#ifndef FASTEST
+                s->prev[str & s->w_mask] = s->head[s->ins_h];
+#endif
+                s->head[s->ins_h] = (Pos)str;
+                str++;
+                s->insert--;
+                if (s->lookahead + s->insert < MIN_MATCH)
+                    break;
+            }
+        }
+        /* If the whole input has less than MIN_MATCH bytes, ins_h is garbage,
+         * but this is not important since only literal bytes will be emitted.
+         */
+
+    } while (s->lookahead < MIN_LOOKAHEAD && s->strm->avail_in != 0);
+
+    /* If the WIN_INIT bytes after the end of the current data have never been
+     * written, then zero those bytes in order to avoid memory check reports of
+     * the use of uninitialized (or uninitialised as Julian writes) bytes by
+     * the longest match routines.  Update the high water mark for the next
+     * time through here.  WIN_INIT is set to MAX_MATCH since the longest match
+     * routines allow scanning to strstart + MAX_MATCH, ignoring lookahead.
+     */
+    if (s->high_water < s->window_size) {
+        ulg curr = s->strstart + (ulg)(s->lookahead);
+        ulg init;
+
+        if (s->high_water < curr) {
+            /* Previous high water mark below current data -- zero WIN_INIT
+             * bytes or up to end of window, whichever is less.
+             */
+            init = s->window_size - curr;
+            if (init > WIN_INIT)
+                init = WIN_INIT;
+            zmemzero(s->window + curr, (unsigned)init);
+            s->high_water = curr + init;
+        }
+        else if (s->high_water < (ulg)curr + WIN_INIT) {
+            /* High water mark at or above current data, but below current data
+             * plus WIN_INIT -- zero out to current data plus WIN_INIT, or up
+             * to end of window, whichever is less.
+             */
+            init = (ulg)curr + WIN_INIT - s->high_water;
+            if (init > s->window_size - s->high_water)
+                init = s->window_size - s->high_water;
+            zmemzero(s->window + s->high_water, (unsigned)init);
+            s->high_water += init;
+        }
+    }
+
+    Assert((ulg)s->strstart <= s->window_size - MIN_LOOKAHEAD,
+           "not enough room for search");
+}
+
+/* ===========================================================================
+ * Flush the current block, with given end-of-file flag.
+ * IN assertion: strstart is set to the end of the current match.
+ */
+#define FLUSH_BLOCK_ONLY(s, last) { \
+   _tr_flush_block(s, (s->block_start >= 0L ? \
+                   (charf *)&s->window[(unsigned)s->block_start] : \
+                   (charf *)Z_NULL), \
+                (ulg)((long)s->strstart - s->block_start), \
+                (last)); \
+   s->block_start = s->strstart; \
+   flush_pending(s->strm); \
+   Tracev((stderr,"[FLUSH]")); \
+}
+
+/* Same but force premature exit if necessary. */
+#define FLUSH_BLOCK(s, last) { \
+   FLUSH_BLOCK_ONLY(s, last); \
+   if (s->strm->avail_out == 0) return (last) ? finish_started : need_more; \
+}
+
+/* ===========================================================================
+ * Copy without compression as much as possible from the input stream, return
+ * the current block state.
+ * This function does not insert new strings in the dictionary since
+ * uncompressible data is probably not useful. This function is used
+ * only for the level=0 compression option.
+ * NOTE: this function should be optimized to avoid extra copying from
+ * window to pending_buf.
+ */
+local block_state deflate_stored(s, flush)
+    deflate_state *s;
+    int flush;
+{
+    /* Stored blocks are limited to 0xffff bytes, pending_buf is limited
+     * to pending_buf_size, and each stored block has a 5 byte header:
+     */
+    ulg max_block_size = 0xffff;
+    ulg max_start;
+
+    if (max_block_size > s->pending_buf_size - 5) {
+        max_block_size = s->pending_buf_size - 5;
+    }
+
+    /* Copy as much as possible from input to output: */
+    for (;;) {
+        /* Fill the window as much as possible: */
+        if (s->lookahead <= 1) {
+
+            Assert(s->strstart < s->w_size+MAX_DIST(s) ||
+                   s->block_start >= (long)s->w_size, "slide too late");
+
+            fill_window(s);
+            if (s->lookahead == 0 && flush == Z_NO_FLUSH) return need_more;
+
+            if (s->lookahead == 0) break; /* flush the current block */
+        }
+        Assert(s->block_start >= 0L, "block gone");
+
+        s->strstart += s->lookahead;
+        s->lookahead = 0;
+
+        /* Emit a stored block if pending_buf will be full: */
+        max_start = s->block_start + max_block_size;
+        if (s->strstart == 0 || (ulg)s->strstart >= max_start) {
+            /* strstart == 0 is possible when wraparound on 16-bit machine */
+            s->lookahead = (uInt)(s->strstart - max_start);
+            s->strstart = (uInt)max_start;
+            FLUSH_BLOCK(s, 0);
+        }
+        /* Flush if we may have to slide, otherwise block_start may become
+         * negative and the data will be gone:
+         */
+        if (s->strstart - (uInt)s->block_start >= MAX_DIST(s)) {
+            FLUSH_BLOCK(s, 0);
+        }
+    }
+    s->insert = 0;
+    if (flush == Z_FINISH) {
+        FLUSH_BLOCK(s, 1);
+        return finish_done;
+    }
+    if ((long)s->strstart > s->block_start)
+        FLUSH_BLOCK(s, 0);
+    return block_done;
+}
+
+/* ===========================================================================
+ * Compress as much as possible from the input stream, return the current
+ * block state.
+ * This function does not perform lazy evaluation of matches and inserts
+ * new strings in the dictionary only for unmatched strings or for short
+ * matches. It is used only for the fast compression options.
+ */
+local block_state deflate_fast(s, flush)
+    deflate_state *s;
+    int flush;
+{
+    IPos hash_head;       /* head of the hash chain */
+    int bflush;           /* set if current block must be flushed */
+
+    for (;;) {
+        /* Make sure that we always have enough lookahead, except
+         * at the end of the input file. We need MAX_MATCH bytes
+         * for the next match, plus MIN_MATCH bytes to insert the
+         * string following the next match.
+         */
+        if (s->lookahead < MIN_LOOKAHEAD) {
+            fill_window(s);
+            if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) {
+                return need_more;
+            }
+            if (s->lookahead == 0) break; /* flush the current block */
+        }
+
+        /* Insert the string window[strstart .. strstart+2] in the
+         * dictionary, and set hash_head to the head of the hash chain:
+         */
+        hash_head = NIL;
+        if (s->lookahead >= MIN_MATCH) {
+            INSERT_STRING(s, s->strstart, hash_head);
+        }
+
+        /* Find the longest match, discarding those <= prev_length.
+         * At this point we have always match_length < MIN_MATCH
+         */
+        if (hash_head != NIL && s->strstart - hash_head <= MAX_DIST(s)) {
+            /* To simplify the code, we prevent matches with the string
+             * of window index 0 (in particular we have to avoid a match
+             * of the string with itself at the start of the input file).
+             */
+            s->match_length = longest_match (s, hash_head);
+            /* longest_match() sets match_start */
+        }
+        if (s->match_length >= MIN_MATCH) {
+            check_match(s, s->strstart, s->match_start, s->match_length);
+
+            _tr_tally_dist(s, s->strstart - s->match_start,
+                           s->match_length - MIN_MATCH, bflush);
+
+            s->lookahead -= s->match_length;
+
+            /* Insert new strings in the hash table only if the match length
+             * is not too large. This saves time but degrades compression.
+             */
+#ifndef FASTEST
+            if (s->match_length <= s->max_insert_length &&
+                s->lookahead >= MIN_MATCH) {
+                s->match_length--; /* string at strstart already in table */
+                do {
+                    s->strstart++;
+                    INSERT_STRING(s, s->strstart, hash_head);
+                    /* strstart never exceeds WSIZE-MAX_MATCH, so there are
+                     * always MIN_MATCH bytes ahead.
+                     */
+                } while (--s->match_length != 0);
+                s->strstart++;
+            } else
+#endif
+            {
+                s->strstart += s->match_length;
+                s->match_length = 0;
+                s->ins_h = s->window[s->strstart];
+                UPDATE_HASH(s, s->ins_h, s->window[s->strstart+1]);
+#if MIN_MATCH != 3
+                Call UPDATE_HASH() MIN_MATCH-3 more times
+#endif
+                /* If lookahead < MIN_MATCH, ins_h is garbage, but it does not
+                 * matter since it will be recomputed at next deflate call.
+                 */
+            }
+        } else {
+            /* No match, output a literal byte */
+            Tracevv((stderr,"%c", s->window[s->strstart]));
+            _tr_tally_lit (s, s->window[s->strstart], bflush);
+            s->lookahead--;
+            s->strstart++;
+        }
+        if (bflush) FLUSH_BLOCK(s, 0);
+    }
+    s->insert = s->strstart < MIN_MATCH-1 ? s->strstart : MIN_MATCH-1;
+    if (flush == Z_FINISH) {
+        FLUSH_BLOCK(s, 1);
+        return finish_done;
+    }
+    if (s->last_lit)
+        FLUSH_BLOCK(s, 0);
+    return block_done;
+}
+
+#ifndef FASTEST
+/* ===========================================================================
+ * Same as above, but achieves better compression. We use a lazy
+ * evaluation for matches: a match is finally adopted only if there is
+ * no better match at the next window position.
+ */
+local block_state deflate_slow(s, flush)
+    deflate_state *s;
+    int flush;
+{
+    IPos hash_head;          /* head of hash chain */
+    int bflush;              /* set if current block must be flushed */
+
+    /* Process the input block. */
+    for (;;) {
+        /* Make sure that we always have enough lookahead, except
+         * at the end of the input file. We need MAX_MATCH bytes
+         * for the next match, plus MIN_MATCH bytes to insert the
+         * string following the next match.
+         */
+        if (s->lookahead < MIN_LOOKAHEAD) {
+            fill_window(s);
+            if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) {
+                return need_more;
+            }
+            if (s->lookahead == 0) break; /* flush the current block */
+        }
+
+        /* Insert the string window[strstart .. strstart+2] in the
+         * dictionary, and set hash_head to the head of the hash chain:
+         */
+        hash_head = NIL;
+        if (s->lookahead >= MIN_MATCH) {
+            INSERT_STRING(s, s->strstart, hash_head);
+        }
+
+        /* Find the longest match, discarding those <= prev_length.
+         */
+        s->prev_length = s->match_length, s->prev_match = s->match_start;
+        s->match_length = MIN_MATCH-1;
+
+        if (hash_head != NIL && s->prev_length < s->max_lazy_match &&
+            s->strstart - hash_head <= MAX_DIST(s)) {
+            /* To simplify the code, we prevent matches with the string
+             * of window index 0 (in particular we have to avoid a match
+             * of the string with itself at the start of the input file).
+             */
+            s->match_length = longest_match (s, hash_head);
+            /* longest_match() sets match_start */
+
+            if (s->match_length <= 5 && (s->strategy == Z_FILTERED
+#if TOO_FAR <= 32767
+                || (s->match_length == MIN_MATCH &&
+                    s->strstart - s->match_start > TOO_FAR)
+#endif
+                )) {
+
+                /* If prev_match is also MIN_MATCH, match_start is garbage
+                 * but we will ignore the current match anyway.
+                 */
+                s->match_length = MIN_MATCH-1;
+            }
+        }
+        /* If there was a match at the previous step and the current
+         * match is not better, output the previous match:
+         */
+        if (s->prev_length >= MIN_MATCH && s->match_length <= s->prev_length) {
+            uInt max_insert = s->strstart + s->lookahead - MIN_MATCH;
+            /* Do not insert strings in hash table beyond this. */
+
+            check_match(s, s->strstart-1, s->prev_match, s->prev_length);
+
+            _tr_tally_dist(s, s->strstart -1 - s->prev_match,
+                           s->prev_length - MIN_MATCH, bflush);
+
+            /* Insert in hash table all strings up to the end of the match.
+             * strstart-1 and strstart are already inserted. If there is not
+             * enough lookahead, the last two strings are not inserted in
+             * the hash table.
+             */
+            s->lookahead -= s->prev_length-1;
+            s->prev_length -= 2;
+            do {
+                if (++s->strstart <= max_insert) {
+                    INSERT_STRING(s, s->strstart, hash_head);
+                }
+            } while (--s->prev_length != 0);
+            s->match_available = 0;
+            s->match_length = MIN_MATCH-1;
+            s->strstart++;
+
+            if (bflush) FLUSH_BLOCK(s, 0);
+
+        } else if (s->match_available) {
+            /* If there was no match at the previous position, output a
+             * single literal. If there was a match but the current match
+             * is longer, truncate the previous match to a single literal.
+             */
+            Tracevv((stderr,"%c", s->window[s->strstart-1]));
+            _tr_tally_lit(s, s->window[s->strstart-1], bflush);
+            if (bflush) {
+                FLUSH_BLOCK_ONLY(s, 0);
+            }
+            s->strstart++;
+            s->lookahead--;
+            if (s->strm->avail_out == 0) return need_more;
+        } else {
+            /* There is no previous match to compare with, wait for
+             * the next step to decide.
+             */
+            s->match_available = 1;
+            s->strstart++;
+            s->lookahead--;
+        }
+    }
+    Assert (flush != Z_NO_FLUSH, "no flush?");
+    if (s->match_available) {
+        Tracevv((stderr,"%c", s->window[s->strstart-1]));
+        _tr_tally_lit(s, s->window[s->strstart-1], bflush);
+        s->match_available = 0;
+    }
+    s->insert = s->strstart < MIN_MATCH-1 ? s->strstart : MIN_MATCH-1;
+    if (flush == Z_FINISH) {
+        FLUSH_BLOCK(s, 1);
+        return finish_done;
+    }
+    if (s->last_lit)
+        FLUSH_BLOCK(s, 0);
+    return block_done;
+}
+#endif /* FASTEST */
+
+/* ===========================================================================
+ * For Z_RLE, simply look for runs of bytes, generate matches only of distance
+ * one.  Do not maintain a hash table.  (It will be regenerated if this run of
+ * deflate switches away from Z_RLE.)
+ */
+local block_state deflate_rle(s, flush)
+    deflate_state *s;
+    int flush;
+{
+    int bflush;             /* set if current block must be flushed */
+    uInt prev;              /* byte at distance one to match */
+    Bytef *scan, *strend;   /* scan goes up to strend for length of run */
+
+    for (;;) {
+        /* Make sure that we always have enough lookahead, except
+         * at the end of the input file. We need MAX_MATCH bytes
+         * for the longest run, plus one for the unrolled loop.
+         */
+        if (s->lookahead <= MAX_MATCH) {
+            fill_window(s);
+            if (s->lookahead <= MAX_MATCH && flush == Z_NO_FLUSH) {
+                return need_more;
+            }
+            if (s->lookahead == 0) break; /* flush the current block */
+        }
+
+        /* See how many times the previous byte repeats */
+        s->match_length = 0;
+        if (s->lookahead >= MIN_MATCH && s->strstart > 0) {
+            scan = s->window + s->strstart - 1;
+            prev = *scan;
+            if (prev == *++scan && prev == *++scan && prev == *++scan) {
+                strend = s->window + s->strstart + MAX_MATCH;
+                do {
+                } while (prev == *++scan && prev == *++scan &&
+                         prev == *++scan && prev == *++scan &&
+                         prev == *++scan && prev == *++scan &&
+                         prev == *++scan && prev == *++scan &&
+                         scan < strend);
+                s->match_length = MAX_MATCH - (int)(strend - scan);
+                if (s->match_length > s->lookahead)
+                    s->match_length = s->lookahead;
+            }
+            Assert(scan <= s->window+(uInt)(s->window_size-1), "wild scan");
+        }
+
+        /* Emit match if have run of MIN_MATCH or longer, else emit literal */
+        if (s->match_length >= MIN_MATCH) {
+            check_match(s, s->strstart, s->strstart - 1, s->match_length);
+
+            _tr_tally_dist(s, 1, s->match_length - MIN_MATCH, bflush);
+
+            s->lookahead -= s->match_length;
+            s->strstart += s->match_length;
+            s->match_length = 0;
+        } else {
+            /* No match, output a literal byte */
+            Tracevv((stderr,"%c", s->window[s->strstart]));
+            _tr_tally_lit (s, s->window[s->strstart], bflush);
+            s->lookahead--;
+            s->strstart++;
+        }
+        if (bflush) FLUSH_BLOCK(s, 0);
+    }
+    s->insert = 0;
+    if (flush == Z_FINISH) {
+        FLUSH_BLOCK(s, 1);
+        return finish_done;
+    }
+    if (s->last_lit)
+        FLUSH_BLOCK(s, 0);
+    return block_done;
+}
+
+/* ===========================================================================
+ * For Z_HUFFMAN_ONLY, do not look for matches.  Do not maintain a hash table.
+ * (It will be regenerated if this run of deflate switches away from Huffman.)
+ */
+local block_state deflate_huff(s, flush)
+    deflate_state *s;
+    int flush;
+{
+    int bflush;             /* set if current block must be flushed */
+
+    for (;;) {
+        /* Make sure that we have a literal to write. */
+        if (s->lookahead == 0) {
+            fill_window(s);
+            if (s->lookahead == 0) {
+                if (flush == Z_NO_FLUSH)
+                    return need_more;
+                break;      /* flush the current block */
+            }
+        }
+
+        /* Output a literal byte */
+        s->match_length = 0;
+        Tracevv((stderr,"%c", s->window[s->strstart]));
+        _tr_tally_lit (s, s->window[s->strstart], bflush);
+        s->lookahead--;
+        s->strstart++;
+        if (bflush) FLUSH_BLOCK(s, 0);
+    }
+    s->insert = 0;
+    if (flush == Z_FINISH) {
+        FLUSH_BLOCK(s, 1);
+        return finish_done;
+    }
+    if (s->last_lit)
+        FLUSH_BLOCK(s, 0);
+    return block_done;
+}
diff --git a/ml/dlib/dlib/external/zlib/deflate.h b/ml/dlib/dlib/external/zlib/deflate.h
new file mode 100644
index 000000000..ce0299edd
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/deflate.h
@@ -0,0 +1,346 @@
+/* deflate.h -- internal compression state
+ * Copyright (C) 1995-2012 Jean-loup Gailly
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* WARNING: this file should *not* be used by applications. It is
+   part of the implementation of the compression library and is
+   subject to change. Applications should only use zlib.h.
+ */
+
+/* @(#) $Id$ */
+
+#ifndef DEFLATE_H
+#define DEFLATE_H
+
+#include "zutil.h"
+
+/* define NO_GZIP when compiling if you want to disable gzip header and
+   trailer creation by deflate().  NO_GZIP would be used to avoid linking in
+   the crc code when it is not needed.  For shared libraries, gzip encoding
+   should be left enabled. */
+#ifndef NO_GZIP
+#  define GZIP
+#endif
+
+/* ===========================================================================
+ * Internal compression state.
+ */
+
+#define LENGTH_CODES 29
+/* number of length codes, not counting the special END_BLOCK code */
+
+#define LITERALS  256
+/* number of literal bytes 0..255 */
+
+#define L_CODES (LITERALS+1+LENGTH_CODES)
+/* number of Literal or Length codes, including the END_BLOCK code */
+
+#define D_CODES   30
+/* number of distance codes */
+
+#define BL_CODES  19
+/* number of codes used to transfer the bit lengths */
+
+#define HEAP_SIZE (2*L_CODES+1)
+/* maximum heap size */
+
+#define MAX_BITS 15
+/* All codes must not exceed MAX_BITS bits */
+
+#define Buf_size 16
+/* size of bit buffer in bi_buf */
+
+#define INIT_STATE    42
+#define EXTRA_STATE   69
+#define NAME_STATE    73
+#define COMMENT_STATE 91
+#define HCRC_STATE   103
+#define BUSY_STATE   113
+#define FINISH_STATE 666
+/* Stream status */
+
+
+/* Data structure describing a single value and its code string. */
+typedef struct ct_data_s {
+    union {
+        ush  freq;       /* frequency count */
+        ush  code;       /* bit string */
+    } fc;
+    union {
+        ush  dad;        /* father node in Huffman tree */
+        ush  len;        /* length of bit string */
+    } dl;
+} FAR ct_data;
+
+#define Freq fc.freq
+#define Code fc.code
+#define Dad  dl.dad
+#define Len  dl.len
+
+typedef struct static_tree_desc_s  static_tree_desc;
+
+typedef struct tree_desc_s {
+    ct_data *dyn_tree;           /* the dynamic tree */
+    int     max_code;            /* largest code with non zero frequency */
+    static_tree_desc *stat_desc; /* the corresponding static tree */
+} FAR tree_desc;
+
+typedef ush Pos;
+typedef Pos FAR Posf;
+typedef unsigned IPos;
+
+/* A Pos is an index in the character window. We use short instead of int to
+ * save space in the various tables. IPos is used only for parameter passing.
+ */
+
+typedef struct internal_state {
+    z_streamp strm;      /* pointer back to this zlib stream */
+    int   status;        /* as the name implies */
+    Bytef *pending_buf;  /* output still pending */
+    ulg   pending_buf_size; /* size of pending_buf */
+    Bytef *pending_out;  /* next pending byte to output to the stream */
+    uInt   pending;      /* nb of bytes in the pending buffer */
+    int   wrap;          /* bit 0 true for zlib, bit 1 true for gzip */
+    gz_headerp  gzhead;  /* gzip header information to write */
+    uInt   gzindex;      /* where in extra, name, or comment */
+    Byte  method;        /* can only be DEFLATED */
+    int   last_flush;    /* value of flush param for previous deflate call */
+
+                /* used by deflate.c: */
+
+    uInt  w_size;        /* LZ77 window size (32K by default) */
+    uInt  w_bits;        /* log2(w_size)  (8..16) */
+    uInt  w_mask;        /* w_size - 1 */
+
+    Bytef *window;
+    /* Sliding window. Input bytes are read into the second half of the window,
+     * and move to the first half later to keep a dictionary of at least wSize
+     * bytes. With this organization, matches are limited to a distance of
+     * wSize-MAX_MATCH bytes, but this ensures that IO is always
+     * performed with a length multiple of the block size. Also, it limits
+     * the window size to 64K, which is quite useful on MSDOS.
+     * To do: use the user input buffer as sliding window.
+     */
+
+    ulg window_size;
+    /* Actual size of window: 2*wSize, except when the user input buffer
+     * is directly used as sliding window.
+     */
+
+    Posf *prev;
+    /* Link to older string with same hash index. To limit the size of this
+     * array to 64K, this link is maintained only for the last 32K strings.
+     * An index in this array is thus a window index modulo 32K.
+     */
+
+    Posf *head; /* Heads of the hash chains or NIL. */
+
+    uInt  ins_h;          /* hash index of string to be inserted */
+    uInt  hash_size;      /* number of elements in hash table */
+    uInt  hash_bits;      /* log2(hash_size) */
+    uInt  hash_mask;      /* hash_size-1 */
+
+    uInt  hash_shift;
+    /* Number of bits by which ins_h must be shifted at each input
+     * step. It must be such that after MIN_MATCH steps, the oldest
+     * byte no longer takes part in the hash key, that is:
+     *   hash_shift * MIN_MATCH >= hash_bits
+     */
+
+    long block_start;
+    /* Window position at the beginning of the current output block. Gets
+     * negative when the window is moved backwards.
+     */
+
+    uInt match_length;           /* length of best match */
+    IPos prev_match;             /* previous match */
+    int match_available;         /* set if previous match exists */
+    uInt strstart;               /* start of string to insert */
+    uInt match_start;            /* start of matching string */
+    uInt lookahead;              /* number of valid bytes ahead in window */
+
+    uInt prev_length;
+    /* Length of the best match at previous step. Matches not greater than this
+     * are discarded. This is used in the lazy match evaluation.
+     */
+
+    uInt max_chain_length;
+    /* To speed up deflation, hash chains are never searched beyond this
+     * length.  A higher limit improves compression ratio but degrades the
+     * speed.
+     */
+
+    uInt max_lazy_match;
+    /* Attempt to find a better match only when the current match is strictly
+     * smaller than this value. This mechanism is used only for compression
+     * levels >= 4.
+     */
+#   define max_insert_length  max_lazy_match
+    /* Insert new strings in the hash table only if the match length is not
+     * greater than this length. This saves time but degrades compression.
+     * max_insert_length is used only for compression levels <= 3.
+     */
+
+    int level;    /* compression level (1..9) */
+    int strategy; /* favor or force Huffman coding*/
+
+    uInt good_match;
+    /* Use a faster search when the previous match is longer than this */
+
+    int nice_match; /* Stop searching when current match exceeds this */
+
+                /* used by trees.c: */
+    /* Didn't use ct_data typedef below to suppress compiler warning */
+    struct ct_data_s dyn_ltree[HEAP_SIZE];   /* literal and length tree */
+    struct ct_data_s dyn_dtree[2*D_CODES+1]; /* distance tree */
+    struct ct_data_s bl_tree[2*BL_CODES+1];  /* Huffman tree for bit lengths */
+
+    struct tree_desc_s l_desc;               /* desc. for literal tree */
+    struct tree_desc_s d_desc;               /* desc. for distance tree */
+    struct tree_desc_s bl_desc;              /* desc. for bit length tree */
+
+    ush bl_count[MAX_BITS+1];
+    /* number of codes at each bit length for an optimal tree */
+
+    int heap[2*L_CODES+1];      /* heap used to build the Huffman trees */
+    int heap_len;               /* number of elements in the heap */
+    int heap_max;               /* element of largest frequency */
+    /* The sons of heap[n] are heap[2*n] and heap[2*n+1]. heap[0] is not used.
+     * The same heap array is used to build all trees.
+     */
+
+    uch depth[2*L_CODES+1];
+    /* Depth of each subtree used as tie breaker for trees of equal frequency
+     */
+
+    uchf *l_buf;          /* buffer for literals or lengths */
+
+    uInt  lit_bufsize;
+    /* Size of match buffer for literals/lengths.  There are 4 reasons for
+     * limiting lit_bufsize to 64K:
+     *   - frequencies can be kept in 16 bit counters
+     *   - if compression is not successful for the first block, all input
+     *     data is still in the window so we can still emit a stored block even
+     *     when input comes from standard input.  (This can also be done for
+     *     all blocks if lit_bufsize is not greater than 32K.)
+     *   - if compression is not successful for a file smaller than 64K, we can
+     *     even emit a stored file instead of a stored block (saving 5 bytes).
+     *     This is applicable only for zip (not gzip or zlib).
+     *   - creating new Huffman trees less frequently may not provide fast
+     *     adaptation to changes in the input data statistics. (Take for
+     *     example a binary file with poorly compressible code followed by
+     *     a highly compressible string table.) Smaller buffer sizes give
+     *     fast adaptation but have of course the overhead of transmitting
+     *     trees more frequently.
+     *   - I can't count above 4
+     */
+
+    uInt last_lit;      /* running index in l_buf */
+
+    ushf *d_buf;
+    /* Buffer for distances. To simplify the code, d_buf and l_buf have
+     * the same number of elements. To use different lengths, an extra flag
+     * array would be necessary.
+     */
+
+    ulg opt_len;        /* bit length of current block with optimal trees */
+    ulg static_len;     /* bit length of current block with static trees */
+    uInt matches;       /* number of string matches in current block */
+    uInt insert;        /* bytes at end of window left to insert */
+
+#ifdef DEBUG
+    ulg compressed_len; /* total bit length of compressed file mod 2^32 */
+    ulg bits_sent;      /* bit length of compressed data sent mod 2^32 */
+#endif
+
+    ush bi_buf;
+    /* Output buffer. bits are inserted starting at the bottom (least
+     * significant bits).
+     */
+    int bi_valid;
+    /* Number of valid bits in bi_buf.  All bits above the last valid bit
+     * are always zero.
+     */
+
+    ulg high_water;
+    /* High water mark offset in window for initialized bytes -- bytes above
+     * this are set to zero in order to avoid memory check warnings when
+     * longest match routines access bytes past the input.  This is then
+     * updated to the new high water mark.
+     */
+
+} FAR deflate_state;
+
+/* Output a byte on the stream.
+ * IN assertion: there is enough room in pending_buf.
+ */
+#define put_byte(s, c) {s->pending_buf[s->pending++] = (c);}
+
+
+#define MIN_LOOKAHEAD (MAX_MATCH+MIN_MATCH+1)
+/* Minimum amount of lookahead, except at the end of the input file.
+ * See deflate.c for comments about the MIN_MATCH+1.
+ */
+
+#define MAX_DIST(s)  ((s)->w_size-MIN_LOOKAHEAD)
+/* In order to simplify the code, particularly on 16 bit machines, match
+ * distances are limited to MAX_DIST instead of WSIZE.
+ */
+
+#define WIN_INIT MAX_MATCH
+/* Number of bytes after end of data in window to initialize in order to avoid
+   memory checker errors from longest match routines */
+
+        /* in trees.c */
+void ZLIB_INTERNAL _tr_init OF((deflate_state *s));
+int ZLIB_INTERNAL _tr_tally OF((deflate_state *s, unsigned dist, unsigned lc));
+void ZLIB_INTERNAL _tr_flush_block OF((deflate_state *s, charf *buf,
+                        ulg stored_len, int last));
+void ZLIB_INTERNAL _tr_flush_bits OF((deflate_state *s));
+void ZLIB_INTERNAL _tr_align OF((deflate_state *s));
+void ZLIB_INTERNAL _tr_stored_block OF((deflate_state *s, charf *buf,
+                        ulg stored_len, int last));
+
+#define d_code(dist) \
+   ((dist) < 256 ? _dist_code[dist] : _dist_code[256+((dist)>>7)])
+/* Mapping from a distance to a distance code. dist is the distance - 1 and
+ * must not have side effects. _dist_code[256] and _dist_code[257] are never
+ * used.
+ */
+
+#ifndef DEBUG
+/* Inline versions of _tr_tally for speed: */
+
+#if defined(GEN_TREES_H) || !defined(STDC)
+  extern uch ZLIB_INTERNAL _length_code[];
+  extern uch ZLIB_INTERNAL _dist_code[];
+#else
+  extern const uch ZLIB_INTERNAL _length_code[];
+  extern const uch ZLIB_INTERNAL _dist_code[];
+#endif
+
+# define _tr_tally_lit(s, c, flush) \
+  { uch cc = (c); \
+    s->d_buf[s->last_lit] = 0; \
+    s->l_buf[s->last_lit++] = cc; \
+    s->dyn_ltree[cc].Freq++; \
+    flush = (s->last_lit == s->lit_bufsize-1); \
+   }
+# define _tr_tally_dist(s, distance, length, flush) \
+  { uch len = (length); \
+    ush dist = (distance); \
+    s->d_buf[s->last_lit] = dist; \
+    s->l_buf[s->last_lit++] = len; \
+    dist--; \
+    s->dyn_ltree[_length_code[len]+LITERALS+1].Freq++; \
+    s->dyn_dtree[d_code(dist)].Freq++; \
+    flush = (s->last_lit == s->lit_bufsize-1); \
+  }
+#else
+# define _tr_tally_lit(s, c, flush) flush = _tr_tally(s, 0, c)
+# define _tr_tally_dist(s, distance, length, flush) \
+              flush = _tr_tally(s, distance, length)
+#endif
+
+#endif /* DEFLATE_H */
diff --git a/ml/dlib/dlib/external/zlib/gzclose.c b/ml/dlib/dlib/external/zlib/gzclose.c
new file mode 100644
index 000000000..caeb99a31
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/gzclose.c
@@ -0,0 +1,25 @@
+/* gzclose.c -- zlib gzclose() function
+ * Copyright (C) 2004, 2010 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "gzguts.h"
+
+/* gzclose() is in a separate file so that it is linked in only if it is used.
+   That way the other gzclose functions can be used instead to avoid linking in
+   unneeded compression or decompression routines. */
+int ZEXPORT gzclose(file)
+    gzFile file;
+{
+#ifndef NO_GZCOMPRESS
+    gz_statep state;
+
+    if (file == NULL)
+        return Z_STREAM_ERROR;
+    state = (gz_statep)file;
+
+    return state->mode == GZ_READ ? gzclose_r(file) : gzclose_w(file);
+#else
+    return gzclose_r(file);
+#endif
+}
diff --git a/ml/dlib/dlib/external/zlib/gzguts.h b/ml/dlib/dlib/external/zlib/gzguts.h
new file mode 100644
index 000000000..2bb0b0499
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/gzguts.h
@@ -0,0 +1,219 @@
+/* gzguts.h -- zlib internal header definitions for gz* operations
+ * Copyright (C) 2004, 2005, 2010, 2011, 2012, 2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#ifdef _MSC_VER
+// Disable the following warnings for Visual Studio
+// This is a warning you get from visual studio 2005 about things in the standard C++
+// library being "deprecated."  I checked the C++ standard and it doesn't say jack 
+// about any of them (I checked the searchable PDF).   So this warning is total Bunk.
+#pragma warning(disable : 4996)
+#endif
+
+#ifdef _LARGEFILE64_SOURCE
+#  ifndef _LARGEFILE_SOURCE
+#    define _LARGEFILE_SOURCE 1
+#  endif
+#  ifdef _FILE_OFFSET_BITS
+#    undef _FILE_OFFSET_BITS
+#  endif
+#endif
+
+#ifdef HAVE_HIDDEN
+#  define ZLIB_INTERNAL __attribute__((visibility ("hidden")))
+#else
+#  define ZLIB_INTERNAL
+#endif
+
+#include <stdio.h>
+#include "zlib.h"
+#ifdef STDC
+#  include <string.h>
+#  include <stdlib.h>
+#  include <limits.h>
+#endif
+#include <fcntl.h>
+
+#ifdef _WIN32
+#  include <stddef.h>
+#endif
+
+#if defined(__TURBOC__) || defined(_MSC_VER) || defined(_WIN32)
+#  include <io.h>
+#else
+#  include <unistd.h>
+#endif
+
+#ifdef WINAPI_FAMILY
+#  define open _open
+#  define read _read
+#  define write _write
+#  define close _close
+#endif
+
+#ifdef NO_DEFLATE       /* for compatibility with old definition */
+#  define NO_GZCOMPRESS
+#endif
+
+#if defined(STDC99) || (defined(__TURBOC__) && __TURBOC__ >= 0x550)
+#  ifndef HAVE_VSNPRINTF
+#    define HAVE_VSNPRINTF
+#  endif
+#endif
+
+#if defined(__CYGWIN__)
+#  ifndef HAVE_VSNPRINTF
+#    define HAVE_VSNPRINTF
+#  endif
+#endif
+
+#if defined(MSDOS) && defined(__BORLANDC__) && (BORLANDC > 0x410)
+#  ifndef HAVE_VSNPRINTF
+#    define HAVE_VSNPRINTF
+#  endif
+#endif
+
+#ifndef HAVE_VSNPRINTF
+#  ifdef MSDOS
+/* vsnprintf may exist on some MS-DOS compilers (DJGPP?),
+   but for now we just assume it doesn't. */
+#    define NO_vsnprintf
+#  endif
+#  ifdef __TURBOC__
+#    define NO_vsnprintf
+#  endif
+#  ifdef WIN32
+/* In Win32, vsnprintf is available as the "non-ANSI" _vsnprintf. */
+#    if !defined(vsnprintf) && !defined(NO_vsnprintf)
+#      if !defined(_MSC_VER) || ( defined(_MSC_VER) && _MSC_VER < 1500 )
+#         define vsnprintf _vsnprintf
+#      endif
+#    endif
+#  endif
+#  ifdef __SASC
+#    define NO_vsnprintf
+#  endif
+#  ifdef VMS
+#    define NO_vsnprintf
+#  endif
+#  ifdef __OS400__
+#    define NO_vsnprintf
+#  endif
+#  ifdef __MVS__
+#    define NO_vsnprintf
+#  endif
+#endif
+
+/* unlike snprintf (which is required in C99, yet still not supported by
+   Microsoft more than a decade later!), _snprintf does not guarantee null
+   termination of the result -- however this is only used in gzlib.c where
+   the result is assured to fit in the space provided */
+#ifdef _MSC_VER
+#  define snprintf _snprintf
+#endif
+
+#ifndef local
+#  define local static
+#endif
+/* compile with -Dlocal if your debugger can't find static symbols */
+
+/* gz* functions always use library allocation functions */
+#ifndef STDC
+  extern voidp  malloc OF((uInt size));
+  extern void   free   OF((voidpf ptr));
+#endif
+
+/* get errno and strerror definition */
+#if defined UNDER_CE
+#  include <windows.h>
+#  define zstrerror() gz_strwinerror((DWORD)GetLastError())
+#else
+#  ifndef NO_STRERROR
+#    include <errno.h>
+#    define zstrerror() strerror(errno)
+#  else
+#    define zstrerror() "stdio error (consult errno)"
+#  endif
+#endif
+
+/* provide prototypes for these when building zlib without LFS */
+#if !defined(_LARGEFILE64_SOURCE) || _LFS64_LARGEFILE-0 == 0
+    ZEXTERN gzFile ZEXPORT gzopen64 OF((const char *, const char *));
+    ZEXTERN z_off64_t ZEXPORT gzseek64 OF((gzFile, z_off64_t, int));
+    ZEXTERN z_off64_t ZEXPORT gztell64 OF((gzFile));
+    ZEXTERN z_off64_t ZEXPORT gzoffset64 OF((gzFile));
+#endif
+
+/* default memLevel */
+#if MAX_MEM_LEVEL >= 8
+#  define DEF_MEM_LEVEL 8
+#else
+#  define DEF_MEM_LEVEL  MAX_MEM_LEVEL
+#endif
+
+/* default i/o buffer size -- double this for output when reading (this and
+   twice this must be able to fit in an unsigned type) */
+#define GZBUFSIZE 8192
+
+/* gzip modes, also provide a little integrity check on the passed structure */
+#define GZ_NONE 0
+#define GZ_READ 7247
+#define GZ_WRITE 31153
+#define GZ_APPEND 1     /* mode set to GZ_WRITE after the file is opened */
+
+/* values for gz_state how */
+#define LOOK 0      /* look for a gzip header */
+#define COPY 1      /* copy input directly */
+#define GZIP 2      /* decompress a gzip stream */
+
+/* internal gzip file state data structure */
+typedef struct {
+        /* exposed contents for gzgetc() macro */
+    struct gzFile_s x;      /* "x" for exposed */
+                            /* x.have: number of bytes available at x.next */
+                            /* x.next: next output data to deliver or write */
+                            /* x.pos: current position in uncompressed data */
+        /* used for both reading and writing */
+    int mode;               /* see gzip modes above */
+    int fd;                 /* file descriptor */
+    char *path;             /* path or fd for error messages */
+    unsigned size;          /* buffer size, zero if not allocated yet */
+    unsigned want;          /* requested buffer size, default is GZBUFSIZE */
+    unsigned char *in;      /* input buffer */
+    unsigned char *out;     /* output buffer (double-sized when reading) */
+    int direct;             /* 0 if processing gzip, 1 if transparent */
+        /* just for reading */
+    int how;                /* 0: get header, 1: copy, 2: decompress */
+    z_off64_t start;        /* where the gzip data started, for rewinding */
+    int eof;                /* true if end of input file reached */
+    int past;               /* true if read requested past end */
+        /* just for writing */
+    int level;              /* compression level */
+    int strategy;           /* compression strategy */
+        /* seek request */
+    z_off64_t skip;         /* amount to skip (already rewound if backwards) */
+    int seek;               /* true if seek request pending */
+        /* error information */
+    int err;                /* error code */
+    char *msg;              /* error message */
+        /* zlib inflate or deflate stream */
+    z_stream strm;          /* stream structure in-place (not a pointer) */
+} gz_state;
+typedef gz_state FAR *gz_statep;
+
+/* shared functions */
+void ZLIB_INTERNAL gz_error OF((gz_statep, int, const char *));
+#if defined UNDER_CE
+char ZLIB_INTERNAL *gz_strwinerror OF((DWORD error));
+#endif
+
+/* GT_OFF(x), where x is an unsigned value, is true if x > maximum z_off64_t
+   value -- needed when comparing unsigned to z_off64_t, which is signed
+   (possible z_off64_t types off_t, off64_t, and long are all signed) */
+#ifdef INT_MAX
+#  define GT_OFF(x) (sizeof(int) == sizeof(z_off64_t) && (x) > INT_MAX)
+#else
+unsigned ZLIB_INTERNAL gz_intmax OF((void));
+#  define GT_OFF(x) (sizeof(int) == sizeof(z_off64_t) && (x) > gz_intmax())
+#endif
diff --git a/ml/dlib/dlib/external/zlib/gzlib.c b/ml/dlib/dlib/external/zlib/gzlib.c
new file mode 100644
index 000000000..fae202ef8
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/gzlib.c
@@ -0,0 +1,634 @@
+/* gzlib.c -- zlib functions common to reading and writing gzip files
+ * Copyright (C) 2004, 2010, 2011, 2012, 2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "gzguts.h"
+
+#if defined(_WIN32) && !defined(__BORLANDC__)
+#  define LSEEK _lseeki64
+#else
+#if defined(_LARGEFILE64_SOURCE) && _LFS64_LARGEFILE-0
+#  define LSEEK lseek64
+#else
+#  define LSEEK lseek
+#endif
+#endif
+
+/* Local functions */
+local void gz_reset OF((gz_statep));
+local gzFile gz_open OF((const void *, int, const char *));
+
+#if defined UNDER_CE
+
+/* Map the Windows error number in ERROR to a locale-dependent error message
+   string and return a pointer to it.  Typically, the values for ERROR come
+   from GetLastError.
+
+   The string pointed to shall not be modified by the application, but may be
+   overwritten by a subsequent call to gz_strwinerror
+
+   The gz_strwinerror function does not change the current setting of
+   GetLastError. */
+char ZLIB_INTERNAL *gz_strwinerror (error)
+     DWORD error;
+{
+    static char buf[1024];
+
+    wchar_t *msgbuf;
+    DWORD lasterr = GetLastError();
+    DWORD chars = FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM
+        | FORMAT_MESSAGE_ALLOCATE_BUFFER,
+        NULL,
+        error,
+        0, /* Default language */
+        (LPVOID)&msgbuf,
+        0,
+        NULL);
+    if (chars != 0) {
+        /* If there is an \r\n appended, zap it.  */
+        if (chars >= 2
+            && msgbuf[chars - 2] == '\r' && msgbuf[chars - 1] == '\n') {
+            chars -= 2;
+            msgbuf[chars] = 0;
+        }
+
+        if (chars > sizeof (buf) - 1) {
+            chars = sizeof (buf) - 1;
+            msgbuf[chars] = 0;
+        }
+
+        wcstombs(buf, msgbuf, chars + 1);
+        LocalFree(msgbuf);
+    }
+    else {
+        sprintf(buf, "unknown win32 error (%ld)", error);
+    }
+
+    SetLastError(lasterr);
+    return buf;
+}
+
+#endif /* UNDER_CE */
+
+/* Reset gzip file state */
+local void gz_reset(state)
+    gz_statep state;
+{
+    state->x.have = 0;              /* no output data available */
+    if (state->mode == GZ_READ) {   /* for reading ... */
+        state->eof = 0;             /* not at end of file */
+        state->past = 0;            /* have not read past end yet */
+        state->how = LOOK;          /* look for gzip header */
+    }
+    state->seek = 0;                /* no seek request pending */
+    gz_error(state, Z_OK, NULL);    /* clear error */
+    state->x.pos = 0;               /* no uncompressed data yet */
+    state->strm.avail_in = 0;       /* no input data yet */
+}
+
+/* Open a gzip file either by name or file descriptor. */
+local gzFile gz_open(path, fd, mode)
+    const void *path;
+    int fd;
+    const char *mode;
+{
+    gz_statep state;
+    size_t len;
+    int oflag;
+#ifdef O_CLOEXEC
+    int cloexec = 0;
+#endif
+#ifdef O_EXCL
+    int exclusive = 0;
+#endif
+
+    /* check input */
+    if (path == NULL)
+        return NULL;
+
+    /* allocate gzFile structure to return */
+    state = (gz_statep)malloc(sizeof(gz_state));
+    if (state == NULL)
+        return NULL;
+    state->size = 0;            /* no buffers allocated yet */
+    state->want = GZBUFSIZE;    /* requested buffer size */
+    state->msg = NULL;          /* no error message yet */
+
+    /* interpret mode */
+    state->mode = GZ_NONE;
+    state->level = Z_DEFAULT_COMPRESSION;
+    state->strategy = Z_DEFAULT_STRATEGY;
+    state->direct = 0;
+    while (*mode) {
+        if (*mode >= '0' && *mode <= '9')
+            state->level = *mode - '0';
+        else
+            switch (*mode) {
+            case 'r':
+                state->mode = GZ_READ;
+                break;
+#ifndef NO_GZCOMPRESS
+            case 'w':
+                state->mode = GZ_WRITE;
+                break;
+            case 'a':
+                state->mode = GZ_APPEND;
+                break;
+#endif
+            case '+':       /* can't read and write at the same time */
+                free(state);
+                return NULL;
+            case 'b':       /* ignore -- will request binary anyway */
+                break;
+#ifdef O_CLOEXEC
+            case 'e':
+                cloexec = 1;
+                break;
+#endif
+#ifdef O_EXCL
+            case 'x':
+                exclusive = 1;
+                break;
+#endif
+            case 'f':
+                state->strategy = Z_FILTERED;
+                break;
+            case 'h':
+                state->strategy = Z_HUFFMAN_ONLY;
+                break;
+            case 'R':
+                state->strategy = Z_RLE;
+                break;
+            case 'F':
+                state->strategy = Z_FIXED;
+                break;
+            case 'T':
+                state->direct = 1;
+                break;
+            default:        /* could consider as an error, but just ignore */
+                ;
+            }
+        mode++;
+    }
+
+    /* must provide an "r", "w", or "a" */
+    if (state->mode == GZ_NONE) {
+        free(state);
+        return NULL;
+    }
+
+    /* can't force transparent read */
+    if (state->mode == GZ_READ) {
+        if (state->direct) {
+            free(state);
+            return NULL;
+        }
+        state->direct = 1;      /* for empty file */
+    }
+
+    /* save the path name for error messages */
+#ifdef _WIN32
+    if (fd == -2) {
+        len = wcstombs(NULL, path, 0);
+        if (len == (size_t)-1)
+            len = 0;
+    }
+    else
+#endif
+        len = strlen((const char *)path);
+    state->path = (char *)malloc(len + 1);
+    if (state->path == NULL) {
+        free(state);
+        return NULL;
+    }
+#ifdef _WIN32
+    if (fd == -2)
+        if (len)
+            wcstombs(state->path, path, len + 1);
+        else
+            *(state->path) = 0;
+    else
+#endif
+#if !defined(NO_snprintf) && !defined(NO_vsnprintf)
+        snprintf(state->path, len + 1, "%s", (const char *)path);
+#else
+        strcpy(state->path, path);
+#endif
+
+    /* compute the flags for open() */
+    oflag =
+#ifdef O_LARGEFILE
+        O_LARGEFILE |
+#endif
+#ifdef O_BINARY
+        O_BINARY |
+#endif
+#ifdef O_CLOEXEC
+        (cloexec ? O_CLOEXEC : 0) |
+#endif
+        (state->mode == GZ_READ ?
+         O_RDONLY :
+         (O_WRONLY | O_CREAT |
+#ifdef O_EXCL
+          (exclusive ? O_EXCL : 0) |
+#endif
+          (state->mode == GZ_WRITE ?
+           O_TRUNC :
+           O_APPEND)));
+
+    /* open the file with the appropriate flags (or just use fd) */
+    state->fd = fd > -1 ? fd : (
+#ifdef _WIN32
+        fd == -2 ? _wopen(path, oflag, 0666) :
+#endif
+        open((const char *)path, oflag, 0666));
+    if (state->fd == -1) {
+        free(state->path);
+        free(state);
+        return NULL;
+    }
+    if (state->mode == GZ_APPEND)
+        state->mode = GZ_WRITE;         /* simplify later checks */
+
+    /* save the current position for rewinding (only if reading) */
+    if (state->mode == GZ_READ) {
+        state->start = LSEEK(state->fd, 0, SEEK_CUR);
+        if (state->start == -1) state->start = 0;
+    }
+
+    /* initialize stream */
+    gz_reset(state);
+
+    /* return stream */
+    return (gzFile)state;
+}
+
+/* -- see zlib.h -- */
+gzFile ZEXPORT gzopen(path, mode)
+    const char *path;
+    const char *mode;
+{
+    return gz_open(path, -1, mode);
+}
+
+/* -- see zlib.h -- */
+gzFile ZEXPORT gzopen64(path, mode)
+    const char *path;
+    const char *mode;
+{
+    return gz_open(path, -1, mode);
+}
+
+/* -- see zlib.h -- */
+gzFile ZEXPORT gzdopen(fd, mode)
+    int fd;
+    const char *mode;
+{
+    char *path;         /* identifier for error messages */
+    gzFile gz;
+
+    if (fd == -1 || (path = (char *)malloc(7 + 3 * sizeof(int))) == NULL)
+        return NULL;
+#if !defined(NO_snprintf) && !defined(NO_vsnprintf)
+    snprintf(path, 7 + 3 * sizeof(int), "<fd:%d>", fd); /* for debugging */
+#else
+    sprintf(path, "<fd:%d>", fd);   /* for debugging */
+#endif
+    gz = gz_open(path, fd, mode);
+    free(path);
+    return gz;
+}
+
+/* -- see zlib.h -- */
+#ifdef _WIN32
+gzFile ZEXPORT gzopen_w(path, mode)
+    const wchar_t *path;
+    const char *mode;
+{
+    return gz_open(path, -2, mode);
+}
+#endif
+
+/* -- see zlib.h -- */
+int ZEXPORT gzbuffer(file, size)
+    gzFile file;
+    unsigned size;
+{
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return -1;
+
+    /* make sure we haven't already allocated memory */
+    if (state->size != 0)
+        return -1;
+
+    /* check and set requested size */
+    if (size < 2)
+        size = 2;               /* need two bytes to check magic header */
+    state->want = size;
+    return 0;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzrewind(file)
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+
+    /* check that we're reading and that there's no error */
+    if (state->mode != GZ_READ ||
+            (state->err != Z_OK && state->err != Z_BUF_ERROR))
+        return -1;
+
+    /* back up and start over */
+    if (LSEEK(state->fd, state->start, SEEK_SET) == -1)
+        return -1;
+    gz_reset(state);
+    return 0;
+}
+
+/* -- see zlib.h -- */
+z_off64_t ZEXPORT gzseek64(file, offset, whence)
+    gzFile file;
+    z_off64_t offset;
+    int whence;
+{
+    unsigned n;
+    z_off64_t ret;
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return -1;
+
+    /* check that there's no error */
+    if (state->err != Z_OK && state->err != Z_BUF_ERROR)
+        return -1;
+
+    /* can only seek from start or relative to current position */
+    if (whence != SEEK_SET && whence != SEEK_CUR)
+        return -1;
+
+    /* normalize offset to a SEEK_CUR specification */
+    if (whence == SEEK_SET)
+        offset -= state->x.pos;
+    else if (state->seek)
+        offset += state->skip;
+    state->seek = 0;
+
+    /* if within raw area while reading, just go there */
+    if (state->mode == GZ_READ && state->how == COPY &&
+            state->x.pos + offset >= 0) {
+        ret = LSEEK(state->fd, offset - state->x.have, SEEK_CUR);
+        if (ret == -1)
+            return -1;
+        state->x.have = 0;
+        state->eof = 0;
+        state->past = 0;
+        state->seek = 0;
+        gz_error(state, Z_OK, NULL);
+        state->strm.avail_in = 0;
+        state->x.pos += offset;
+        return state->x.pos;
+    }
+
+    /* calculate skip amount, rewinding if needed for back seek when reading */
+    if (offset < 0) {
+        if (state->mode != GZ_READ)         /* writing -- can't go backwards */
+            return -1;
+        offset += state->x.pos;
+        if (offset < 0)                     /* before start of file! */
+            return -1;
+        if (gzrewind(file) == -1)           /* rewind, then skip to offset */
+            return -1;
+    }
+
+    /* if reading, skip what's in output buffer (one less gzgetc() check) */
+    if (state->mode == GZ_READ) {
+        n = GT_OFF(state->x.have) || (z_off64_t)state->x.have > offset ?
+            (unsigned)offset : state->x.have;
+        state->x.have -= n;
+        state->x.next += n;
+        state->x.pos += n;
+        offset -= n;
+    }
+
+    /* request skip (if not zero) */
+    if (offset) {
+        state->seek = 1;
+        state->skip = offset;
+    }
+    return state->x.pos + offset;
+}
+
+/* -- see zlib.h -- */
+z_off_t ZEXPORT gzseek(file, offset, whence)
+    gzFile file;
+    z_off_t offset;
+    int whence;
+{
+    z_off64_t ret;
+
+    ret = gzseek64(file, (z_off64_t)offset, whence);
+    return ret == (z_off_t)ret ? (z_off_t)ret : -1;
+}
+
+/* -- see zlib.h -- */
+z_off64_t ZEXPORT gztell64(file)
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return -1;
+
+    /* return position */
+    return state->x.pos + (state->seek ? state->skip : 0);
+}
+
+/* -- see zlib.h -- */
+z_off_t ZEXPORT gztell(file)
+    gzFile file;
+{
+    z_off64_t ret;
+
+    ret = gztell64(file);
+    return ret == (z_off_t)ret ? (z_off_t)ret : -1;
+}
+
+/* -- see zlib.h -- */
+z_off64_t ZEXPORT gzoffset64(file)
+    gzFile file;
+{
+    z_off64_t offset;
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return -1;
+
+    /* compute and return effective offset in file */
+    offset = LSEEK(state->fd, 0, SEEK_CUR);
+    if (offset == -1)
+        return -1;
+    if (state->mode == GZ_READ)             /* reading */
+        offset -= state->strm.avail_in;     /* don't count buffered input */
+    return offset;
+}
+
+/* -- see zlib.h -- */
+z_off_t ZEXPORT gzoffset(file)
+    gzFile file;
+{
+    z_off64_t ret;
+
+    ret = gzoffset64(file);
+    return ret == (z_off_t)ret ? (z_off_t)ret : -1;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzeof(file)
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return 0;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return 0;
+
+    /* return end-of-file state */
+    return state->mode == GZ_READ ? state->past : 0;
+}
+
+/* -- see zlib.h -- */
+const char * ZEXPORT gzerror(file, errnum)
+    gzFile file;
+    int *errnum;
+{
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return NULL;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return NULL;
+
+    /* return error information */
+    if (errnum != NULL)
+        *errnum = state->err;
+    return state->err == Z_MEM_ERROR ? "out of memory" :
+                                       (state->msg == NULL ? "" : state->msg);
+}
+
+/* -- see zlib.h -- */
+void ZEXPORT gzclearerr(file)
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure and check integrity */
+    if (file == NULL)
+        return;
+    state = (gz_statep)file;
+    if (state->mode != GZ_READ && state->mode != GZ_WRITE)
+        return;
+
+    /* clear error and end-of-file */
+    if (state->mode == GZ_READ) {
+        state->eof = 0;
+        state->past = 0;
+    }
+    gz_error(state, Z_OK, NULL);
+}
+
+/* Create an error message in allocated memory and set state->err and
+   state->msg accordingly.  Free any previous error message already there.  Do
+   not try to free or allocate space if the error is Z_MEM_ERROR (out of
+   memory).  Simply save the error message as a static string.  If there is an
+   allocation failure constructing the error message, then convert the error to
+   out of memory. */
+void ZLIB_INTERNAL gz_error(state, err, msg)
+    gz_statep state;
+    int err;
+    const char *msg;
+{
+    /* free previously allocated message and clear */
+    if (state->msg != NULL) {
+        if (state->err != Z_MEM_ERROR)
+            free(state->msg);
+        state->msg = NULL;
+    }
+
+    /* if fatal, set state->x.have to 0 so that the gzgetc() macro fails */
+    if (err != Z_OK && err != Z_BUF_ERROR)
+        state->x.have = 0;
+
+    /* set error code, and if no message, then done */
+    state->err = err;
+    if (msg == NULL)
+        return;
+
+    /* for an out of memory error, return literal string when requested */
+    if (err == Z_MEM_ERROR)
+        return;
+
+    /* construct error message with path */
+    if ((state->msg = (char *)malloc(strlen(state->path) + strlen(msg) + 3)) ==
+            NULL) {
+        state->err = Z_MEM_ERROR;
+        return;
+    }
+#if !defined(NO_snprintf) && !defined(NO_vsnprintf)
+    snprintf(state->msg, strlen(state->path) + strlen(msg) + 3,
+             "%s%s%s", state->path, ": ", msg);
+#else
+    strcpy(state->msg, state->path);
+    strcat(state->msg, ": ");
+    strcat(state->msg, msg);
+#endif
+    return;
+}
+
+#ifndef INT_MAX
+/* portably return maximum value for an int (when limits.h presumed not
+   available) -- we need to do this to cover cases where 2's complement not
+   used, since C standard permits 1's complement and sign-bit representations,
+   otherwise we could just use ((unsigned)-1) >> 1 */
+unsigned ZLIB_INTERNAL gz_intmax()
+{
+    unsigned p, q;
+
+    p = 1;
+    do {
+        q = p;
+        p <<= 1;
+        p++;
+    } while (p > q);
+    return q >> 1;
+}
+#endif
diff --git a/ml/dlib/dlib/external/zlib/gzread.c b/ml/dlib/dlib/external/zlib/gzread.c
new file mode 100644
index 000000000..bf4538eb2
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/gzread.c
@@ -0,0 +1,594 @@
+/* gzread.c -- zlib functions for reading gzip files
+ * Copyright (C) 2004, 2005, 2010, 2011, 2012, 2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "gzguts.h"
+
+/* Local functions */
+local int gz_load OF((gz_statep, unsigned char *, unsigned, unsigned *));
+local int gz_avail OF((gz_statep));
+local int gz_look OF((gz_statep));
+local int gz_decomp OF((gz_statep));
+local int gz_fetch OF((gz_statep));
+local int gz_skip OF((gz_statep, z_off64_t));
+
+/* Use read() to load a buffer -- return -1 on error, otherwise 0.  Read from
+   state->fd, and update state->eof, state->err, and state->msg as appropriate.
+   This function needs to loop on read(), since read() is not guaranteed to
+   read the number of bytes requested, depending on the type of descriptor. */
+local int gz_load(state, buf, len, have)
+    gz_statep state;
+    unsigned char *buf;
+    unsigned len;
+    unsigned *have;
+{
+    int ret;
+
+    *have = 0;
+    do {
+        ret = read(state->fd, buf + *have, len - *have);
+        if (ret <= 0)
+            break;
+        *have += ret;
+    } while (*have < len);
+    if (ret < 0) {
+        gz_error(state, Z_ERRNO, zstrerror());
+        return -1;
+    }
+    if (ret == 0)
+        state->eof = 1;
+    return 0;
+}
+
+/* Load up input buffer and set eof flag if last data loaded -- return -1 on
+   error, 0 otherwise.  Note that the eof flag is set when the end of the input
+   file is reached, even though there may be unused data in the buffer.  Once
+   that data has been used, no more attempts will be made to read the file.
+   If strm->avail_in != 0, then the current data is moved to the beginning of
+   the input buffer, and then the remainder of the buffer is loaded with the
+   available data from the input file. */
+local int gz_avail(state)
+    gz_statep state;
+{
+    unsigned got;
+    z_streamp strm = &(state->strm);
+
+    if (state->err != Z_OK && state->err != Z_BUF_ERROR)
+        return -1;
+    if (state->eof == 0) {
+        if (strm->avail_in) {       /* copy what's there to the start */
+            unsigned char *p = state->in;
+            unsigned const char *q = strm->next_in;
+            unsigned n = strm->avail_in;
+            do {
+                *p++ = *q++;
+            } while (--n);
+        }
+        if (gz_load(state, state->in + strm->avail_in,
+                    state->size - strm->avail_in, &got) == -1)
+            return -1;
+        strm->avail_in += got;
+        strm->next_in = state->in;
+    }
+    return 0;
+}
+
+/* Look for gzip header, set up for inflate or copy.  state->x.have must be 0.
+   If this is the first time in, allocate required memory.  state->how will be
+   left unchanged if there is no more input data available, will be set to COPY
+   if there is no gzip header and direct copying will be performed, or it will
+   be set to GZIP for decompression.  If direct copying, then leftover input
+   data from the input buffer will be copied to the output buffer.  In that
+   case, all further file reads will be directly to either the output buffer or
+   a user buffer.  If decompressing, the inflate state will be initialized.
+   gz_look() will return 0 on success or -1 on failure. */
+local int gz_look(state)
+    gz_statep state;
+{
+    z_streamp strm = &(state->strm);
+
+    /* allocate read buffers and inflate memory */
+    if (state->size == 0) {
+        /* allocate buffers */
+        state->in = (unsigned char *)malloc(state->want);
+        state->out = (unsigned char *)malloc(state->want << 1);
+        if (state->in == NULL || state->out == NULL) {
+            if (state->out != NULL)
+                free(state->out);
+            if (state->in != NULL)
+                free(state->in);
+            gz_error(state, Z_MEM_ERROR, "out of memory");
+            return -1;
+        }
+        state->size = state->want;
+
+        /* allocate inflate memory */
+        state->strm.zalloc = Z_NULL;
+        state->strm.zfree = Z_NULL;
+        state->strm.opaque = Z_NULL;
+        state->strm.avail_in = 0;
+        state->strm.next_in = Z_NULL;
+        if (inflateInit2(&(state->strm), 15 + 16) != Z_OK) {    /* gunzip */
+            free(state->out);
+            free(state->in);
+            state->size = 0;
+            gz_error(state, Z_MEM_ERROR, "out of memory");
+            return -1;
+        }
+    }
+
+    /* get at least the magic bytes in the input buffer */
+    if (strm->avail_in < 2) {
+        if (gz_avail(state) == -1)
+            return -1;
+        if (strm->avail_in == 0)
+            return 0;
+    }
+
+    /* look for gzip magic bytes -- if there, do gzip decoding (note: there is
+       a logical dilemma here when considering the case of a partially written
+       gzip file, to wit, if a single 31 byte is written, then we cannot tell
+       whether this is a single-byte file, or just a partially written gzip
+       file -- for here we assume that if a gzip file is being written, then
+       the header will be written in a single operation, so that reading a
+       single byte is sufficient indication that it is not a gzip file) */
+    if (strm->avail_in > 1 &&
+            strm->next_in[0] == 31 && strm->next_in[1] == 139) {
+        inflateReset(strm);
+        state->how = GZIP;
+        state->direct = 0;
+        return 0;
+    }
+
+    /* no gzip header -- if we were decoding gzip before, then this is trailing
+       garbage.  Ignore the trailing garbage and finish. */
+    if (state->direct == 0) {
+        strm->avail_in = 0;
+        state->eof = 1;
+        state->x.have = 0;
+        return 0;
+    }
+
+    /* doing raw i/o, copy any leftover input to output -- this assumes that
+       the output buffer is larger than the input buffer, which also assures
+       space for gzungetc() */
+    state->x.next = state->out;
+    if (strm->avail_in) {
+        memcpy(state->x.next, strm->next_in, strm->avail_in);
+        state->x.have = strm->avail_in;
+        strm->avail_in = 0;
+    }
+    state->how = COPY;
+    state->direct = 1;
+    return 0;
+}
+
+/* Decompress from input to the provided next_out and avail_out in the state.
+   On return, state->x.have and state->x.next point to the just decompressed
+   data.  If the gzip stream completes, state->how is reset to LOOK to look for
+   the next gzip stream or raw data, once state->x.have is depleted.  Returns 0
+   on success, -1 on failure. */
+local int gz_decomp(state)
+    gz_statep state;
+{
+    int ret = Z_OK;
+    unsigned had;
+    z_streamp strm = &(state->strm);
+
+    /* fill output buffer up to end of deflate stream */
+    had = strm->avail_out;
+    do {
+        /* get more input for inflate() */
+        if (strm->avail_in == 0 && gz_avail(state) == -1)
+            return -1;
+        if (strm->avail_in == 0) {
+            gz_error(state, Z_BUF_ERROR, "unexpected end of file");
+            break;
+        }
+
+        /* decompress and handle errors */
+        ret = inflate(strm, Z_NO_FLUSH);
+        if (ret == Z_STREAM_ERROR || ret == Z_NEED_DICT) {
+            gz_error(state, Z_STREAM_ERROR,
+                     "internal error: inflate stream corrupt");
+            return -1;
+        }
+        if (ret == Z_MEM_ERROR) {
+            gz_error(state, Z_MEM_ERROR, "out of memory");
+            return -1;
+        }
+        if (ret == Z_DATA_ERROR) {              /* deflate stream invalid */
+            gz_error(state, Z_DATA_ERROR,
+                     strm->msg == NULL ? "compressed data error" : strm->msg);
+            return -1;
+        }
+    } while (strm->avail_out && ret != Z_STREAM_END);
+
+    /* update available output */
+    state->x.have = had - strm->avail_out;
+    state->x.next = strm->next_out - state->x.have;
+
+    /* if the gzip stream completed successfully, look for another */
+    if (ret == Z_STREAM_END)
+        state->how = LOOK;
+
+    /* good decompression */
+    return 0;
+}
+
+/* Fetch data and put it in the output buffer.  Assumes state->x.have is 0.
+   Data is either copied from the input file or decompressed from the input
+   file depending on state->how.  If state->how is LOOK, then a gzip header is
+   looked for to determine whether to copy or decompress.  Returns -1 on error,
+   otherwise 0.  gz_fetch() will leave state->how as COPY or GZIP unless the
+   end of the input file has been reached and all data has been processed.  */
+local int gz_fetch(state)
+    gz_statep state;
+{
+    z_streamp strm = &(state->strm);
+
+    do {
+        switch(state->how) {
+        case LOOK:      /* -> LOOK, COPY (only if never GZIP), or GZIP */
+            if (gz_look(state) == -1)
+                return -1;
+            if (state->how == LOOK)
+                return 0;
+            break;
+        case COPY:      /* -> COPY */
+            if (gz_load(state, state->out, state->size << 1, &(state->x.have))
+                    == -1)
+                return -1;
+            state->x.next = state->out;
+            return 0;
+        case GZIP:      /* -> GZIP or LOOK (if end of gzip stream) */
+            strm->avail_out = state->size << 1;
+            strm->next_out = state->out;
+            if (gz_decomp(state) == -1)
+                return -1;
+        }
+    } while (state->x.have == 0 && (!state->eof || strm->avail_in));
+    return 0;
+}
+
+/* Skip len uncompressed bytes of output.  Return -1 on error, 0 on success. */
+local int gz_skip(state, len)
+    gz_statep state;
+    z_off64_t len;
+{
+    unsigned n;
+
+    /* skip over len bytes or reach end-of-file, whichever comes first */
+    while (len)
+        /* skip over whatever is in output buffer */
+        if (state->x.have) {
+            n = GT_OFF(state->x.have) || (z_off64_t)state->x.have > len ?
+                (unsigned)len : state->x.have;
+            state->x.have -= n;
+            state->x.next += n;
+            state->x.pos += n;
+            len -= n;
+        }
+
+        /* output buffer empty -- return if we're at the end of the input */
+        else if (state->eof && state->strm.avail_in == 0)
+            break;
+
+        /* need more data to skip -- load up output buffer */
+        else {
+            /* get more output, looking for header if required */
+            if (gz_fetch(state) == -1)
+                return -1;
+        }
+    return 0;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzread(file, buf, len)
+    gzFile file;
+    voidp buf;
+    unsigned len;
+{
+    unsigned got, n;
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that we're reading and that there's no (serious) error */
+    if (state->mode != GZ_READ ||
+            (state->err != Z_OK && state->err != Z_BUF_ERROR))
+        return -1;
+
+    /* since an int is returned, make sure len fits in one, otherwise return
+       with an error (this avoids the flaw in the interface) */
+    if ((int)len < 0) {
+        gz_error(state, Z_DATA_ERROR, "requested length does not fit in int");
+        return -1;
+    }
+
+    /* if len is zero, avoid unnecessary operations */
+    if (len == 0)
+        return 0;
+
+    /* process a skip request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_skip(state, state->skip) == -1)
+            return -1;
+    }
+
+    /* get len bytes to buf, or less than len if at the end */
+    got = 0;
+    do {
+        /* first just try copying data from the output buffer */
+        if (state->x.have) {
+            n = state->x.have > len ? len : state->x.have;
+            memcpy(buf, state->x.next, n);
+            state->x.next += n;
+            state->x.have -= n;
+        }
+
+        /* output buffer empty -- return if we're at the end of the input */
+        else if (state->eof && strm->avail_in == 0) {
+            state->past = 1;        /* tried to read past end */
+            break;
+        }
+
+        /* need output data -- for small len or new stream load up our output
+           buffer */
+        else if (state->how == LOOK || len < (state->size << 1)) {
+            /* get more output, looking for header if required */
+            if (gz_fetch(state) == -1)
+                return -1;
+            continue;       /* no progress yet -- go back to copy above */
+            /* the copy above assures that we will leave with space in the
+               output buffer, allowing at least one gzungetc() to succeed */
+        }
+
+        /* large len -- read directly into user buffer */
+        else if (state->how == COPY) {      /* read directly */
+            if (gz_load(state, (unsigned char *)buf, len, &n) == -1)
+                return -1;
+        }
+
+        /* large len -- decompress directly into user buffer */
+        else {  /* state->how == GZIP */
+            strm->avail_out = len;
+            strm->next_out = (unsigned char *)buf;
+            if (gz_decomp(state) == -1)
+                return -1;
+            n = state->x.have;
+            state->x.have = 0;
+        }
+
+        /* update progress */
+        len -= n;
+        buf = (char *)buf + n;
+        got += n;
+        state->x.pos += n;
+    } while (len);
+
+    /* return number of bytes read into user buffer (will fit in int) */
+    return (int)got;
+}
+
+/* -- see zlib.h -- */
+#ifdef Z_PREFIX_SET
+#  undef z_gzgetc
+#else
+#  undef gzgetc
+#endif
+int ZEXPORT gzgetc(file)
+    gzFile file;
+{
+    int ret;
+    unsigned char buf[1];
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+
+    /* check that we're reading and that there's no (serious) error */
+    if (state->mode != GZ_READ ||
+        (state->err != Z_OK && state->err != Z_BUF_ERROR))
+        return -1;
+
+    /* try output buffer (no need to check for skip request) */
+    if (state->x.have) {
+        state->x.have--;
+        state->x.pos++;
+        return *(state->x.next)++;
+    }
+
+    /* nothing there -- try gzread() */
+    ret = gzread(file, buf, 1);
+    return ret < 1 ? -1 : buf[0];
+}
+
+int ZEXPORT gzgetc_(file)
+gzFile file;
+{
+    return gzgetc(file);
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzungetc(c, file)
+    int c;
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+
+    /* check that we're reading and that there's no (serious) error */
+    if (state->mode != GZ_READ ||
+        (state->err != Z_OK && state->err != Z_BUF_ERROR))
+        return -1;
+
+    /* process a skip request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_skip(state, state->skip) == -1)
+            return -1;
+    }
+
+    /* can't push EOF */
+    if (c < 0)
+        return -1;
+
+    /* if output buffer empty, put byte at end (allows more pushing) */
+    if (state->x.have == 0) {
+        state->x.have = 1;
+        state->x.next = state->out + (state->size << 1) - 1;
+        state->x.next[0] = c;
+        state->x.pos--;
+        state->past = 0;
+        return c;
+    }
+
+    /* if no room, give up (must have already done a gzungetc()) */
+    if (state->x.have == (state->size << 1)) {
+        gz_error(state, Z_DATA_ERROR, "out of room to push characters");
+        return -1;
+    }
+
+    /* slide output data if needed and insert byte before existing data */
+    if (state->x.next == state->out) {
+        unsigned char *src = state->out + state->x.have;
+        unsigned char *dest = state->out + (state->size << 1);
+        while (src > state->out)
+            *--dest = *--src;
+        state->x.next = dest;
+    }
+    state->x.have++;
+    state->x.next--;
+    state->x.next[0] = c;
+    state->x.pos--;
+    state->past = 0;
+    return c;
+}
+
+/* -- see zlib.h -- */
+char * ZEXPORT gzgets(file, buf, len)
+    gzFile file;
+    char *buf;
+    int len;
+{
+    unsigned left, n;
+    char *str;
+    unsigned char *eol;
+    gz_statep state;
+
+    /* check parameters and get internal structure */
+    if (file == NULL || buf == NULL || len < 1)
+        return NULL;
+    state = (gz_statep)file;
+
+    /* check that we're reading and that there's no (serious) error */
+    if (state->mode != GZ_READ ||
+        (state->err != Z_OK && state->err != Z_BUF_ERROR))
+        return NULL;
+
+    /* process a skip request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_skip(state, state->skip) == -1)
+            return NULL;
+    }
+
+    /* copy output bytes up to new line or len - 1, whichever comes first --
+       append a terminating zero to the string (we don't check for a zero in
+       the contents, let the user worry about that) */
+    str = buf;
+    left = (unsigned)len - 1;
+    if (left) do {
+        /* assure that something is in the output buffer */
+        if (state->x.have == 0 && gz_fetch(state) == -1)
+            return NULL;                /* error */
+        if (state->x.have == 0) {       /* end of file */
+            state->past = 1;            /* read past end */
+            break;                      /* return what we have */
+        }
+
+        /* look for end-of-line in current output buffer */
+        n = state->x.have > left ? left : state->x.have;
+        eol = (unsigned char *)memchr(state->x.next, '\n', n);
+        if (eol != NULL)
+            n = (unsigned)(eol - state->x.next) + 1;
+
+        /* copy through end-of-line, or remainder if not found */
+        memcpy(buf, state->x.next, n);
+        state->x.have -= n;
+        state->x.next += n;
+        state->x.pos += n;
+        left -= n;
+        buf += n;
+    } while (left && eol == NULL);
+
+    /* return terminated string, or if nothing, end of file */
+    if (buf == str)
+        return NULL;
+    buf[0] = 0;
+    return str;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzdirect(file)
+    gzFile file;
+{
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return 0;
+    state = (gz_statep)file;
+
+    /* if the state is not known, but we can find out, then do so (this is
+       mainly for right after a gzopen() or gzdopen()) */
+    if (state->mode == GZ_READ && state->how == LOOK && state->x.have == 0)
+        (void)gz_look(state);
+
+    /* return 1 if transparent, 0 if processing a gzip stream */
+    return state->direct;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzclose_r(file)
+    gzFile file;
+{
+    int ret, err;
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return Z_STREAM_ERROR;
+    state = (gz_statep)file;
+
+    /* check that we're reading */
+    if (state->mode != GZ_READ)
+        return Z_STREAM_ERROR;
+
+    /* free memory and close file */
+    if (state->size) {
+        inflateEnd(&(state->strm));
+        free(state->out);
+        free(state->in);
+    }
+    err = state->err == Z_BUF_ERROR ? Z_BUF_ERROR : Z_OK;
+    gz_error(state, Z_OK, NULL);
+    free(state->path);
+    ret = close(state->fd);
+    free(state);
+    return ret ? Z_ERRNO : err;
+}
diff --git a/ml/dlib/dlib/external/zlib/gzwrite.c b/ml/dlib/dlib/external/zlib/gzwrite.c
new file mode 100644
index 000000000..aa767fbf6
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/gzwrite.c
@@ -0,0 +1,577 @@
+/* gzwrite.c -- zlib functions for writing gzip files
+ * Copyright (C) 2004, 2005, 2010, 2011, 2012, 2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "gzguts.h"
+
+/* Local functions */
+local int gz_init OF((gz_statep));
+local int gz_comp OF((gz_statep, int));
+local int gz_zero OF((gz_statep, z_off64_t));
+
+/* Initialize state for writing a gzip file.  Mark initialization by setting
+   state->size to non-zero.  Return -1 on failure or 0 on success. */
+local int gz_init(state)
+    gz_statep state;
+{
+    int ret;
+    z_streamp strm = &(state->strm);
+
+    /* allocate input buffer */
+    state->in = (unsigned char *)malloc(state->want);
+    if (state->in == NULL) {
+        gz_error(state, Z_MEM_ERROR, "out of memory");
+        return -1;
+    }
+
+    /* only need output buffer and deflate state if compressing */
+    if (!state->direct) {
+        /* allocate output buffer */
+        state->out = (unsigned char *)malloc(state->want);
+        if (state->out == NULL) {
+            free(state->in);
+            gz_error(state, Z_MEM_ERROR, "out of memory");
+            return -1;
+        }
+
+        /* allocate deflate memory, set up for gzip compression */
+        strm->zalloc = Z_NULL;
+        strm->zfree = Z_NULL;
+        strm->opaque = Z_NULL;
+        ret = deflateInit2(strm, state->level, Z_DEFLATED,
+                           MAX_WBITS + 16, DEF_MEM_LEVEL, state->strategy);
+        if (ret != Z_OK) {
+            free(state->out);
+            free(state->in);
+            gz_error(state, Z_MEM_ERROR, "out of memory");
+            return -1;
+        }
+    }
+
+    /* mark state as initialized */
+    state->size = state->want;
+
+    /* initialize write buffer if compressing */
+    if (!state->direct) {
+        strm->avail_out = state->size;
+        strm->next_out = state->out;
+        state->x.next = strm->next_out;
+    }
+    return 0;
+}
+
+/* Compress whatever is at avail_in and next_in and write to the output file.
+   Return -1 if there is an error writing to the output file, otherwise 0.
+   flush is assumed to be a valid deflate() flush value.  If flush is Z_FINISH,
+   then the deflate() state is reset to start a new gzip stream.  If gz->direct
+   is true, then simply write to the output file without compressing, and
+   ignore flush. */
+local int gz_comp(state, flush)
+    gz_statep state;
+    int flush;
+{
+    int ret, got;
+    unsigned have;
+    z_streamp strm = &(state->strm);
+
+    /* allocate memory if this is the first time through */
+    if (state->size == 0 && gz_init(state) == -1)
+        return -1;
+
+    /* write directly if requested */
+    if (state->direct) {
+        got = write(state->fd, strm->next_in, strm->avail_in);
+        if (got < 0 || (unsigned)got != strm->avail_in) {
+            gz_error(state, Z_ERRNO, zstrerror());
+            return -1;
+        }
+        strm->avail_in = 0;
+        return 0;
+    }
+
+    /* run deflate() on provided input until it produces no more output */
+    ret = Z_OK;
+    do {
+        /* write out current buffer contents if full, or if flushing, but if
+           doing Z_FINISH then don't write until we get to Z_STREAM_END */
+        if (strm->avail_out == 0 || (flush != Z_NO_FLUSH &&
+            (flush != Z_FINISH || ret == Z_STREAM_END))) {
+            have = (unsigned)(strm->next_out - state->x.next);
+            if (have && ((got = write(state->fd, state->x.next, have)) < 0 ||
+                         (unsigned)got != have)) {
+                gz_error(state, Z_ERRNO, zstrerror());
+                return -1;
+            }
+            if (strm->avail_out == 0) {
+                strm->avail_out = state->size;
+                strm->next_out = state->out;
+            }
+            state->x.next = strm->next_out;
+        }
+
+        /* compress */
+        have = strm->avail_out;
+        ret = deflate(strm, flush);
+        if (ret == Z_STREAM_ERROR) {
+            gz_error(state, Z_STREAM_ERROR,
+                      "internal error: deflate stream corrupt");
+            return -1;
+        }
+        have -= strm->avail_out;
+    } while (have);
+
+    /* if that completed a deflate stream, allow another to start */
+    if (flush == Z_FINISH)
+        deflateReset(strm);
+
+    /* all done, no errors */
+    return 0;
+}
+
+/* Compress len zeros to output.  Return -1 on error, 0 on success. */
+local int gz_zero(state, len)
+    gz_statep state;
+    z_off64_t len;
+{
+    int first;
+    unsigned n;
+    z_streamp strm = &(state->strm);
+
+    /* consume whatever's left in the input buffer */
+    if (strm->avail_in && gz_comp(state, Z_NO_FLUSH) == -1)
+        return -1;
+
+    /* compress len zeros (len guaranteed > 0) */
+    first = 1;
+    while (len) {
+        n = GT_OFF(state->size) || (z_off64_t)state->size > len ?
+            (unsigned)len : state->size;
+        if (first) {
+            memset(state->in, 0, n);
+            first = 0;
+        }
+        strm->avail_in = n;
+        strm->next_in = state->in;
+        state->x.pos += n;
+        if (gz_comp(state, Z_NO_FLUSH) == -1)
+            return -1;
+        len -= n;
+    }
+    return 0;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzwrite(file, buf, len)
+    gzFile file;
+    voidpc buf;
+    unsigned len;
+{
+    unsigned put = len;
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return 0;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return 0;
+
+    /* since an int is returned, make sure len fits in one, otherwise return
+       with an error (this avoids the flaw in the interface) */
+    if ((int)len < 0) {
+        gz_error(state, Z_DATA_ERROR, "requested length does not fit in int");
+        return 0;
+    }
+
+    /* if len is zero, avoid unnecessary operations */
+    if (len == 0)
+        return 0;
+
+    /* allocate memory if this is the first time through */
+    if (state->size == 0 && gz_init(state) == -1)
+        return 0;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return 0;
+    }
+
+    /* for small len, copy to input buffer, otherwise compress directly */
+    if (len < state->size) {
+        /* copy to input buffer, compress when full */
+        do {
+            unsigned have, copy;
+
+            if (strm->avail_in == 0)
+                strm->next_in = state->in;
+            have = (unsigned)((strm->next_in + strm->avail_in) - state->in);
+            copy = state->size - have;
+            if (copy > len)
+                copy = len;
+            memcpy(state->in + have, buf, copy);
+            strm->avail_in += copy;
+            state->x.pos += copy;
+            buf = (const char *)buf + copy;
+            len -= copy;
+            if (len && gz_comp(state, Z_NO_FLUSH) == -1)
+                return 0;
+        } while (len);
+    }
+    else {
+        /* consume whatever's left in the input buffer */
+        if (strm->avail_in && gz_comp(state, Z_NO_FLUSH) == -1)
+            return 0;
+
+        /* directly compress user buffer to file */
+        strm->avail_in = len;
+        strm->next_in = (z_const Bytef *)buf;
+        state->x.pos += len;
+        if (gz_comp(state, Z_NO_FLUSH) == -1)
+            return 0;
+    }
+
+    /* input was all buffered or compressed (put will fit in int) */
+    return (int)put;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzputc(file, c)
+    gzFile file;
+    int c;
+{
+    unsigned have;
+    unsigned char buf[1];
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return -1;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return -1;
+    }
+
+    /* try writing to input buffer for speed (state->size == 0 if buffer not
+       initialized) */
+    if (state->size) {
+        if (strm->avail_in == 0)
+            strm->next_in = state->in;
+        have = (unsigned)((strm->next_in + strm->avail_in) - state->in);
+        if (have < state->size) {
+            state->in[have] = c;
+            strm->avail_in++;
+            state->x.pos++;
+            return c & 0xff;
+        }
+    }
+
+    /* no room in buffer or not initialized, use gz_write() */
+    buf[0] = c;
+    if (gzwrite(file, buf, 1) != 1)
+        return -1;
+    return c & 0xff;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzputs(file, str)
+    gzFile file;
+    const char *str;
+{
+    int ret;
+    unsigned len;
+
+    /* write string */
+    len = (unsigned)strlen(str);
+    ret = gzwrite(file, str, len);
+    return ret == 0 && len != 0 ? -1 : ret;
+}
+
+#if defined(STDC) || defined(Z_HAVE_STDARG_H)
+#include <stdarg.h>
+
+/* -- see zlib.h -- */
+int ZEXPORTVA gzvprintf(gzFile file, const char *format, va_list va)
+{
+    int size, len;
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return 0;
+
+    /* make sure we have some buffer space */
+    if (state->size == 0 && gz_init(state) == -1)
+        return 0;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return 0;
+    }
+
+    /* consume whatever's left in the input buffer */
+    if (strm->avail_in && gz_comp(state, Z_NO_FLUSH) == -1)
+        return 0;
+
+    /* do the printf() into the input buffer, put length in len */
+    size = (int)(state->size);
+    state->in[size - 1] = 0;
+#ifdef NO_vsnprintf
+#  ifdef HAS_vsprintf_void
+    (void)vsprintf((char *)(state->in), format, va);
+    for (len = 0; len < size; len++)
+        if (state->in[len] == 0) break;
+#  else
+    len = vsprintf((char *)(state->in), format, va);
+#  endif
+#else
+#  ifdef HAS_vsnprintf_void
+    (void)vsnprintf((char *)(state->in), size, format, va);
+    len = strlen((char *)(state->in));
+#  else
+    len = vsnprintf((char *)(state->in), size, format, va);
+#  endif
+#endif
+
+    /* check that printf() results fit in buffer */
+    if (len <= 0 || len >= (int)size || state->in[size - 1] != 0)
+        return 0;
+
+    /* update buffer and position, defer compression until needed */
+    strm->avail_in = (unsigned)len;
+    strm->next_in = state->in;
+    state->x.pos += len;
+    return len;
+}
+
+int ZEXPORTVA gzprintf(gzFile file, const char *format, ...)
+{
+    va_list va;
+    int ret;
+
+    va_start(va, format);
+    ret = gzvprintf(file, format, va);
+    va_end(va);
+    return ret;
+}
+
+#else /* !STDC && !Z_HAVE_STDARG_H */
+
+/* -- see zlib.h -- */
+int ZEXPORTVA gzprintf (file, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
+                       a11, a12, a13, a14, a15, a16, a17, a18, a19, a20)
+    gzFile file;
+    const char *format;
+    int a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
+        a11, a12, a13, a14, a15, a16, a17, a18, a19, a20;
+{
+    int size, len;
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that can really pass pointer in ints */
+    if (sizeof(int) != sizeof(void *))
+        return 0;
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return 0;
+
+    /* make sure we have some buffer space */
+    if (state->size == 0 && gz_init(state) == -1)
+        return 0;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return 0;
+    }
+
+    /* consume whatever's left in the input buffer */
+    if (strm->avail_in && gz_comp(state, Z_NO_FLUSH) == -1)
+        return 0;
+
+    /* do the printf() into the input buffer, put length in len */
+    size = (int)(state->size);
+    state->in[size - 1] = 0;
+#ifdef NO_snprintf
+#  ifdef HAS_sprintf_void
+    sprintf((char *)(state->in), format, a1, a2, a3, a4, a5, a6, a7, a8,
+            a9, a10, a11, a12, a13, a14, a15, a16, a17, a18, a19, a20);
+    for (len = 0; len < size; len++)
+        if (state->in[len] == 0) break;
+#  else
+    len = sprintf((char *)(state->in), format, a1, a2, a3, a4, a5, a6, a7, a8,
+                  a9, a10, a11, a12, a13, a14, a15, a16, a17, a18, a19, a20);
+#  endif
+#else
+#  ifdef HAS_snprintf_void
+    snprintf((char *)(state->in), size, format, a1, a2, a3, a4, a5, a6, a7, a8,
+             a9, a10, a11, a12, a13, a14, a15, a16, a17, a18, a19, a20);
+    len = strlen((char *)(state->in));
+#  else
+    len = snprintf((char *)(state->in), size, format, a1, a2, a3, a4, a5, a6,
+                   a7, a8, a9, a10, a11, a12, a13, a14, a15, a16, a17, a18,
+                   a19, a20);
+#  endif
+#endif
+
+    /* check that printf() results fit in buffer */
+    if (len <= 0 || len >= (int)size || state->in[size - 1] != 0)
+        return 0;
+
+    /* update buffer and position, defer compression until needed */
+    strm->avail_in = (unsigned)len;
+    strm->next_in = state->in;
+    state->x.pos += len;
+    return len;
+}
+
+#endif
+
+/* -- see zlib.h -- */
+int ZEXPORT gzflush(file, flush)
+    gzFile file;
+    int flush;
+{
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return -1;
+    state = (gz_statep)file;
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return Z_STREAM_ERROR;
+
+    /* check flush parameter */
+    if (flush < 0 || flush > Z_FINISH)
+        return Z_STREAM_ERROR;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return -1;
+    }
+
+    /* compress remaining data with requested flush */
+    gz_comp(state, flush);
+    return state->err;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzsetparams(file, level, strategy)
+    gzFile file;
+    int level;
+    int strategy;
+{
+    gz_statep state;
+    z_streamp strm;
+
+    /* get internal structure */
+    if (file == NULL)
+        return Z_STREAM_ERROR;
+    state = (gz_statep)file;
+    strm = &(state->strm);
+
+    /* check that we're writing and that there's no error */
+    if (state->mode != GZ_WRITE || state->err != Z_OK)
+        return Z_STREAM_ERROR;
+
+    /* if no change is requested, then do nothing */
+    if (level == state->level && strategy == state->strategy)
+        return Z_OK;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            return -1;
+    }
+
+    /* change compression parameters for subsequent input */
+    if (state->size) {
+        /* flush previous input with previous parameters before changing */
+        if (strm->avail_in && gz_comp(state, Z_PARTIAL_FLUSH) == -1)
+            return state->err;
+        deflateParams(strm, level, strategy);
+    }
+    state->level = level;
+    state->strategy = strategy;
+    return Z_OK;
+}
+
+/* -- see zlib.h -- */
+int ZEXPORT gzclose_w(file)
+    gzFile file;
+{
+    int ret = Z_OK;
+    gz_statep state;
+
+    /* get internal structure */
+    if (file == NULL)
+        return Z_STREAM_ERROR;
+    state = (gz_statep)file;
+
+    /* check that we're writing */
+    if (state->mode != GZ_WRITE)
+        return Z_STREAM_ERROR;
+
+    /* check for seek request */
+    if (state->seek) {
+        state->seek = 0;
+        if (gz_zero(state, state->skip) == -1)
+            ret = state->err;
+    }
+
+    /* flush, free memory, and close file */
+    if (gz_comp(state, Z_FINISH) == -1)
+        ret = state->err;
+    if (state->size) {
+        if (!state->direct) {
+            (void)deflateEnd(&(state->strm));
+            free(state->out);
+        }
+        free(state->in);
+    }
+    gz_error(state, Z_OK, NULL);
+    free(state->path);
+    if (close(state->fd) == -1)
+        ret = Z_ERRNO;
+    free(state);
+    return ret;
+}
diff --git a/ml/dlib/dlib/external/zlib/infback.c b/ml/dlib/dlib/external/zlib/infback.c
new file mode 100644
index 000000000..f3833c2e4
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/infback.c
@@ -0,0 +1,640 @@
+/* infback.c -- inflate using a call-back interface
+ * Copyright (C) 1995-2011 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/*
+   This code is largely copied from inflate.c.  Normally either infback.o or
+   inflate.o would be linked into an application--not both.  The interface
+   with inffast.c is retained so that optimized assembler-coded versions of
+   inflate_fast() can be used with either inflate.c or infback.c.
+ */
+
+#include "zutil.h"
+#include "inftrees.h"
+#include "inflate.h"
+#include "inffast.h"
+
+/* function prototypes */
+local void fixedtables OF((struct inflate_state FAR *state));
+
+/*
+   strm provides memory allocation functions in zalloc and zfree, or
+   Z_NULL to use the library memory allocation functions.
+
+   windowBits is in the range 8..15, and window is a user-supplied
+   window and output buffer that is 2**windowBits bytes.
+ */
+int ZEXPORT inflateBackInit_(strm, windowBits, window, version, stream_size)
+z_streamp strm;
+int windowBits;
+unsigned char FAR *window;
+const char *version;
+int stream_size;
+{
+    struct inflate_state FAR *state;
+
+    if (version == Z_NULL || version[0] != ZLIB_VERSION[0] ||
+        stream_size != (int)(sizeof(z_stream)))
+        return Z_VERSION_ERROR;
+    if (strm == Z_NULL || window == Z_NULL ||
+        windowBits < 8 || windowBits > 15)
+        return Z_STREAM_ERROR;
+    strm->msg = Z_NULL;                 /* in case we return an error */
+    if (strm->zalloc == (alloc_func)0) {
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+        strm->zalloc = zcalloc;
+        strm->opaque = (voidpf)0;
+#endif
+    }
+    if (strm->zfree == (free_func)0)
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+    strm->zfree = zcfree;
+#endif
+    state = (struct inflate_state FAR *)ZALLOC(strm, 1,
+                                               sizeof(struct inflate_state));
+    if (state == Z_NULL) return Z_MEM_ERROR;
+    Tracev((stderr, "inflate: allocated\n"));
+    strm->state = (struct internal_state FAR *)state;
+    state->dmax = 32768U;
+    state->wbits = windowBits;
+    state->wsize = 1U << windowBits;
+    state->window = window;
+    state->wnext = 0;
+    state->whave = 0;
+    return Z_OK;
+}
+
+/*
+   Return state with length and distance decoding tables and index sizes set to
+   fixed code decoding.  Normally this returns fixed tables from inffixed.h.
+   If BUILDFIXED is defined, then instead this routine builds the tables the
+   first time it's called, and returns those tables the first time and
+   thereafter.  This reduces the size of the code by about 2K bytes, in
+   exchange for a little execution time.  However, BUILDFIXED should not be
+   used for threaded applications, since the rewriting of the tables and virgin
+   may not be thread-safe.
+ */
+local void fixedtables(state)
+struct inflate_state FAR *state;
+{
+#ifdef BUILDFIXED
+    static int virgin = 1;
+    static code *lenfix, *distfix;
+    static code fixed[544];
+
+    /* build fixed huffman tables if first call (may not be thread safe) */
+    if (virgin) {
+        unsigned sym, bits;
+        static code *next;
+
+        /* literal/length table */
+        sym = 0;
+        while (sym < 144) state->lens[sym++] = 8;
+        while (sym < 256) state->lens[sym++] = 9;
+        while (sym < 280) state->lens[sym++] = 7;
+        while (sym < 288) state->lens[sym++] = 8;
+        next = fixed;
+        lenfix = next;
+        bits = 9;
+        inflate_table(LENS, state->lens, 288, &(next), &(bits), state->work);
+
+        /* distance table */
+        sym = 0;
+        while (sym < 32) state->lens[sym++] = 5;
+        distfix = next;
+        bits = 5;
+        inflate_table(DISTS, state->lens, 32, &(next), &(bits), state->work);
+
+        /* do this just once */
+        virgin = 0;
+    }
+#else /* !BUILDFIXED */
+#   include "inffixed.h"
+#endif /* BUILDFIXED */
+    state->lencode = lenfix;
+    state->lenbits = 9;
+    state->distcode = distfix;
+    state->distbits = 5;
+}
+
+/* Macros for inflateBack(): */
+
+/* Load returned state from inflate_fast() */
+#define LOAD() \
+    do { \
+        put = strm->next_out; \
+        left = strm->avail_out; \
+        next = strm->next_in; \
+        have = strm->avail_in; \
+        hold = state->hold; \
+        bits = state->bits; \
+    } while (0)
+
+/* Set state from registers for inflate_fast() */
+#define RESTORE() \
+    do { \
+        strm->next_out = put; \
+        strm->avail_out = left; \
+        strm->next_in = next; \
+        strm->avail_in = have; \
+        state->hold = hold; \
+        state->bits = bits; \
+    } while (0)
+
+/* Clear the input bit accumulator */
+#define INITBITS() \
+    do { \
+        hold = 0; \
+        bits = 0; \
+    } while (0)
+
+/* Assure that some input is available.  If input is requested, but denied,
+   then return a Z_BUF_ERROR from inflateBack(). */
+#define PULL() \
+    do { \
+        if (have == 0) { \
+            have = in(in_desc, &next); \
+            if (have == 0) { \
+                next = Z_NULL; \
+                ret = Z_BUF_ERROR; \
+                goto inf_leave; \
+            } \
+        } \
+    } while (0)
+
+/* Get a byte of input into the bit accumulator, or return from inflateBack()
+   with an error if there is no input available. */
+#define PULLBYTE() \
+    do { \
+        PULL(); \
+        have--; \
+        hold += (unsigned long)(*next++) << bits; \
+        bits += 8; \
+    } while (0)
+
+/* Assure that there are at least n bits in the bit accumulator.  If there is
+   not enough available input to do that, then return from inflateBack() with
+   an error. */
+#define NEEDBITS(n) \
+    do { \
+        while (bits < (unsigned)(n)) \
+            PULLBYTE(); \
+    } while (0)
+
+/* Return the low n bits of the bit accumulator (n < 16) */
+#define BITS(n) \
+    ((unsigned)hold & ((1U << (n)) - 1))
+
+/* Remove n bits from the bit accumulator */
+#define DROPBITS(n) \
+    do { \
+        hold >>= (n); \
+        bits -= (unsigned)(n); \
+    } while (0)
+
+/* Remove zero to seven bits as needed to go to a byte boundary */
+#define BYTEBITS() \
+    do { \
+        hold >>= bits & 7; \
+        bits -= bits & 7; \
+    } while (0)
+
+/* Assure that some output space is available, by writing out the window
+   if it's full.  If the write fails, return from inflateBack() with a
+   Z_BUF_ERROR. */
+#define ROOM() \
+    do { \
+        if (left == 0) { \
+            put = state->window; \
+            left = state->wsize; \
+            state->whave = left; \
+            if (out(out_desc, put, left)) { \
+                ret = Z_BUF_ERROR; \
+                goto inf_leave; \
+            } \
+        } \
+    } while (0)
+
+/*
+   strm provides the memory allocation functions and window buffer on input,
+   and provides information on the unused input on return.  For Z_DATA_ERROR
+   returns, strm will also provide an error message.
+
+   in() and out() are the call-back input and output functions.  When
+   inflateBack() needs more input, it calls in().  When inflateBack() has
+   filled the window with output, or when it completes with data in the
+   window, it calls out() to write out the data.  The application must not
+   change the provided input until in() is called again or inflateBack()
+   returns.  The application must not change the window/output buffer until
+   inflateBack() returns.
+
+   in() and out() are called with a descriptor parameter provided in the
+   inflateBack() call.  This parameter can be a structure that provides the
+   information required to do the read or write, as well as accumulated
+   information on the input and output such as totals and check values.
+
+   in() should return zero on failure.  out() should return non-zero on
+   failure.  If either in() or out() fails, than inflateBack() returns a
+   Z_BUF_ERROR.  strm->next_in can be checked for Z_NULL to see whether it
+   was in() or out() that caused in the error.  Otherwise,  inflateBack()
+   returns Z_STREAM_END on success, Z_DATA_ERROR for an deflate format
+   error, or Z_MEM_ERROR if it could not allocate memory for the state.
+   inflateBack() can also return Z_STREAM_ERROR if the input parameters
+   are not correct, i.e. strm is Z_NULL or the state was not initialized.
+ */
+int ZEXPORT inflateBack(strm, in, in_desc, out, out_desc)
+z_streamp strm;
+in_func in;
+void FAR *in_desc;
+out_func out;
+void FAR *out_desc;
+{
+    struct inflate_state FAR *state;
+    z_const unsigned char FAR *next;    /* next input */
+    unsigned char FAR *put;     /* next output */
+    unsigned have, left;        /* available input and output */
+    unsigned long hold;         /* bit buffer */
+    unsigned bits;              /* bits in bit buffer */
+    unsigned copy;              /* number of stored or match bytes to copy */
+    unsigned char FAR *from;    /* where to copy match bytes from */
+    code here;                  /* current decoding table entry */
+    code last;                  /* parent table entry */
+    unsigned len;               /* length to copy for repeats, bits to drop */
+    int ret;                    /* return code */
+    static const unsigned short order[19] = /* permutation of code lengths */
+        {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15};
+
+    /* Check that the strm exists and that the state was initialized */
+    if (strm == Z_NULL || strm->state == Z_NULL)
+        return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+
+    /* Reset the state */
+    strm->msg = Z_NULL;
+    state->mode = TYPE;
+    state->last = 0;
+    state->whave = 0;
+    next = strm->next_in;
+    have = next != Z_NULL ? strm->avail_in : 0;
+    hold = 0;
+    bits = 0;
+    put = state->window;
+    left = state->wsize;
+
+    /* Inflate until end of block marked as last */
+    for (;;)
+        switch (state->mode) {
+        case TYPE:
+            /* determine and dispatch block type */
+            if (state->last) {
+                BYTEBITS();
+                state->mode = DONE;
+                break;
+            }
+            NEEDBITS(3);
+            state->last = BITS(1);
+            DROPBITS(1);
+            switch (BITS(2)) {
+            case 0:                             /* stored block */
+                Tracev((stderr, "inflate:     stored block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = STORED;
+                break;
+            case 1:                             /* fixed block */
+                fixedtables(state);
+                Tracev((stderr, "inflate:     fixed codes block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = LEN;              /* decode codes */
+                break;
+            case 2:                             /* dynamic block */
+                Tracev((stderr, "inflate:     dynamic codes block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = TABLE;
+                break;
+            case 3:
+                strm->msg = (char *)"invalid block type";
+                state->mode = BAD;
+            }
+            DROPBITS(2);
+            break;
+
+        case STORED:
+            /* get and verify stored block length */
+            BYTEBITS();                         /* go to byte boundary */
+            NEEDBITS(32);
+            if ((hold & 0xffff) != ((hold >> 16) ^ 0xffff)) {
+                strm->msg = (char *)"invalid stored block lengths";
+                state->mode = BAD;
+                break;
+            }
+            state->length = (unsigned)hold & 0xffff;
+            Tracev((stderr, "inflate:       stored length %u\n",
+                    state->length));
+            INITBITS();
+
+            /* copy stored block from input to output */
+            while (state->length != 0) {
+                copy = state->length;
+                PULL();
+                ROOM();
+                if (copy > have) copy = have;
+                if (copy > left) copy = left;
+                zmemcpy(put, next, copy);
+                have -= copy;
+                next += copy;
+                left -= copy;
+                put += copy;
+                state->length -= copy;
+            }
+            Tracev((stderr, "inflate:       stored end\n"));
+            state->mode = TYPE;
+            break;
+
+        case TABLE:
+            /* get dynamic table entries descriptor */
+            NEEDBITS(14);
+            state->nlen = BITS(5) + 257;
+            DROPBITS(5);
+            state->ndist = BITS(5) + 1;
+            DROPBITS(5);
+            state->ncode = BITS(4) + 4;
+            DROPBITS(4);
+#ifndef PKZIP_BUG_WORKAROUND
+            if (state->nlen > 286 || state->ndist > 30) {
+                strm->msg = (char *)"too many length or distance symbols";
+                state->mode = BAD;
+                break;
+            }
+#endif
+            Tracev((stderr, "inflate:       table sizes ok\n"));
+
+            /* get code length code lengths (not a typo) */
+            state->have = 0;
+            while (state->have < state->ncode) {
+                NEEDBITS(3);
+                state->lens[order[state->have++]] = (unsigned short)BITS(3);
+                DROPBITS(3);
+            }
+            while (state->have < 19)
+                state->lens[order[state->have++]] = 0;
+            state->next = state->codes;
+            state->lencode = (code const FAR *)(state->next);
+            state->lenbits = 7;
+            ret = inflate_table(CODES, state->lens, 19, &(state->next),
+                                &(state->lenbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid code lengths set";
+                state->mode = BAD;
+                break;
+            }
+            Tracev((stderr, "inflate:       code lengths ok\n"));
+
+            /* get length and distance code code lengths */
+            state->have = 0;
+            while (state->have < state->nlen + state->ndist) {
+                for (;;) {
+                    here = state->lencode[BITS(state->lenbits)];
+                    if ((unsigned)(here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                if (here.val < 16) {
+                    DROPBITS(here.bits);
+                    state->lens[state->have++] = here.val;
+                }
+                else {
+                    if (here.val == 16) {
+                        NEEDBITS(here.bits + 2);
+                        DROPBITS(here.bits);
+                        if (state->have == 0) {
+                            strm->msg = (char *)"invalid bit length repeat";
+                            state->mode = BAD;
+                            break;
+                        }
+                        len = (unsigned)(state->lens[state->have - 1]);
+                        copy = 3 + BITS(2);
+                        DROPBITS(2);
+                    }
+                    else if (here.val == 17) {
+                        NEEDBITS(here.bits + 3);
+                        DROPBITS(here.bits);
+                        len = 0;
+                        copy = 3 + BITS(3);
+                        DROPBITS(3);
+                    }
+                    else {
+                        NEEDBITS(here.bits + 7);
+                        DROPBITS(here.bits);
+                        len = 0;
+                        copy = 11 + BITS(7);
+                        DROPBITS(7);
+                    }
+                    if (state->have + copy > state->nlen + state->ndist) {
+                        strm->msg = (char *)"invalid bit length repeat";
+                        state->mode = BAD;
+                        break;
+                    }
+                    while (copy--)
+                        state->lens[state->have++] = (unsigned short)len;
+                }
+            }
+
+            /* handle error breaks in while */
+            if (state->mode == BAD) break;
+
+            /* check for end-of-block code (better have one) */
+            if (state->lens[256] == 0) {
+                strm->msg = (char *)"invalid code -- missing end-of-block";
+                state->mode = BAD;
+                break;
+            }
+
+            /* build code tables -- note: do not change the lenbits or distbits
+               values here (9 and 6) without reading the comments in inftrees.h
+               concerning the ENOUGH constants, which depend on those values */
+            state->next = state->codes;
+            state->lencode = (code const FAR *)(state->next);
+            state->lenbits = 9;
+            ret = inflate_table(LENS, state->lens, state->nlen, &(state->next),
+                                &(state->lenbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid literal/lengths set";
+                state->mode = BAD;
+                break;
+            }
+            state->distcode = (code const FAR *)(state->next);
+            state->distbits = 6;
+            ret = inflate_table(DISTS, state->lens + state->nlen, state->ndist,
+                            &(state->next), &(state->distbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid distances set";
+                state->mode = BAD;
+                break;
+            }
+            Tracev((stderr, "inflate:       codes ok\n"));
+            state->mode = LEN;
+
+        case LEN:
+            /* use inflate_fast() if we have enough input and output */
+            if (have >= 6 && left >= 258) {
+                RESTORE();
+                if (state->whave < state->wsize)
+                    state->whave = state->wsize - left;
+                inflate_fast(strm, state->wsize);
+                LOAD();
+                break;
+            }
+
+            /* get a literal, length, or end-of-block code */
+            for (;;) {
+                here = state->lencode[BITS(state->lenbits)];
+                if ((unsigned)(here.bits) <= bits) break;
+                PULLBYTE();
+            }
+            if (here.op && (here.op & 0xf0) == 0) {
+                last = here;
+                for (;;) {
+                    here = state->lencode[last.val +
+                            (BITS(last.bits + last.op) >> last.bits)];
+                    if ((unsigned)(last.bits + here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                DROPBITS(last.bits);
+            }
+            DROPBITS(here.bits);
+            state->length = (unsigned)here.val;
+
+            /* process literal */
+            if (here.op == 0) {
+                Tracevv((stderr, here.val >= 0x20 && here.val < 0x7f ?
+                        "inflate:         literal '%c'\n" :
+                        "inflate:         literal 0x%02x\n", here.val));
+                ROOM();
+                *put++ = (unsigned char)(state->length);
+                left--;
+                state->mode = LEN;
+                break;
+            }
+
+            /* process end of block */
+            if (here.op & 32) {
+                Tracevv((stderr, "inflate:         end of block\n"));
+                state->mode = TYPE;
+                break;
+            }
+
+            /* invalid code */
+            if (here.op & 64) {
+                strm->msg = (char *)"invalid literal/length code";
+                state->mode = BAD;
+                break;
+            }
+
+            /* length code -- get extra bits, if any */
+            state->extra = (unsigned)(here.op) & 15;
+            if (state->extra != 0) {
+                NEEDBITS(state->extra);
+                state->length += BITS(state->extra);
+                DROPBITS(state->extra);
+            }
+            Tracevv((stderr, "inflate:         length %u\n", state->length));
+
+            /* get distance code */
+            for (;;) {
+                here = state->distcode[BITS(state->distbits)];
+                if ((unsigned)(here.bits) <= bits) break;
+                PULLBYTE();
+            }
+            if ((here.op & 0xf0) == 0) {
+                last = here;
+                for (;;) {
+                    here = state->distcode[last.val +
+                            (BITS(last.bits + last.op) >> last.bits)];
+                    if ((unsigned)(last.bits + here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                DROPBITS(last.bits);
+            }
+            DROPBITS(here.bits);
+            if (here.op & 64) {
+                strm->msg = (char *)"invalid distance code";
+                state->mode = BAD;
+                break;
+            }
+            state->offset = (unsigned)here.val;
+
+            /* get distance extra bits, if any */
+            state->extra = (unsigned)(here.op) & 15;
+            if (state->extra != 0) {
+                NEEDBITS(state->extra);
+                state->offset += BITS(state->extra);
+                DROPBITS(state->extra);
+            }
+            if (state->offset > state->wsize - (state->whave < state->wsize ?
+                                                left : 0)) {
+                strm->msg = (char *)"invalid distance too far back";
+                state->mode = BAD;
+                break;
+            }
+            Tracevv((stderr, "inflate:         distance %u\n", state->offset));
+
+            /* copy match from window to output */
+            do {
+                ROOM();
+                copy = state->wsize - state->offset;
+                if (copy < left) {
+                    from = put + copy;
+                    copy = left - copy;
+                }
+                else {
+                    from = put - state->offset;
+                    copy = left;
+                }
+                if (copy > state->length) copy = state->length;
+                state->length -= copy;
+                left -= copy;
+                do {
+                    *put++ = *from++;
+                } while (--copy);
+            } while (state->length != 0);
+            break;
+
+        case DONE:
+            /* inflate stream terminated properly -- write leftover output */
+            ret = Z_STREAM_END;
+            if (left < state->wsize) {
+                if (out(out_desc, state->window, state->wsize - left))
+                    ret = Z_BUF_ERROR;
+            }
+            goto inf_leave;
+
+        case BAD:
+            ret = Z_DATA_ERROR;
+            goto inf_leave;
+
+        default:                /* can't happen, but makes compilers happy */
+            ret = Z_STREAM_ERROR;
+            goto inf_leave;
+        }
+
+    /* Return unused input */
+  inf_leave:
+    strm->next_in = next;
+    strm->avail_in = have;
+    return ret;
+}
+
+int ZEXPORT inflateBackEnd(strm)
+z_streamp strm;
+{
+    if (strm == Z_NULL || strm->state == Z_NULL || strm->zfree == (free_func)0)
+        return Z_STREAM_ERROR;
+    ZFREE(strm, strm->state);
+    strm->state = Z_NULL;
+    Tracev((stderr, "inflate: end\n"));
+    return Z_OK;
+}
diff --git a/ml/dlib/dlib/external/zlib/inffast.c b/ml/dlib/dlib/external/zlib/inffast.c
new file mode 100644
index 000000000..bda59ceb6
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inffast.c
@@ -0,0 +1,340 @@
+/* inffast.c -- fast decoding
+ * Copyright (C) 1995-2008, 2010, 2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "zutil.h"
+#include "inftrees.h"
+#include "inflate.h"
+#include "inffast.h"
+
+#ifndef ASMINF
+
+/* Allow machine dependent optimization for post-increment or pre-increment.
+   Based on testing to date,
+   Pre-increment preferred for:
+   - PowerPC G3 (Adler)
+   - MIPS R5000 (Randers-Pehrson)
+   Post-increment preferred for:
+   - none
+   No measurable difference:
+   - Pentium III (Anderson)
+   - M68060 (Nikl)
+ */
+#ifdef POSTINC
+#  define OFF 0
+#  define PUP(a) *(a)++
+#else
+#  define OFF 1
+#  define PUP(a) *++(a)
+#endif
+
+/*
+   Decode literal, length, and distance codes and write out the resulting
+   literal and match bytes until either not enough input or output is
+   available, an end-of-block is encountered, or a data error is encountered.
+   When large enough input and output buffers are supplied to inflate(), for
+   example, a 16K input buffer and a 64K output buffer, more than 95% of the
+   inflate execution time is spent in this routine.
+
+   Entry assumptions:
+
+        state->mode == LEN
+        strm->avail_in >= 6
+        strm->avail_out >= 258
+        start >= strm->avail_out
+        state->bits < 8
+
+   On return, state->mode is one of:
+
+        LEN -- ran out of enough output space or enough available input
+        TYPE -- reached end of block code, inflate() to interpret next block
+        BAD -- error in block data
+
+   Notes:
+
+    - The maximum input bits used by a length/distance pair is 15 bits for the
+      length code, 5 bits for the length extra, 15 bits for the distance code,
+      and 13 bits for the distance extra.  This totals 48 bits, or six bytes.
+      Therefore if strm->avail_in >= 6, then there is enough input to avoid
+      checking for available input while decoding.
+
+    - The maximum bytes that a single length/distance pair can output is 258
+      bytes, which is the maximum length that can be coded.  inflate_fast()
+      requires strm->avail_out >= 258 for each loop to avoid checking for
+      output space.
+ */
+void ZLIB_INTERNAL inflate_fast(strm, start)
+z_streamp strm;
+unsigned start;         /* inflate()'s starting value for strm->avail_out */
+{
+    struct inflate_state FAR *state;
+    z_const unsigned char FAR *in;      /* local strm->next_in */
+    z_const unsigned char FAR *last;    /* have enough input while in < last */
+    unsigned char FAR *out;     /* local strm->next_out */
+    unsigned char FAR *beg;     /* inflate()'s initial strm->next_out */
+    unsigned char FAR *end;     /* while out < end, enough space available */
+#ifdef INFLATE_STRICT
+    unsigned dmax;              /* maximum distance from zlib header */
+#endif
+    unsigned wsize;             /* window size or zero if not using window */
+    unsigned whave;             /* valid bytes in the window */
+    unsigned wnext;             /* window write index */
+    unsigned char FAR *window;  /* allocated sliding window, if wsize != 0 */
+    unsigned long hold;         /* local strm->hold */
+    unsigned bits;              /* local strm->bits */
+    code const FAR *lcode;      /* local strm->lencode */
+    code const FAR *dcode;      /* local strm->distcode */
+    unsigned lmask;             /* mask for first level of length codes */
+    unsigned dmask;             /* mask for first level of distance codes */
+    code here;                  /* retrieved table entry */
+    unsigned op;                /* code bits, operation, extra bits, or */
+                                /*  window position, window bytes to copy */
+    unsigned len;               /* match length, unused bytes */
+    unsigned dist;              /* match distance */
+    unsigned char FAR *from;    /* where to copy match from */
+
+    /* copy state to local variables */
+    state = (struct inflate_state FAR *)strm->state;
+    in = strm->next_in - OFF;
+    last = in + (strm->avail_in - 5);
+    out = strm->next_out - OFF;
+    beg = out - (start - strm->avail_out);
+    end = out + (strm->avail_out - 257);
+#ifdef INFLATE_STRICT
+    dmax = state->dmax;
+#endif
+    wsize = state->wsize;
+    whave = state->whave;
+    wnext = state->wnext;
+    window = state->window;
+    hold = state->hold;
+    bits = state->bits;
+    lcode = state->lencode;
+    dcode = state->distcode;
+    lmask = (1U << state->lenbits) - 1;
+    dmask = (1U << state->distbits) - 1;
+
+    /* decode literals and length/distances until end-of-block or not enough
+       input data or output space */
+    do {
+        if (bits < 15) {
+            hold += (unsigned long)(PUP(in)) << bits;
+            bits += 8;
+            hold += (unsigned long)(PUP(in)) << bits;
+            bits += 8;
+        }
+        here = lcode[hold & lmask];
+      dolen:
+        op = (unsigned)(here.bits);
+        hold >>= op;
+        bits -= op;
+        op = (unsigned)(here.op);
+        if (op == 0) {                          /* literal */
+            Tracevv((stderr, here.val >= 0x20 && here.val < 0x7f ?
+                    "inflate:         literal '%c'\n" :
+                    "inflate:         literal 0x%02x\n", here.val));
+            PUP(out) = (unsigned char)(here.val);
+        }
+        else if (op & 16) {                     /* length base */
+            len = (unsigned)(here.val);
+            op &= 15;                           /* number of extra bits */
+            if (op) {
+                if (bits < op) {
+                    hold += (unsigned long)(PUP(in)) << bits;
+                    bits += 8;
+                }
+                len += (unsigned)hold & ((1U << op) - 1);
+                hold >>= op;
+                bits -= op;
+            }
+            Tracevv((stderr, "inflate:         length %u\n", len));
+            if (bits < 15) {
+                hold += (unsigned long)(PUP(in)) << bits;
+                bits += 8;
+                hold += (unsigned long)(PUP(in)) << bits;
+                bits += 8;
+            }
+            here = dcode[hold & dmask];
+          dodist:
+            op = (unsigned)(here.bits);
+            hold >>= op;
+            bits -= op;
+            op = (unsigned)(here.op);
+            if (op & 16) {                      /* distance base */
+                dist = (unsigned)(here.val);
+                op &= 15;                       /* number of extra bits */
+                if (bits < op) {
+                    hold += (unsigned long)(PUP(in)) << bits;
+                    bits += 8;
+                    if (bits < op) {
+                        hold += (unsigned long)(PUP(in)) << bits;
+                        bits += 8;
+                    }
+                }
+                dist += (unsigned)hold & ((1U << op) - 1);
+#ifdef INFLATE_STRICT
+                if (dist > dmax) {
+                    strm->msg = (char *)"invalid distance too far back";
+                    state->mode = BAD;
+                    break;
+                }
+#endif
+                hold >>= op;
+                bits -= op;
+                Tracevv((stderr, "inflate:         distance %u\n", dist));
+                op = (unsigned)(out - beg);     /* max distance in output */
+                if (dist > op) {                /* see if copy from window */
+                    op = dist - op;             /* distance back in window */
+                    if (op > whave) {
+                        if (state->sane) {
+                            strm->msg =
+                                (char *)"invalid distance too far back";
+                            state->mode = BAD;
+                            break;
+                        }
+#ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR
+                        if (len <= op - whave) {
+                            do {
+                                PUP(out) = 0;
+                            } while (--len);
+                            continue;
+                        }
+                        len -= op - whave;
+                        do {
+                            PUP(out) = 0;
+                        } while (--op > whave);
+                        if (op == 0) {
+                            from = out - dist;
+                            do {
+                                PUP(out) = PUP(from);
+                            } while (--len);
+                            continue;
+                        }
+#endif
+                    }
+                    from = window - OFF;
+                    if (wnext == 0) {           /* very common case */
+                        from += wsize - op;
+                        if (op < len) {         /* some from window */
+                            len -= op;
+                            do {
+                                PUP(out) = PUP(from);
+                            } while (--op);
+                            from = out - dist;  /* rest from output */
+                        }
+                    }
+                    else if (wnext < op) {      /* wrap around window */
+                        from += wsize + wnext - op;
+                        op -= wnext;
+                        if (op < len) {         /* some from end of window */
+                            len -= op;
+                            do {
+                                PUP(out) = PUP(from);
+                            } while (--op);
+                            from = window - OFF;
+                            if (wnext < len) {  /* some from start of window */
+                                op = wnext;
+                                len -= op;
+                                do {
+                                    PUP(out) = PUP(from);
+                                } while (--op);
+                                from = out - dist;      /* rest from output */
+                            }
+                        }
+                    }
+                    else {                      /* contiguous in window */
+                        from += wnext - op;
+                        if (op < len) {         /* some from window */
+                            len -= op;
+                            do {
+                                PUP(out) = PUP(from);
+                            } while (--op);
+                            from = out - dist;  /* rest from output */
+                        }
+                    }
+                    while (len > 2) {
+                        PUP(out) = PUP(from);
+                        PUP(out) = PUP(from);
+                        PUP(out) = PUP(from);
+                        len -= 3;
+                    }
+                    if (len) {
+                        PUP(out) = PUP(from);
+                        if (len > 1)
+                            PUP(out) = PUP(from);
+                    }
+                }
+                else {
+                    from = out - dist;          /* copy direct from output */
+                    do {                        /* minimum length is three */
+                        PUP(out) = PUP(from);
+                        PUP(out) = PUP(from);
+                        PUP(out) = PUP(from);
+                        len -= 3;
+                    } while (len > 2);
+                    if (len) {
+                        PUP(out) = PUP(from);
+                        if (len > 1)
+                            PUP(out) = PUP(from);
+                    }
+                }
+            }
+            else if ((op & 64) == 0) {          /* 2nd level distance code */
+                here = dcode[here.val + (hold & ((1U << op) - 1))];
+                goto dodist;
+            }
+            else {
+                strm->msg = (char *)"invalid distance code";
+                state->mode = BAD;
+                break;
+            }
+        }
+        else if ((op & 64) == 0) {              /* 2nd level length code */
+            here = lcode[here.val + (hold & ((1U << op) - 1))];
+            goto dolen;
+        }
+        else if (op & 32) {                     /* end-of-block */
+            Tracevv((stderr, "inflate:         end of block\n"));
+            state->mode = TYPE;
+            break;
+        }
+        else {
+            strm->msg = (char *)"invalid literal/length code";
+            state->mode = BAD;
+            break;
+        }
+    } while (in < last && out < end);
+
+    /* return unused bytes (on entry, bits < 8, so in won't go too far back) */
+    len = bits >> 3;
+    in -= len;
+    bits -= len << 3;
+    hold &= (1U << bits) - 1;
+
+    /* update state and return */
+    strm->next_in = in + OFF;
+    strm->next_out = out + OFF;
+    strm->avail_in = (unsigned)(in < last ? 5 + (last - in) : 5 - (in - last));
+    strm->avail_out = (unsigned)(out < end ?
+                                 257 + (end - out) : 257 - (out - end));
+    state->hold = hold;
+    state->bits = bits;
+    return;
+}
+
+/*
+   inflate_fast() speedups that turned out slower (on a PowerPC G3 750CXe):
+   - Using bit fields for code structure
+   - Different op definition to avoid & for extra bits (do & for table bits)
+   - Three separate decoding do-loops for direct, window, and wnext == 0
+   - Special case for distance > 1 copies to do overlapped load and store copy
+   - Explicit branch predictions (based on measured branch probabilities)
+   - Deferring match copy and interspersed it with decoding subsequent codes
+   - Swapping literal/length else
+   - Swapping window/direct else
+   - Larger unrolled copy loops (three is about right)
+   - Moving len -= 3 statement into middle of loop
+ */
+
+#endif /* !ASMINF */
diff --git a/ml/dlib/dlib/external/zlib/inffast.h b/ml/dlib/dlib/external/zlib/inffast.h
new file mode 100644
index 000000000..e5c1aa4ca
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inffast.h
@@ -0,0 +1,11 @@
+/* inffast.h -- header to use inffast.c
+ * Copyright (C) 1995-2003, 2010 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* WARNING: this file should *not* be used by applications. It is
+   part of the implementation of the compression library and is
+   subject to change. Applications should only use zlib.h.
+ */
+
+void ZLIB_INTERNAL inflate_fast OF((z_streamp strm, unsigned start));
diff --git a/ml/dlib/dlib/external/zlib/inffixed.h b/ml/dlib/dlib/external/zlib/inffixed.h
new file mode 100644
index 000000000..d62832776
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inffixed.h
@@ -0,0 +1,94 @@
+    /* inffixed.h -- table for decoding fixed codes
+     * Generated automatically by makefixed().
+     */
+
+    /* WARNING: this file should *not* be used by applications.
+       It is part of the implementation of this library and is
+       subject to change. Applications should only use zlib.h.
+     */
+
+    static const code lenfix[512] = {
+        {96,7,0},{0,8,80},{0,8,16},{20,8,115},{18,7,31},{0,8,112},{0,8,48},
+        {0,9,192},{16,7,10},{0,8,96},{0,8,32},{0,9,160},{0,8,0},{0,8,128},
+        {0,8,64},{0,9,224},{16,7,6},{0,8,88},{0,8,24},{0,9,144},{19,7,59},
+        {0,8,120},{0,8,56},{0,9,208},{17,7,17},{0,8,104},{0,8,40},{0,9,176},
+        {0,8,8},{0,8,136},{0,8,72},{0,9,240},{16,7,4},{0,8,84},{0,8,20},
+        {21,8,227},{19,7,43},{0,8,116},{0,8,52},{0,9,200},{17,7,13},{0,8,100},
+        {0,8,36},{0,9,168},{0,8,4},{0,8,132},{0,8,68},{0,9,232},{16,7,8},
+        {0,8,92},{0,8,28},{0,9,152},{20,7,83},{0,8,124},{0,8,60},{0,9,216},
+        {18,7,23},{0,8,108},{0,8,44},{0,9,184},{0,8,12},{0,8,140},{0,8,76},
+        {0,9,248},{16,7,3},{0,8,82},{0,8,18},{21,8,163},{19,7,35},{0,8,114},
+        {0,8,50},{0,9,196},{17,7,11},{0,8,98},{0,8,34},{0,9,164},{0,8,2},
+        {0,8,130},{0,8,66},{0,9,228},{16,7,7},{0,8,90},{0,8,26},{0,9,148},
+        {20,7,67},{0,8,122},{0,8,58},{0,9,212},{18,7,19},{0,8,106},{0,8,42},
+        {0,9,180},{0,8,10},{0,8,138},{0,8,74},{0,9,244},{16,7,5},{0,8,86},
+        {0,8,22},{64,8,0},{19,7,51},{0,8,118},{0,8,54},{0,9,204},{17,7,15},
+        {0,8,102},{0,8,38},{0,9,172},{0,8,6},{0,8,134},{0,8,70},{0,9,236},
+        {16,7,9},{0,8,94},{0,8,30},{0,9,156},{20,7,99},{0,8,126},{0,8,62},
+        {0,9,220},{18,7,27},{0,8,110},{0,8,46},{0,9,188},{0,8,14},{0,8,142},
+        {0,8,78},{0,9,252},{96,7,0},{0,8,81},{0,8,17},{21,8,131},{18,7,31},
+        {0,8,113},{0,8,49},{0,9,194},{16,7,10},{0,8,97},{0,8,33},{0,9,162},
+        {0,8,1},{0,8,129},{0,8,65},{0,9,226},{16,7,6},{0,8,89},{0,8,25},
+        {0,9,146},{19,7,59},{0,8,121},{0,8,57},{0,9,210},{17,7,17},{0,8,105},
+        {0,8,41},{0,9,178},{0,8,9},{0,8,137},{0,8,73},{0,9,242},{16,7,4},
+        {0,8,85},{0,8,21},{16,8,258},{19,7,43},{0,8,117},{0,8,53},{0,9,202},
+        {17,7,13},{0,8,101},{0,8,37},{0,9,170},{0,8,5},{0,8,133},{0,8,69},
+        {0,9,234},{16,7,8},{0,8,93},{0,8,29},{0,9,154},{20,7,83},{0,8,125},
+        {0,8,61},{0,9,218},{18,7,23},{0,8,109},{0,8,45},{0,9,186},{0,8,13},
+        {0,8,141},{0,8,77},{0,9,250},{16,7,3},{0,8,83},{0,8,19},{21,8,195},
+        {19,7,35},{0,8,115},{0,8,51},{0,9,198},{17,7,11},{0,8,99},{0,8,35},
+        {0,9,166},{0,8,3},{0,8,131},{0,8,67},{0,9,230},{16,7,7},{0,8,91},
+        {0,8,27},{0,9,150},{20,7,67},{0,8,123},{0,8,59},{0,9,214},{18,7,19},
+        {0,8,107},{0,8,43},{0,9,182},{0,8,11},{0,8,139},{0,8,75},{0,9,246},
+        {16,7,5},{0,8,87},{0,8,23},{64,8,0},{19,7,51},{0,8,119},{0,8,55},
+        {0,9,206},{17,7,15},{0,8,103},{0,8,39},{0,9,174},{0,8,7},{0,8,135},
+        {0,8,71},{0,9,238},{16,7,9},{0,8,95},{0,8,31},{0,9,158},{20,7,99},
+        {0,8,127},{0,8,63},{0,9,222},{18,7,27},{0,8,111},{0,8,47},{0,9,190},
+        {0,8,15},{0,8,143},{0,8,79},{0,9,254},{96,7,0},{0,8,80},{0,8,16},
+        {20,8,115},{18,7,31},{0,8,112},{0,8,48},{0,9,193},{16,7,10},{0,8,96},
+        {0,8,32},{0,9,161},{0,8,0},{0,8,128},{0,8,64},{0,9,225},{16,7,6},
+        {0,8,88},{0,8,24},{0,9,145},{19,7,59},{0,8,120},{0,8,56},{0,9,209},
+        {17,7,17},{0,8,104},{0,8,40},{0,9,177},{0,8,8},{0,8,136},{0,8,72},
+        {0,9,241},{16,7,4},{0,8,84},{0,8,20},{21,8,227},{19,7,43},{0,8,116},
+        {0,8,52},{0,9,201},{17,7,13},{0,8,100},{0,8,36},{0,9,169},{0,8,4},
+        {0,8,132},{0,8,68},{0,9,233},{16,7,8},{0,8,92},{0,8,28},{0,9,153},
+        {20,7,83},{0,8,124},{0,8,60},{0,9,217},{18,7,23},{0,8,108},{0,8,44},
+        {0,9,185},{0,8,12},{0,8,140},{0,8,76},{0,9,249},{16,7,3},{0,8,82},
+        {0,8,18},{21,8,163},{19,7,35},{0,8,114},{0,8,50},{0,9,197},{17,7,11},
+        {0,8,98},{0,8,34},{0,9,165},{0,8,2},{0,8,130},{0,8,66},{0,9,229},
+        {16,7,7},{0,8,90},{0,8,26},{0,9,149},{20,7,67},{0,8,122},{0,8,58},
+        {0,9,213},{18,7,19},{0,8,106},{0,8,42},{0,9,181},{0,8,10},{0,8,138},
+        {0,8,74},{0,9,245},{16,7,5},{0,8,86},{0,8,22},{64,8,0},{19,7,51},
+        {0,8,118},{0,8,54},{0,9,205},{17,7,15},{0,8,102},{0,8,38},{0,9,173},
+        {0,8,6},{0,8,134},{0,8,70},{0,9,237},{16,7,9},{0,8,94},{0,8,30},
+        {0,9,157},{20,7,99},{0,8,126},{0,8,62},{0,9,221},{18,7,27},{0,8,110},
+        {0,8,46},{0,9,189},{0,8,14},{0,8,142},{0,8,78},{0,9,253},{96,7,0},
+        {0,8,81},{0,8,17},{21,8,131},{18,7,31},{0,8,113},{0,8,49},{0,9,195},
+        {16,7,10},{0,8,97},{0,8,33},{0,9,163},{0,8,1},{0,8,129},{0,8,65},
+        {0,9,227},{16,7,6},{0,8,89},{0,8,25},{0,9,147},{19,7,59},{0,8,121},
+        {0,8,57},{0,9,211},{17,7,17},{0,8,105},{0,8,41},{0,9,179},{0,8,9},
+        {0,8,137},{0,8,73},{0,9,243},{16,7,4},{0,8,85},{0,8,21},{16,8,258},
+        {19,7,43},{0,8,117},{0,8,53},{0,9,203},{17,7,13},{0,8,101},{0,8,37},
+        {0,9,171},{0,8,5},{0,8,133},{0,8,69},{0,9,235},{16,7,8},{0,8,93},
+        {0,8,29},{0,9,155},{20,7,83},{0,8,125},{0,8,61},{0,9,219},{18,7,23},
+        {0,8,109},{0,8,45},{0,9,187},{0,8,13},{0,8,141},{0,8,77},{0,9,251},
+        {16,7,3},{0,8,83},{0,8,19},{21,8,195},{19,7,35},{0,8,115},{0,8,51},
+        {0,9,199},{17,7,11},{0,8,99},{0,8,35},{0,9,167},{0,8,3},{0,8,131},
+        {0,8,67},{0,9,231},{16,7,7},{0,8,91},{0,8,27},{0,9,151},{20,7,67},
+        {0,8,123},{0,8,59},{0,9,215},{18,7,19},{0,8,107},{0,8,43},{0,9,183},
+        {0,8,11},{0,8,139},{0,8,75},{0,9,247},{16,7,5},{0,8,87},{0,8,23},
+        {64,8,0},{19,7,51},{0,8,119},{0,8,55},{0,9,207},{17,7,15},{0,8,103},
+        {0,8,39},{0,9,175},{0,8,7},{0,8,135},{0,8,71},{0,9,239},{16,7,9},
+        {0,8,95},{0,8,31},{0,9,159},{20,7,99},{0,8,127},{0,8,63},{0,9,223},
+        {18,7,27},{0,8,111},{0,8,47},{0,9,191},{0,8,15},{0,8,143},{0,8,79},
+        {0,9,255}
+    };
+
+    static const code distfix[32] = {
+        {16,5,1},{23,5,257},{19,5,17},{27,5,4097},{17,5,5},{25,5,1025},
+        {21,5,65},{29,5,16385},{16,5,3},{24,5,513},{20,5,33},{28,5,8193},
+        {18,5,9},{26,5,2049},{22,5,129},{64,5,0},{16,5,2},{23,5,385},
+        {19,5,25},{27,5,6145},{17,5,7},{25,5,1537},{21,5,97},{29,5,24577},
+        {16,5,4},{24,5,769},{20,5,49},{28,5,12289},{18,5,13},{26,5,3073},
+        {22,5,193},{64,5,0}
+    };
diff --git a/ml/dlib/dlib/external/zlib/inflate.c b/ml/dlib/dlib/external/zlib/inflate.c
new file mode 100644
index 000000000..870f89bb4
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inflate.c
@@ -0,0 +1,1512 @@
+/* inflate.c -- zlib decompression
+ * Copyright (C) 1995-2012 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/*
+ * Change history:
+ *
+ * 1.2.beta0    24 Nov 2002
+ * - First version -- complete rewrite of inflate to simplify code, avoid
+ *   creation of window when not needed, minimize use of window when it is
+ *   needed, make inffast.c even faster, implement gzip decoding, and to
+ *   improve code readability and style over the previous zlib inflate code
+ *
+ * 1.2.beta1    25 Nov 2002
+ * - Use pointers for available input and output checking in inffast.c
+ * - Remove input and output counters in inffast.c
+ * - Change inffast.c entry and loop from avail_in >= 7 to >= 6
+ * - Remove unnecessary second byte pull from length extra in inffast.c
+ * - Unroll direct copy to three copies per loop in inffast.c
+ *
+ * 1.2.beta2    4 Dec 2002
+ * - Change external routine names to reduce potential conflicts
+ * - Correct filename to inffixed.h for fixed tables in inflate.c
+ * - Make hbuf[] unsigned char to match parameter type in inflate.c
+ * - Change strm->next_out[-state->offset] to *(strm->next_out - state->offset)
+ *   to avoid negation problem on Alphas (64 bit) in inflate.c
+ *
+ * 1.2.beta3    22 Dec 2002
+ * - Add comments on state->bits assertion in inffast.c
+ * - Add comments on op field in inftrees.h
+ * - Fix bug in reuse of allocated window after inflateReset()
+ * - Remove bit fields--back to byte structure for speed
+ * - Remove distance extra == 0 check in inflate_fast()--only helps for lengths
+ * - Change post-increments to pre-increments in inflate_fast(), PPC biased?
+ * - Add compile time option, POSTINC, to use post-increments instead (Intel?)
+ * - Make MATCH copy in inflate() much faster for when inflate_fast() not used
+ * - Use local copies of stream next and avail values, as well as local bit
+ *   buffer and bit count in inflate()--for speed when inflate_fast() not used
+ *
+ * 1.2.beta4    1 Jan 2003
+ * - Split ptr - 257 statements in inflate_table() to avoid compiler warnings
+ * - Move a comment on output buffer sizes from inffast.c to inflate.c
+ * - Add comments in inffast.c to introduce the inflate_fast() routine
+ * - Rearrange window copies in inflate_fast() for speed and simplification
+ * - Unroll last copy for window match in inflate_fast()
+ * - Use local copies of window variables in inflate_fast() for speed
+ * - Pull out common wnext == 0 case for speed in inflate_fast()
+ * - Make op and len in inflate_fast() unsigned for consistency
+ * - Add FAR to lcode and dcode declarations in inflate_fast()
+ * - Simplified bad distance check in inflate_fast()
+ * - Added inflateBackInit(), inflateBack(), and inflateBackEnd() in new
+ *   source file infback.c to provide a call-back interface to inflate for
+ *   programs like gzip and unzip -- uses window as output buffer to avoid
+ *   window copying
+ *
+ * 1.2.beta5    1 Jan 2003
+ * - Improved inflateBack() interface to allow the caller to provide initial
+ *   input in strm.
+ * - Fixed stored blocks bug in inflateBack()
+ *
+ * 1.2.beta6    4 Jan 2003
+ * - Added comments in inffast.c on effectiveness of POSTINC
+ * - Typecasting all around to reduce compiler warnings
+ * - Changed loops from while (1) or do {} while (1) to for (;;), again to
+ *   make compilers happy
+ * - Changed type of window in inflateBackInit() to unsigned char *
+ *
+ * 1.2.beta7    27 Jan 2003
+ * - Changed many types to unsigned or unsigned short to avoid warnings
+ * - Added inflateCopy() function
+ *
+ * 1.2.0        9 Mar 2003
+ * - Changed inflateBack() interface to provide separate opaque descriptors
+ *   for the in() and out() functions
+ * - Changed inflateBack() argument and in_func typedef to swap the length
+ *   and buffer address return values for the input function
+ * - Check next_in and next_out for Z_NULL on entry to inflate()
+ *
+ * The history for versions after 1.2.0 are in ChangeLog in zlib distribution.
+ */
+
+#include "zutil.h"
+#include "inftrees.h"
+#include "inflate.h"
+#include "inffast.h"
+
+#ifdef MAKEFIXED
+#  ifndef BUILDFIXED
+#    define BUILDFIXED
+#  endif
+#endif
+
+/* function prototypes */
+local void fixedtables OF((struct inflate_state FAR *state));
+local int updatewindow OF((z_streamp strm, const unsigned char FAR *end,
+                           unsigned copy));
+#ifdef BUILDFIXED
+   void makefixed OF((void));
+#endif
+local unsigned syncsearch OF((unsigned FAR *have, const unsigned char FAR *buf,
+                              unsigned len));
+
+int ZEXPORT inflateResetKeep(strm)
+z_streamp strm;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    strm->total_in = strm->total_out = state->total = 0;
+    strm->msg = Z_NULL;
+    if (state->wrap)        /* to support ill-conceived Java test suite */
+        strm->adler = state->wrap & 1;
+    state->mode = HEAD;
+    state->last = 0;
+    state->havedict = 0;
+    state->dmax = 32768U;
+    state->head = Z_NULL;
+    state->hold = 0;
+    state->bits = 0;
+    state->lencode = state->distcode = state->next = state->codes;
+    state->sane = 1;
+    state->back = -1;
+    Tracev((stderr, "inflate: reset\n"));
+    return Z_OK;
+}
+
+int ZEXPORT inflateReset(strm)
+z_streamp strm;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    state->wsize = 0;
+    state->whave = 0;
+    state->wnext = 0;
+    return inflateResetKeep(strm);
+}
+
+int ZEXPORT inflateReset2(strm, windowBits)
+z_streamp strm;
+int windowBits;
+{
+    int wrap;
+    struct inflate_state FAR *state;
+
+    /* get the state */
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+
+    /* extract wrap request from windowBits parameter */
+    if (windowBits < 0) {
+        wrap = 0;
+        windowBits = -windowBits;
+    }
+    else {
+        wrap = (windowBits >> 4) + 1;
+#ifdef GUNZIP
+        if (windowBits < 48)
+            windowBits &= 15;
+#endif
+    }
+
+    /* set number of window bits, free window if different */
+    if (windowBits && (windowBits < 8 || windowBits > 15))
+        return Z_STREAM_ERROR;
+    if (state->window != Z_NULL && state->wbits != (unsigned)windowBits) {
+        ZFREE(strm, state->window);
+        state->window = Z_NULL;
+    }
+
+    /* update state and reset the rest of it */
+    state->wrap = wrap;
+    state->wbits = (unsigned)windowBits;
+    return inflateReset(strm);
+}
+
+int ZEXPORT inflateInit2_(strm, windowBits, version, stream_size)
+z_streamp strm;
+int windowBits;
+const char *version;
+int stream_size;
+{
+    int ret;
+    struct inflate_state FAR *state;
+
+    if (version == Z_NULL || version[0] != ZLIB_VERSION[0] ||
+        stream_size != (int)(sizeof(z_stream)))
+        return Z_VERSION_ERROR;
+    if (strm == Z_NULL) return Z_STREAM_ERROR;
+    strm->msg = Z_NULL;                 /* in case we return an error */
+    if (strm->zalloc == (alloc_func)0) {
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+        strm->zalloc = zcalloc;
+        strm->opaque = (voidpf)0;
+#endif
+    }
+    if (strm->zfree == (free_func)0)
+#ifdef Z_SOLO
+        return Z_STREAM_ERROR;
+#else
+        strm->zfree = zcfree;
+#endif
+    state = (struct inflate_state FAR *)
+            ZALLOC(strm, 1, sizeof(struct inflate_state));
+    if (state == Z_NULL) return Z_MEM_ERROR;
+    Tracev((stderr, "inflate: allocated\n"));
+    strm->state = (struct internal_state FAR *)state;
+    state->window = Z_NULL;
+    ret = inflateReset2(strm, windowBits);
+    if (ret != Z_OK) {
+        ZFREE(strm, state);
+        strm->state = Z_NULL;
+    }
+    return ret;
+}
+
+int ZEXPORT inflateInit_(strm, version, stream_size)
+z_streamp strm;
+const char *version;
+int stream_size;
+{
+    return inflateInit2_(strm, DEF_WBITS, version, stream_size);
+}
+
+int ZEXPORT inflatePrime(strm, bits, value)
+z_streamp strm;
+int bits;
+int value;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    if (bits < 0) {
+        state->hold = 0;
+        state->bits = 0;
+        return Z_OK;
+    }
+    if (bits > 16 || state->bits + bits > 32) return Z_STREAM_ERROR;
+    value &= (1L << bits) - 1;
+    state->hold += value << state->bits;
+    state->bits += bits;
+    return Z_OK;
+}
+
+/*
+   Return state with length and distance decoding tables and index sizes set to
+   fixed code decoding.  Normally this returns fixed tables from inffixed.h.
+   If BUILDFIXED is defined, then instead this routine builds the tables the
+   first time it's called, and returns those tables the first time and
+   thereafter.  This reduces the size of the code by about 2K bytes, in
+   exchange for a little execution time.  However, BUILDFIXED should not be
+   used for threaded applications, since the rewriting of the tables and virgin
+   may not be thread-safe.
+ */
+local void fixedtables(state)
+struct inflate_state FAR *state;
+{
+#ifdef BUILDFIXED
+    static int virgin = 1;
+    static code *lenfix, *distfix;
+    static code fixed[544];
+
+    /* build fixed huffman tables if first call (may not be thread safe) */
+    if (virgin) {
+        unsigned sym, bits;
+        static code *next;
+
+        /* literal/length table */
+        sym = 0;
+        while (sym < 144) state->lens[sym++] = 8;
+        while (sym < 256) state->lens[sym++] = 9;
+        while (sym < 280) state->lens[sym++] = 7;
+        while (sym < 288) state->lens[sym++] = 8;
+        next = fixed;
+        lenfix = next;
+        bits = 9;
+        inflate_table(LENS, state->lens, 288, &(next), &(bits), state->work);
+
+        /* distance table */
+        sym = 0;
+        while (sym < 32) state->lens[sym++] = 5;
+        distfix = next;
+        bits = 5;
+        inflate_table(DISTS, state->lens, 32, &(next), &(bits), state->work);
+
+        /* do this just once */
+        virgin = 0;
+    }
+#else /* !BUILDFIXED */
+#   include "inffixed.h"
+#endif /* BUILDFIXED */
+    state->lencode = lenfix;
+    state->lenbits = 9;
+    state->distcode = distfix;
+    state->distbits = 5;
+}
+
+#ifdef MAKEFIXED
+#include <stdio.h>
+
+/*
+   Write out the inffixed.h that is #include'd above.  Defining MAKEFIXED also
+   defines BUILDFIXED, so the tables are built on the fly.  makefixed() writes
+   those tables to stdout, which would be piped to inffixed.h.  A small program
+   can simply call makefixed to do this:
+
+    void makefixed(void);
+
+    int main(void)
+    {
+        makefixed();
+        return 0;
+    }
+
+   Then that can be linked with zlib built with MAKEFIXED defined and run:
+
+    a.out > inffixed.h
+ */
+void makefixed()
+{
+    unsigned low, size;
+    struct inflate_state state;
+
+    fixedtables(&state);
+    puts("    /* inffixed.h -- table for decoding fixed codes");
+    puts("     * Generated automatically by makefixed().");
+    puts("     */");
+    puts("");
+    puts("    /* WARNING: this file should *not* be used by applications.");
+    puts("       It is part of the implementation of this library and is");
+    puts("       subject to change. Applications should only use zlib.h.");
+    puts("     */");
+    puts("");
+    size = 1U << 9;
+    printf("    static const code lenfix[%u] = {", size);
+    low = 0;
+    for (;;) {
+        if ((low % 7) == 0) printf("\n        ");
+        printf("{%u,%u,%d}", (low & 127) == 99 ? 64 : state.lencode[low].op,
+               state.lencode[low].bits, state.lencode[low].val);
+        if (++low == size) break;
+        putchar(',');
+    }
+    puts("\n    };");
+    size = 1U << 5;
+    printf("\n    static const code distfix[%u] = {", size);
+    low = 0;
+    for (;;) {
+        if ((low % 6) == 0) printf("\n        ");
+        printf("{%u,%u,%d}", state.distcode[low].op, state.distcode[low].bits,
+               state.distcode[low].val);
+        if (++low == size) break;
+        putchar(',');
+    }
+    puts("\n    };");
+}
+#endif /* MAKEFIXED */
+
+/*
+   Update the window with the last wsize (normally 32K) bytes written before
+   returning.  If window does not exist yet, create it.  This is only called
+   when a window is already in use, or when output has been written during this
+   inflate call, but the end of the deflate stream has not been reached yet.
+   It is also called to create a window for dictionary data when a dictionary
+   is loaded.
+
+   Providing output buffers larger than 32K to inflate() should provide a speed
+   advantage, since only the last 32K of output is copied to the sliding window
+   upon return from inflate(), and since all distances after the first 32K of
+   output will fall in the output data, making match copies simpler and faster.
+   The advantage may be dependent on the size of the processor's data caches.
+ */
+local int updatewindow(strm, end, copy)
+z_streamp strm;
+const Bytef *end;
+unsigned copy;
+{
+    struct inflate_state FAR *state;
+    unsigned dist;
+
+    state = (struct inflate_state FAR *)strm->state;
+
+    /* if it hasn't been done already, allocate space for the window */
+    if (state->window == Z_NULL) {
+        state->window = (unsigned char FAR *)
+                        ZALLOC(strm, 1U << state->wbits,
+                               sizeof(unsigned char));
+        if (state->window == Z_NULL) return 1;
+    }
+
+    /* if window not in use yet, initialize */
+    if (state->wsize == 0) {
+        state->wsize = 1U << state->wbits;
+        state->wnext = 0;
+        state->whave = 0;
+    }
+
+    /* copy state->wsize or less output bytes into the circular window */
+    if (copy >= state->wsize) {
+        zmemcpy(state->window, end - state->wsize, state->wsize);
+        state->wnext = 0;
+        state->whave = state->wsize;
+    }
+    else {
+        dist = state->wsize - state->wnext;
+        if (dist > copy) dist = copy;
+        zmemcpy(state->window + state->wnext, end - copy, dist);
+        copy -= dist;
+        if (copy) {
+            zmemcpy(state->window, end - copy, copy);
+            state->wnext = copy;
+            state->whave = state->wsize;
+        }
+        else {
+            state->wnext += dist;
+            if (state->wnext == state->wsize) state->wnext = 0;
+            if (state->whave < state->wsize) state->whave += dist;
+        }
+    }
+    return 0;
+}
+
+/* Macros for inflate(): */
+
+/* check function to use adler32() for zlib or crc32() for gzip */
+#ifdef GUNZIP
+#  define UPDATE(check, buf, len) \
+    (state->flags ? crc32(check, buf, len) : adler32(check, buf, len))
+#else
+#  define UPDATE(check, buf, len) adler32(check, buf, len)
+#endif
+
+/* check macros for header crc */
+#ifdef GUNZIP
+#  define CRC2(check, word) \
+    do { \
+        hbuf[0] = (unsigned char)(word); \
+        hbuf[1] = (unsigned char)((word) >> 8); \
+        check = crc32(check, hbuf, 2); \
+    } while (0)
+
+#  define CRC4(check, word) \
+    do { \
+        hbuf[0] = (unsigned char)(word); \
+        hbuf[1] = (unsigned char)((word) >> 8); \
+        hbuf[2] = (unsigned char)((word) >> 16); \
+        hbuf[3] = (unsigned char)((word) >> 24); \
+        check = crc32(check, hbuf, 4); \
+    } while (0)
+#endif
+
+/* Load registers with state in inflate() for speed */
+#define LOAD() \
+    do { \
+        put = strm->next_out; \
+        left = strm->avail_out; \
+        next = strm->next_in; \
+        have = strm->avail_in; \
+        hold = state->hold; \
+        bits = state->bits; \
+    } while (0)
+
+/* Restore state from registers in inflate() */
+#define RESTORE() \
+    do { \
+        strm->next_out = put; \
+        strm->avail_out = left; \
+        strm->next_in = next; \
+        strm->avail_in = have; \
+        state->hold = hold; \
+        state->bits = bits; \
+    } while (0)
+
+/* Clear the input bit accumulator */
+#define INITBITS() \
+    do { \
+        hold = 0; \
+        bits = 0; \
+    } while (0)
+
+/* Get a byte of input into the bit accumulator, or return from inflate()
+   if there is no input available. */
+#define PULLBYTE() \
+    do { \
+        if (have == 0) goto inf_leave; \
+        have--; \
+        hold += (unsigned long)(*next++) << bits; \
+        bits += 8; \
+    } while (0)
+
+/* Assure that there are at least n bits in the bit accumulator.  If there is
+   not enough available input to do that, then return from inflate(). */
+#define NEEDBITS(n) \
+    do { \
+        while (bits < (unsigned)(n)) \
+            PULLBYTE(); \
+    } while (0)
+
+/* Return the low n bits of the bit accumulator (n < 16) */
+#define BITS(n) \
+    ((unsigned)hold & ((1U << (n)) - 1))
+
+/* Remove n bits from the bit accumulator */
+#define DROPBITS(n) \
+    do { \
+        hold >>= (n); \
+        bits -= (unsigned)(n); \
+    } while (0)
+
+/* Remove zero to seven bits as needed to go to a byte boundary */
+#define BYTEBITS() \
+    do { \
+        hold >>= bits & 7; \
+        bits -= bits & 7; \
+    } while (0)
+
+/*
+   inflate() uses a state machine to process as much input data and generate as
+   much output data as possible before returning.  The state machine is
+   structured roughly as follows:
+
+    for (;;) switch (state) {
+    ...
+    case STATEn:
+        if (not enough input data or output space to make progress)
+            return;
+        ... make progress ...
+        state = STATEm;
+        break;
+    ...
+    }
+
+   so when inflate() is called again, the same case is attempted again, and
+   if the appropriate resources are provided, the machine proceeds to the
+   next state.  The NEEDBITS() macro is usually the way the state evaluates
+   whether it can proceed or should return.  NEEDBITS() does the return if
+   the requested bits are not available.  The typical use of the BITS macros
+   is:
+
+        NEEDBITS(n);
+        ... do something with BITS(n) ...
+        DROPBITS(n);
+
+   where NEEDBITS(n) either returns from inflate() if there isn't enough
+   input left to load n bits into the accumulator, or it continues.  BITS(n)
+   gives the low n bits in the accumulator.  When done, DROPBITS(n) drops
+   the low n bits off the accumulator.  INITBITS() clears the accumulator
+   and sets the number of available bits to zero.  BYTEBITS() discards just
+   enough bits to put the accumulator on a byte boundary.  After BYTEBITS()
+   and a NEEDBITS(8), then BITS(8) would return the next byte in the stream.
+
+   NEEDBITS(n) uses PULLBYTE() to get an available byte of input, or to return
+   if there is no input available.  The decoding of variable length codes uses
+   PULLBYTE() directly in order to pull just enough bytes to decode the next
+   code, and no more.
+
+   Some states loop until they get enough input, making sure that enough
+   state information is maintained to continue the loop where it left off
+   if NEEDBITS() returns in the loop.  For example, want, need, and keep
+   would all have to actually be part of the saved state in case NEEDBITS()
+   returns:
+
+    case STATEw:
+        while (want < need) {
+            NEEDBITS(n);
+            keep[want++] = BITS(n);
+            DROPBITS(n);
+        }
+        state = STATEx;
+    case STATEx:
+
+   As shown above, if the next state is also the next case, then the break
+   is omitted.
+
+   A state may also return if there is not enough output space available to
+   complete that state.  Those states are copying stored data, writing a
+   literal byte, and copying a matching string.
+
+   When returning, a "goto inf_leave" is used to update the total counters,
+   update the check value, and determine whether any progress has been made
+   during that inflate() call in order to return the proper return code.
+   Progress is defined as a change in either strm->avail_in or strm->avail_out.
+   When there is a window, goto inf_leave will update the window with the last
+   output written.  If a goto inf_leave occurs in the middle of decompression
+   and there is no window currently, goto inf_leave will create one and copy
+   output to the window for the next call of inflate().
+
+   In this implementation, the flush parameter of inflate() only affects the
+   return code (per zlib.h).  inflate() always writes as much as possible to
+   strm->next_out, given the space available and the provided input--the effect
+   documented in zlib.h of Z_SYNC_FLUSH.  Furthermore, inflate() always defers
+   the allocation of and copying into a sliding window until necessary, which
+   provides the effect documented in zlib.h for Z_FINISH when the entire input
+   stream available.  So the only thing the flush parameter actually does is:
+   when flush is set to Z_FINISH, inflate() cannot return Z_OK.  Instead it
+   will return Z_BUF_ERROR if it has not reached the end of the stream.
+ */
+
+int ZEXPORT inflate(strm, flush)
+z_streamp strm;
+int flush;
+{
+    struct inflate_state FAR *state;
+    z_const unsigned char FAR *next;    /* next input */
+    unsigned char FAR *put;     /* next output */
+    unsigned have, left;        /* available input and output */
+    unsigned long hold;         /* bit buffer */
+    unsigned bits;              /* bits in bit buffer */
+    unsigned in, out;           /* save starting available input and output */
+    unsigned copy;              /* number of stored or match bytes to copy */
+    unsigned char FAR *from;    /* where to copy match bytes from */
+    code here;                  /* current decoding table entry */
+    code last;                  /* parent table entry */
+    unsigned len;               /* length to copy for repeats, bits to drop */
+    int ret;                    /* return code */
+#ifdef GUNZIP
+    unsigned char hbuf[4];      /* buffer for gzip header crc calculation */
+#endif
+    static const unsigned short order[19] = /* permutation of code lengths */
+        {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15};
+
+    if (strm == Z_NULL || strm->state == Z_NULL || strm->next_out == Z_NULL ||
+        (strm->next_in == Z_NULL && strm->avail_in != 0))
+        return Z_STREAM_ERROR;
+
+    state = (struct inflate_state FAR *)strm->state;
+    if (state->mode == TYPE) state->mode = TYPEDO;      /* skip check */
+    LOAD();
+    in = have;
+    out = left;
+    ret = Z_OK;
+    for (;;)
+        switch (state->mode) {
+        case HEAD:
+            if (state->wrap == 0) {
+                state->mode = TYPEDO;
+                break;
+            }
+            NEEDBITS(16);
+#ifdef GUNZIP
+            if ((state->wrap & 2) && hold == 0x8b1f) {  /* gzip header */
+                state->check = crc32(0L, Z_NULL, 0);
+                CRC2(state->check, hold);
+                INITBITS();
+                state->mode = FLAGS;
+                break;
+            }
+            state->flags = 0;           /* expect zlib header */
+            if (state->head != Z_NULL)
+                state->head->done = -1;
+            if (!(state->wrap & 1) ||   /* check if zlib header allowed */
+#else
+            if (
+#endif
+                ((BITS(8) << 8) + (hold >> 8)) % 31) {
+                strm->msg = (char *)"incorrect header check";
+                state->mode = BAD;
+                break;
+            }
+            if (BITS(4) != Z_DEFLATED) {
+                strm->msg = (char *)"unknown compression method";
+                state->mode = BAD;
+                break;
+            }
+            DROPBITS(4);
+            len = BITS(4) + 8;
+            if (state->wbits == 0)
+                state->wbits = len;
+            else if (len > state->wbits) {
+                strm->msg = (char *)"invalid window size";
+                state->mode = BAD;
+                break;
+            }
+            state->dmax = 1U << len;
+            Tracev((stderr, "inflate:   zlib header ok\n"));
+            strm->adler = state->check = adler32(0L, Z_NULL, 0);
+            state->mode = hold & 0x200 ? DICTID : TYPE;
+            INITBITS();
+            break;
+#ifdef GUNZIP
+        case FLAGS:
+            NEEDBITS(16);
+            state->flags = (int)(hold);
+            if ((state->flags & 0xff) != Z_DEFLATED) {
+                strm->msg = (char *)"unknown compression method";
+                state->mode = BAD;
+                break;
+            }
+            if (state->flags & 0xe000) {
+                strm->msg = (char *)"unknown header flags set";
+                state->mode = BAD;
+                break;
+            }
+            if (state->head != Z_NULL)
+                state->head->text = (int)((hold >> 8) & 1);
+            if (state->flags & 0x0200) CRC2(state->check, hold);
+            INITBITS();
+            state->mode = TIME;
+        case TIME:
+            NEEDBITS(32);
+            if (state->head != Z_NULL)
+                state->head->time = hold;
+            if (state->flags & 0x0200) CRC4(state->check, hold);
+            INITBITS();
+            state->mode = OS;
+        case OS:
+            NEEDBITS(16);
+            if (state->head != Z_NULL) {
+                state->head->xflags = (int)(hold & 0xff);
+                state->head->os = (int)(hold >> 8);
+            }
+            if (state->flags & 0x0200) CRC2(state->check, hold);
+            INITBITS();
+            state->mode = EXLEN;
+        case EXLEN:
+            if (state->flags & 0x0400) {
+                NEEDBITS(16);
+                state->length = (unsigned)(hold);
+                if (state->head != Z_NULL)
+                    state->head->extra_len = (unsigned)hold;
+                if (state->flags & 0x0200) CRC2(state->check, hold);
+                INITBITS();
+            }
+            else if (state->head != Z_NULL)
+                state->head->extra = Z_NULL;
+            state->mode = EXTRA;
+        case EXTRA:
+            if (state->flags & 0x0400) {
+                copy = state->length;
+                if (copy > have) copy = have;
+                if (copy) {
+                    if (state->head != Z_NULL &&
+                        state->head->extra != Z_NULL) {
+                        len = state->head->extra_len - state->length;
+                        zmemcpy(state->head->extra + len, next,
+                                len + copy > state->head->extra_max ?
+                                state->head->extra_max - len : copy);
+                    }
+                    if (state->flags & 0x0200)
+                        state->check = crc32(state->check, next, copy);
+                    have -= copy;
+                    next += copy;
+                    state->length -= copy;
+                }
+                if (state->length) goto inf_leave;
+            }
+            state->length = 0;
+            state->mode = NAME;
+        case NAME:
+            if (state->flags & 0x0800) {
+                if (have == 0) goto inf_leave;
+                copy = 0;
+                do {
+                    len = (unsigned)(next[copy++]);
+                    if (state->head != Z_NULL &&
+                            state->head->name != Z_NULL &&
+                            state->length < state->head->name_max)
+                        state->head->name[state->length++] = len;
+                } while (len && copy < have);
+                if (state->flags & 0x0200)
+                    state->check = crc32(state->check, next, copy);
+                have -= copy;
+                next += copy;
+                if (len) goto inf_leave;
+            }
+            else if (state->head != Z_NULL)
+                state->head->name = Z_NULL;
+            state->length = 0;
+            state->mode = COMMENT;
+        case COMMENT:
+            if (state->flags & 0x1000) {
+                if (have == 0) goto inf_leave;
+                copy = 0;
+                do {
+                    len = (unsigned)(next[copy++]);
+                    if (state->head != Z_NULL &&
+                            state->head->comment != Z_NULL &&
+                            state->length < state->head->comm_max)
+                        state->head->comment[state->length++] = len;
+                } while (len && copy < have);
+                if (state->flags & 0x0200)
+                    state->check = crc32(state->check, next, copy);
+                have -= copy;
+                next += copy;
+                if (len) goto inf_leave;
+            }
+            else if (state->head != Z_NULL)
+                state->head->comment = Z_NULL;
+            state->mode = HCRC;
+        case HCRC:
+            if (state->flags & 0x0200) {
+                NEEDBITS(16);
+                if (hold != (state->check & 0xffff)) {
+                    strm->msg = (char *)"header crc mismatch";
+                    state->mode = BAD;
+                    break;
+                }
+                INITBITS();
+            }
+            if (state->head != Z_NULL) {
+                state->head->hcrc = (int)((state->flags >> 9) & 1);
+                state->head->done = 1;
+            }
+            strm->adler = state->check = crc32(0L, Z_NULL, 0);
+            state->mode = TYPE;
+            break;
+#endif
+        case DICTID:
+            NEEDBITS(32);
+            strm->adler = state->check = ZSWAP32(hold);
+            INITBITS();
+            state->mode = DICT;
+        case DICT:
+            if (state->havedict == 0) {
+                RESTORE();
+                return Z_NEED_DICT;
+            }
+            strm->adler = state->check = adler32(0L, Z_NULL, 0);
+            state->mode = TYPE;
+        case TYPE:
+            if (flush == Z_BLOCK || flush == Z_TREES) goto inf_leave;
+        case TYPEDO:
+            if (state->last) {
+                BYTEBITS();
+                state->mode = CHECK;
+                break;
+            }
+            NEEDBITS(3);
+            state->last = BITS(1);
+            DROPBITS(1);
+            switch (BITS(2)) {
+            case 0:                             /* stored block */
+                Tracev((stderr, "inflate:     stored block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = STORED;
+                break;
+            case 1:                             /* fixed block */
+                fixedtables(state);
+                Tracev((stderr, "inflate:     fixed codes block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = LEN_;             /* decode codes */
+                if (flush == Z_TREES) {
+                    DROPBITS(2);
+                    goto inf_leave;
+                }
+                break;
+            case 2:                             /* dynamic block */
+                Tracev((stderr, "inflate:     dynamic codes block%s\n",
+                        state->last ? " (last)" : ""));
+                state->mode = TABLE;
+                break;
+            case 3:
+                strm->msg = (char *)"invalid block type";
+                state->mode = BAD;
+            }
+            DROPBITS(2);
+            break;
+        case STORED:
+            BYTEBITS();                         /* go to byte boundary */
+            NEEDBITS(32);
+            if ((hold & 0xffff) != ((hold >> 16) ^ 0xffff)) {
+                strm->msg = (char *)"invalid stored block lengths";
+                state->mode = BAD;
+                break;
+            }
+            state->length = (unsigned)hold & 0xffff;
+            Tracev((stderr, "inflate:       stored length %u\n",
+                    state->length));
+            INITBITS();
+            state->mode = COPY_;
+            if (flush == Z_TREES) goto inf_leave;
+        case COPY_:
+            state->mode = COPY;
+        case COPY:
+            copy = state->length;
+            if (copy) {
+                if (copy > have) copy = have;
+                if (copy > left) copy = left;
+                if (copy == 0) goto inf_leave;
+                zmemcpy(put, next, copy);
+                have -= copy;
+                next += copy;
+                left -= copy;
+                put += copy;
+                state->length -= copy;
+                break;
+            }
+            Tracev((stderr, "inflate:       stored end\n"));
+            state->mode = TYPE;
+            break;
+        case TABLE:
+            NEEDBITS(14);
+            state->nlen = BITS(5) + 257;
+            DROPBITS(5);
+            state->ndist = BITS(5) + 1;
+            DROPBITS(5);
+            state->ncode = BITS(4) + 4;
+            DROPBITS(4);
+#ifndef PKZIP_BUG_WORKAROUND
+            if (state->nlen > 286 || state->ndist > 30) {
+                strm->msg = (char *)"too many length or distance symbols";
+                state->mode = BAD;
+                break;
+            }
+#endif
+            Tracev((stderr, "inflate:       table sizes ok\n"));
+            state->have = 0;
+            state->mode = LENLENS;
+        case LENLENS:
+            while (state->have < state->ncode) {
+                NEEDBITS(3);
+                state->lens[order[state->have++]] = (unsigned short)BITS(3);
+                DROPBITS(3);
+            }
+            while (state->have < 19)
+                state->lens[order[state->have++]] = 0;
+            state->next = state->codes;
+            state->lencode = (const code FAR *)(state->next);
+            state->lenbits = 7;
+            ret = inflate_table(CODES, state->lens, 19, &(state->next),
+                                &(state->lenbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid code lengths set";
+                state->mode = BAD;
+                break;
+            }
+            Tracev((stderr, "inflate:       code lengths ok\n"));
+            state->have = 0;
+            state->mode = CODELENS;
+        case CODELENS:
+            while (state->have < state->nlen + state->ndist) {
+                for (;;) {
+                    here = state->lencode[BITS(state->lenbits)];
+                    if ((unsigned)(here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                if (here.val < 16) {
+                    DROPBITS(here.bits);
+                    state->lens[state->have++] = here.val;
+                }
+                else {
+                    if (here.val == 16) {
+                        NEEDBITS(here.bits + 2);
+                        DROPBITS(here.bits);
+                        if (state->have == 0) {
+                            strm->msg = (char *)"invalid bit length repeat";
+                            state->mode = BAD;
+                            break;
+                        }
+                        len = state->lens[state->have - 1];
+                        copy = 3 + BITS(2);
+                        DROPBITS(2);
+                    }
+                    else if (here.val == 17) {
+                        NEEDBITS(here.bits + 3);
+                        DROPBITS(here.bits);
+                        len = 0;
+                        copy = 3 + BITS(3);
+                        DROPBITS(3);
+                    }
+                    else {
+                        NEEDBITS(here.bits + 7);
+                        DROPBITS(here.bits);
+                        len = 0;
+                        copy = 11 + BITS(7);
+                        DROPBITS(7);
+                    }
+                    if (state->have + copy > state->nlen + state->ndist) {
+                        strm->msg = (char *)"invalid bit length repeat";
+                        state->mode = BAD;
+                        break;
+                    }
+                    while (copy--)
+                        state->lens[state->have++] = (unsigned short)len;
+                }
+            }
+
+            /* handle error breaks in while */
+            if (state->mode == BAD) break;
+
+            /* check for end-of-block code (better have one) */
+            if (state->lens[256] == 0) {
+                strm->msg = (char *)"invalid code -- missing end-of-block";
+                state->mode = BAD;
+                break;
+            }
+
+            /* build code tables -- note: do not change the lenbits or distbits
+               values here (9 and 6) without reading the comments in inftrees.h
+               concerning the ENOUGH constants, which depend on those values */
+            state->next = state->codes;
+            state->lencode = (const code FAR *)(state->next);
+            state->lenbits = 9;
+            ret = inflate_table(LENS, state->lens, state->nlen, &(state->next),
+                                &(state->lenbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid literal/lengths set";
+                state->mode = BAD;
+                break;
+            }
+            state->distcode = (const code FAR *)(state->next);
+            state->distbits = 6;
+            ret = inflate_table(DISTS, state->lens + state->nlen, state->ndist,
+                            &(state->next), &(state->distbits), state->work);
+            if (ret) {
+                strm->msg = (char *)"invalid distances set";
+                state->mode = BAD;
+                break;
+            }
+            Tracev((stderr, "inflate:       codes ok\n"));
+            state->mode = LEN_;
+            if (flush == Z_TREES) goto inf_leave;
+        case LEN_:
+            state->mode = LEN;
+        case LEN:
+            if (have >= 6 && left >= 258) {
+                RESTORE();
+                inflate_fast(strm, out);
+                LOAD();
+                if (state->mode == TYPE)
+                    state->back = -1;
+                break;
+            }
+            state->back = 0;
+            for (;;) {
+                here = state->lencode[BITS(state->lenbits)];
+                if ((unsigned)(here.bits) <= bits) break;
+                PULLBYTE();
+            }
+            if (here.op && (here.op & 0xf0) == 0) {
+                last = here;
+                for (;;) {
+                    here = state->lencode[last.val +
+                            (BITS(last.bits + last.op) >> last.bits)];
+                    if ((unsigned)(last.bits + here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                DROPBITS(last.bits);
+                state->back += last.bits;
+            }
+            DROPBITS(here.bits);
+            state->back += here.bits;
+            state->length = (unsigned)here.val;
+            if ((int)(here.op) == 0) {
+                Tracevv((stderr, here.val >= 0x20 && here.val < 0x7f ?
+                        "inflate:         literal '%c'\n" :
+                        "inflate:         literal 0x%02x\n", here.val));
+                state->mode = LIT;
+                break;
+            }
+            if (here.op & 32) {
+                Tracevv((stderr, "inflate:         end of block\n"));
+                state->back = -1;
+                state->mode = TYPE;
+                break;
+            }
+            if (here.op & 64) {
+                strm->msg = (char *)"invalid literal/length code";
+                state->mode = BAD;
+                break;
+            }
+            state->extra = (unsigned)(here.op) & 15;
+            state->mode = LENEXT;
+        case LENEXT:
+            if (state->extra) {
+                NEEDBITS(state->extra);
+                state->length += BITS(state->extra);
+                DROPBITS(state->extra);
+                state->back += state->extra;
+            }
+            Tracevv((stderr, "inflate:         length %u\n", state->length));
+            state->was = state->length;
+            state->mode = DIST;
+        case DIST:
+            for (;;) {
+                here = state->distcode[BITS(state->distbits)];
+                if ((unsigned)(here.bits) <= bits) break;
+                PULLBYTE();
+            }
+            if ((here.op & 0xf0) == 0) {
+                last = here;
+                for (;;) {
+                    here = state->distcode[last.val +
+                            (BITS(last.bits + last.op) >> last.bits)];
+                    if ((unsigned)(last.bits + here.bits) <= bits) break;
+                    PULLBYTE();
+                }
+                DROPBITS(last.bits);
+                state->back += last.bits;
+            }
+            DROPBITS(here.bits);
+            state->back += here.bits;
+            if (here.op & 64) {
+                strm->msg = (char *)"invalid distance code";
+                state->mode = BAD;
+                break;
+            }
+            state->offset = (unsigned)here.val;
+            state->extra = (unsigned)(here.op) & 15;
+            state->mode = DISTEXT;
+        case DISTEXT:
+            if (state->extra) {
+                NEEDBITS(state->extra);
+                state->offset += BITS(state->extra);
+                DROPBITS(state->extra);
+                state->back += state->extra;
+            }
+#ifdef INFLATE_STRICT
+            if (state->offset > state->dmax) {
+                strm->msg = (char *)"invalid distance too far back";
+                state->mode = BAD;
+                break;
+            }
+#endif
+            Tracevv((stderr, "inflate:         distance %u\n", state->offset));
+            state->mode = MATCH;
+        case MATCH:
+            if (left == 0) goto inf_leave;
+            copy = out - left;
+            if (state->offset > copy) {         /* copy from window */
+                copy = state->offset - copy;
+                if (copy > state->whave) {
+                    if (state->sane) {
+                        strm->msg = (char *)"invalid distance too far back";
+                        state->mode = BAD;
+                        break;
+                    }
+#ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR
+                    Trace((stderr, "inflate.c too far\n"));
+                    copy -= state->whave;
+                    if (copy > state->length) copy = state->length;
+                    if (copy > left) copy = left;
+                    left -= copy;
+                    state->length -= copy;
+                    do {
+                        *put++ = 0;
+                    } while (--copy);
+                    if (state->length == 0) state->mode = LEN;
+                    break;
+#endif
+                }
+                if (copy > state->wnext) {
+                    copy -= state->wnext;
+                    from = state->window + (state->wsize - copy);
+                }
+                else
+                    from = state->window + (state->wnext - copy);
+                if (copy > state->length) copy = state->length;
+            }
+            else {                              /* copy from output */
+                from = put - state->offset;
+                copy = state->length;
+            }
+            if (copy > left) copy = left;
+            left -= copy;
+            state->length -= copy;
+            do {
+                *put++ = *from++;
+            } while (--copy);
+            if (state->length == 0) state->mode = LEN;
+            break;
+        case LIT:
+            if (left == 0) goto inf_leave;
+            *put++ = (unsigned char)(state->length);
+            left--;
+            state->mode = LEN;
+            break;
+        case CHECK:
+            if (state->wrap) {
+                NEEDBITS(32);
+                out -= left;
+                strm->total_out += out;
+                state->total += out;
+                if (out)
+                    strm->adler = state->check =
+                        UPDATE(state->check, put - out, out);
+                out = left;
+                if ((
+#ifdef GUNZIP
+                     state->flags ? hold :
+#endif
+                     ZSWAP32(hold)) != state->check) {
+                    strm->msg = (char *)"incorrect data check";
+                    state->mode = BAD;
+                    break;
+                }
+                INITBITS();
+                Tracev((stderr, "inflate:   check matches trailer\n"));
+            }
+#ifdef GUNZIP
+            state->mode = LENGTH;
+        case LENGTH:
+            if (state->wrap && state->flags) {
+                NEEDBITS(32);
+                if (hold != (state->total & 0xffffffffUL)) {
+                    strm->msg = (char *)"incorrect length check";
+                    state->mode = BAD;
+                    break;
+                }
+                INITBITS();
+                Tracev((stderr, "inflate:   length matches trailer\n"));
+            }
+#endif
+            state->mode = DONE;
+        case DONE:
+            ret = Z_STREAM_END;
+            goto inf_leave;
+        case BAD:
+            ret = Z_DATA_ERROR;
+            goto inf_leave;
+        case MEM:
+            return Z_MEM_ERROR;
+        case SYNC:
+        default:
+            return Z_STREAM_ERROR;
+        }
+
+    /*
+       Return from inflate(), updating the total counts and the check value.
+       If there was no progress during the inflate() call, return a buffer
+       error.  Call updatewindow() to create and/or update the window state.
+       Note: a memory error from inflate() is non-recoverable.
+     */
+  inf_leave:
+    RESTORE();
+    if (state->wsize || (out != strm->avail_out && state->mode < BAD &&
+            (state->mode < CHECK || flush != Z_FINISH)))
+        if (updatewindow(strm, strm->next_out, out - strm->avail_out)) {
+            state->mode = MEM;
+            return Z_MEM_ERROR;
+        }
+    in -= strm->avail_in;
+    out -= strm->avail_out;
+    strm->total_in += in;
+    strm->total_out += out;
+    state->total += out;
+    if (state->wrap && out)
+        strm->adler = state->check =
+            UPDATE(state->check, strm->next_out - out, out);
+    strm->data_type = state->bits + (state->last ? 64 : 0) +
+                      (state->mode == TYPE ? 128 : 0) +
+                      (state->mode == LEN_ || state->mode == COPY_ ? 256 : 0);
+    if (((in == 0 && out == 0) || flush == Z_FINISH) && ret == Z_OK)
+        ret = Z_BUF_ERROR;
+    return ret;
+}
+
+int ZEXPORT inflateEnd(strm)
+z_streamp strm;
+{
+    struct inflate_state FAR *state;
+    if (strm == Z_NULL || strm->state == Z_NULL || strm->zfree == (free_func)0)
+        return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    if (state->window != Z_NULL) ZFREE(strm, state->window);
+    ZFREE(strm, strm->state);
+    strm->state = Z_NULL;
+    Tracev((stderr, "inflate: end\n"));
+    return Z_OK;
+}
+
+int ZEXPORT inflateGetDictionary(strm, dictionary, dictLength)
+z_streamp strm;
+Bytef *dictionary;
+uInt *dictLength;
+{
+    struct inflate_state FAR *state;
+
+    /* check state */
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+
+    /* copy dictionary */
+    if (state->whave && dictionary != Z_NULL) {
+        zmemcpy(dictionary, state->window + state->wnext,
+                state->whave - state->wnext);
+        zmemcpy(dictionary + state->whave - state->wnext,
+                state->window, state->wnext);
+    }
+    if (dictLength != Z_NULL)
+        *dictLength = state->whave;
+    return Z_OK;
+}
+
+int ZEXPORT inflateSetDictionary(strm, dictionary, dictLength)
+z_streamp strm;
+const Bytef *dictionary;
+uInt dictLength;
+{
+    struct inflate_state FAR *state;
+    unsigned long dictid;
+    int ret;
+
+    /* check state */
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    if (state->wrap != 0 && state->mode != DICT)
+        return Z_STREAM_ERROR;
+
+    /* check for correct dictionary identifier */
+    if (state->mode == DICT) {
+        dictid = adler32(0L, Z_NULL, 0);
+        dictid = adler32(dictid, dictionary, dictLength);
+        if (dictid != state->check)
+            return Z_DATA_ERROR;
+    }
+
+    /* copy dictionary to window using updatewindow(), which will amend the
+       existing dictionary if appropriate */
+    ret = updatewindow(strm, dictionary + dictLength, dictLength);
+    if (ret) {
+        state->mode = MEM;
+        return Z_MEM_ERROR;
+    }
+    state->havedict = 1;
+    Tracev((stderr, "inflate:   dictionary set\n"));
+    return Z_OK;
+}
+
+int ZEXPORT inflateGetHeader(strm, head)
+z_streamp strm;
+gz_headerp head;
+{
+    struct inflate_state FAR *state;
+
+    /* check state */
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    if ((state->wrap & 2) == 0) return Z_STREAM_ERROR;
+
+    /* save header structure */
+    state->head = head;
+    head->done = 0;
+    return Z_OK;
+}
+
+/*
+   Search buf[0..len-1] for the pattern: 0, 0, 0xff, 0xff.  Return when found
+   or when out of input.  When called, *have is the number of pattern bytes
+   found in order so far, in 0..3.  On return *have is updated to the new
+   state.  If on return *have equals four, then the pattern was found and the
+   return value is how many bytes were read including the last byte of the
+   pattern.  If *have is less than four, then the pattern has not been found
+   yet and the return value is len.  In the latter case, syncsearch() can be
+   called again with more data and the *have state.  *have is initialized to
+   zero for the first call.
+ */
+local unsigned syncsearch(have, buf, len)
+unsigned FAR *have;
+const unsigned char FAR *buf;
+unsigned len;
+{
+    unsigned got;
+    unsigned next;
+
+    got = *have;
+    next = 0;
+    while (next < len && got < 4) {
+        if ((int)(buf[next]) == (got < 2 ? 0 : 0xff))
+            got++;
+        else if (buf[next])
+            got = 0;
+        else
+            got = 4 - got;
+        next++;
+    }
+    *have = got;
+    return next;
+}
+
+int ZEXPORT inflateSync(strm)
+z_streamp strm;
+{
+    unsigned len;               /* number of bytes to look at or looked at */
+    unsigned long in, out;      /* temporary to save total_in and total_out */
+    unsigned char buf[4];       /* to restore bit buffer to byte string */
+    struct inflate_state FAR *state;
+
+    /* check parameters */
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    if (strm->avail_in == 0 && state->bits < 8) return Z_BUF_ERROR;
+
+    /* if first time, start search in bit buffer */
+    if (state->mode != SYNC) {
+        state->mode = SYNC;
+        state->hold <<= state->bits & 7;
+        state->bits -= state->bits & 7;
+        len = 0;
+        while (state->bits >= 8) {
+            buf[len++] = (unsigned char)(state->hold);
+            state->hold >>= 8;
+            state->bits -= 8;
+        }
+        state->have = 0;
+        syncsearch(&(state->have), buf, len);
+    }
+
+    /* search available input */
+    len = syncsearch(&(state->have), strm->next_in, strm->avail_in);
+    strm->avail_in -= len;
+    strm->next_in += len;
+    strm->total_in += len;
+
+    /* return no joy or set up to restart inflate() on a new block */
+    if (state->have != 4) return Z_DATA_ERROR;
+    in = strm->total_in;  out = strm->total_out;
+    inflateReset(strm);
+    strm->total_in = in;  strm->total_out = out;
+    state->mode = TYPE;
+    return Z_OK;
+}
+
+/*
+   Returns true if inflate is currently at the end of a block generated by
+   Z_SYNC_FLUSH or Z_FULL_FLUSH. This function is used by one PPP
+   implementation to provide an additional safety check. PPP uses
+   Z_SYNC_FLUSH but removes the length bytes of the resulting empty stored
+   block. When decompressing, PPP checks that at the end of input packet,
+   inflate is waiting for these length bytes.
+ */
+int ZEXPORT inflateSyncPoint(strm)
+z_streamp strm;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    return state->mode == STORED && state->bits == 0;
+}
+
+int ZEXPORT inflateCopy(dest, source)
+z_streamp dest;
+z_streamp source;
+{
+    struct inflate_state FAR *state;
+    struct inflate_state FAR *copy;
+    unsigned char FAR *window;
+    unsigned wsize;
+
+    /* check input */
+    if (dest == Z_NULL || source == Z_NULL || source->state == Z_NULL ||
+        source->zalloc == (alloc_func)0 || source->zfree == (free_func)0)
+        return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)source->state;
+
+    /* allocate space */
+    copy = (struct inflate_state FAR *)
+           ZALLOC(source, 1, sizeof(struct inflate_state));
+    if (copy == Z_NULL) return Z_MEM_ERROR;
+    window = Z_NULL;
+    if (state->window != Z_NULL) {
+        window = (unsigned char FAR *)
+                 ZALLOC(source, 1U << state->wbits, sizeof(unsigned char));
+        if (window == Z_NULL) {
+            ZFREE(source, copy);
+            return Z_MEM_ERROR;
+        }
+    }
+
+    /* copy state */
+    zmemcpy((voidpf)dest, (voidpf)source, sizeof(z_stream));
+    zmemcpy((voidpf)copy, (voidpf)state, sizeof(struct inflate_state));
+    if (state->lencode >= state->codes &&
+        state->lencode <= state->codes + ENOUGH - 1) {
+        copy->lencode = copy->codes + (state->lencode - state->codes);
+        copy->distcode = copy->codes + (state->distcode - state->codes);
+    }
+    copy->next = copy->codes + (state->next - state->codes);
+    if (window != Z_NULL) {
+        wsize = 1U << state->wbits;
+        zmemcpy(window, state->window, wsize);
+    }
+    copy->window = window;
+    dest->state = (struct internal_state FAR *)copy;
+    return Z_OK;
+}
+
+int ZEXPORT inflateUndermine(strm, subvert)
+z_streamp strm;
+int subvert;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
+    state = (struct inflate_state FAR *)strm->state;
+    state->sane = !subvert;
+#ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR
+    return Z_OK;
+#else
+    state->sane = 1;
+    return Z_DATA_ERROR;
+#endif
+}
+
+long ZEXPORT inflateMark(strm)
+z_streamp strm;
+{
+    struct inflate_state FAR *state;
+
+    if (strm == Z_NULL || strm->state == Z_NULL) return -1L << 16;
+    state = (struct inflate_state FAR *)strm->state;
+    return ((long)(state->back) << 16) +
+        (state->mode == COPY ? state->length :
+            (state->mode == MATCH ? state->was - state->length : 0));
+}
diff --git a/ml/dlib/dlib/external/zlib/inflate.h b/ml/dlib/dlib/external/zlib/inflate.h
new file mode 100644
index 000000000..95f4986d4
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inflate.h
@@ -0,0 +1,122 @@
+/* inflate.h -- internal inflate state definition
+ * Copyright (C) 1995-2009 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* WARNING: this file should *not* be used by applications. It is
+   part of the implementation of the compression library and is
+   subject to change. Applications should only use zlib.h.
+ */
+
+/* define NO_GZIP when compiling if you want to disable gzip header and
+   trailer decoding by inflate().  NO_GZIP would be used to avoid linking in
+   the crc code when it is not needed.  For shared libraries, gzip decoding
+   should be left enabled. */
+#ifndef NO_GZIP
+#  define GUNZIP
+#endif
+
+/* Possible inflate modes between inflate() calls */
+typedef enum {
+    HEAD,       /* i: waiting for magic header */
+    FLAGS,      /* i: waiting for method and flags (gzip) */
+    TIME,       /* i: waiting for modification time (gzip) */
+    OS,         /* i: waiting for extra flags and operating system (gzip) */
+    EXLEN,      /* i: waiting for extra length (gzip) */
+    EXTRA,      /* i: waiting for extra bytes (gzip) */
+    NAME,       /* i: waiting for end of file name (gzip) */
+    COMMENT,    /* i: waiting for end of comment (gzip) */
+    HCRC,       /* i: waiting for header crc (gzip) */
+    DICTID,     /* i: waiting for dictionary check value */
+    DICT,       /* waiting for inflateSetDictionary() call */
+        TYPE,       /* i: waiting for type bits, including last-flag bit */
+        TYPEDO,     /* i: same, but skip check to exit inflate on new block */
+        STORED,     /* i: waiting for stored size (length and complement) */
+        COPY_,      /* i/o: same as COPY below, but only first time in */
+        COPY,       /* i/o: waiting for input or output to copy stored block */
+        TABLE,      /* i: waiting for dynamic block table lengths */
+        LENLENS,    /* i: waiting for code length code lengths */
+        CODELENS,   /* i: waiting for length/lit and distance code lengths */
+            LEN_,       /* i: same as LEN below, but only first time in */
+            LEN,        /* i: waiting for length/lit/eob code */
+            LENEXT,     /* i: waiting for length extra bits */
+            DIST,       /* i: waiting for distance code */
+            DISTEXT,    /* i: waiting for distance extra bits */
+            MATCH,      /* o: waiting for output space to copy string */
+            LIT,        /* o: waiting for output space to write literal */
+    CHECK,      /* i: waiting for 32-bit check value */
+    LENGTH,     /* i: waiting for 32-bit length (gzip) */
+    DONE,       /* finished check, done -- remain here until reset */
+    BAD,        /* got a data error -- remain here until reset */
+    MEM,        /* got an inflate() memory error -- remain here until reset */
+    SYNC        /* looking for synchronization bytes to restart inflate() */
+} inflate_mode;
+
+/*
+    State transitions between above modes -
+
+    (most modes can go to BAD or MEM on error -- not shown for clarity)
+
+    Process header:
+        HEAD -> (gzip) or (zlib) or (raw)
+        (gzip) -> FLAGS -> TIME -> OS -> EXLEN -> EXTRA -> NAME -> COMMENT ->
+                  HCRC -> TYPE
+        (zlib) -> DICTID or TYPE
+        DICTID -> DICT -> TYPE
+        (raw) -> TYPEDO
+    Read deflate blocks:
+            TYPE -> TYPEDO -> STORED or TABLE or LEN_ or CHECK
+            STORED -> COPY_ -> COPY -> TYPE
+            TABLE -> LENLENS -> CODELENS -> LEN_
+            LEN_ -> LEN
+    Read deflate codes in fixed or dynamic block:
+                LEN -> LENEXT or LIT or TYPE
+                LENEXT -> DIST -> DISTEXT -> MATCH -> LEN
+                LIT -> LEN
+    Process trailer:
+        CHECK -> LENGTH -> DONE
+ */
+
+/* state maintained between inflate() calls.  Approximately 10K bytes. */
+struct inflate_state {
+    inflate_mode mode;          /* current inflate mode */
+    int last;                   /* true if processing last block */
+    int wrap;                   /* bit 0 true for zlib, bit 1 true for gzip */
+    int havedict;               /* true if dictionary provided */
+    int flags;                  /* gzip header method and flags (0 if zlib) */
+    unsigned dmax;              /* zlib header max distance (INFLATE_STRICT) */
+    unsigned long check;        /* protected copy of check value */
+    unsigned long total;        /* protected copy of output count */
+    gz_headerp head;            /* where to save gzip header information */
+        /* sliding window */
+    unsigned wbits;             /* log base 2 of requested window size */
+    unsigned wsize;             /* window size or zero if not using window */
+    unsigned whave;             /* valid bytes in the window */
+    unsigned wnext;             /* window write index */
+    unsigned char FAR *window;  /* allocated sliding window, if needed */
+        /* bit accumulator */
+    unsigned long hold;         /* input bit accumulator */
+    unsigned bits;              /* number of bits in "in" */
+        /* for string and stored block copying */
+    unsigned length;            /* literal or length of data to copy */
+    unsigned offset;            /* distance back to copy string from */
+        /* for table and code decoding */
+    unsigned extra;             /* extra bits needed */
+        /* fixed and dynamic code tables */
+    code const FAR *lencode;    /* starting table for length/literal codes */
+    code const FAR *distcode;   /* starting table for distance codes */
+    unsigned lenbits;           /* index bits for lencode */
+    unsigned distbits;          /* index bits for distcode */
+        /* dynamic table building */
+    unsigned ncode;             /* number of code length code lengths */
+    unsigned nlen;              /* number of length code lengths */
+    unsigned ndist;             /* number of distance code lengths */
+    unsigned have;              /* number of code lengths in lens[] */
+    code FAR *next;             /* next available space in codes[] */
+    unsigned short lens[320];   /* temporary storage for code lengths */
+    unsigned short work[288];   /* work area for code table building */
+    code codes[ENOUGH];         /* space for code tables */
+    int sane;                   /* if false, allow invalid distance too far */
+    int back;                   /* bits back of last unprocessed length/lit */
+    unsigned was;               /* initial length of match */
+};
diff --git a/ml/dlib/dlib/external/zlib/inftrees.c b/ml/dlib/dlib/external/zlib/inftrees.c
new file mode 100644
index 000000000..44d89cf24
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inftrees.c
@@ -0,0 +1,306 @@
+/* inftrees.c -- generate Huffman trees for efficient decoding
+ * Copyright (C) 1995-2013 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "zutil.h"
+#include "inftrees.h"
+
+#define MAXBITS 15
+
+const char inflate_copyright[] =
+   " inflate 1.2.8 Copyright 1995-2013 Mark Adler ";
+/*
+  If you use the zlib library in a product, an acknowledgment is welcome
+  in the documentation of your product. If for some reason you cannot
+  include such an acknowledgment, I would appreciate that you keep this
+  copyright string in the executable of your product.
+ */
+
+/*
+   Build a set of tables to decode the provided canonical Huffman code.
+   The code lengths are lens[0..codes-1].  The result starts at *table,
+   whose indices are 0..2^bits-1.  work is a writable array of at least
+   lens shorts, which is used as a work area.  type is the type of code
+   to be generated, CODES, LENS, or DISTS.  On return, zero is success,
+   -1 is an invalid code, and +1 means that ENOUGH isn't enough.  table
+   on return points to the next available entry's address.  bits is the
+   requested root table index bits, and on return it is the actual root
+   table index bits.  It will differ if the request is greater than the
+   longest code or if it is less than the shortest code.
+ */
+int ZLIB_INTERNAL inflate_table(type, lens, codes, table, bits, work)
+codetype type;
+unsigned short FAR *lens;
+unsigned codes;
+code FAR * FAR *table;
+unsigned FAR *bits;
+unsigned short FAR *work;
+{
+    unsigned len;               /* a code's length in bits */
+    unsigned sym;               /* index of code symbols */
+    unsigned min, max;          /* minimum and maximum code lengths */
+    unsigned root;              /* number of index bits for root table */
+    unsigned curr;              /* number of index bits for current table */
+    unsigned drop;              /* code bits to drop for sub-table */
+    int left;                   /* number of prefix codes available */
+    unsigned used;              /* code entries in table used */
+    unsigned huff;              /* Huffman code */
+    unsigned incr;              /* for incrementing code, index */
+    unsigned fill;              /* index for replicating entries */
+    unsigned low;               /* low bits for current root entry */
+    unsigned mask;              /* mask for low root bits */
+    code here;                  /* table entry for duplication */
+    code FAR *next;             /* next available space in table */
+    const unsigned short FAR *base;     /* base value table to use */
+    const unsigned short FAR *extra;    /* extra bits table to use */
+    int end;                    /* use base and extra for symbol > end */
+    unsigned short count[MAXBITS+1];    /* number of codes of each length */
+    unsigned short offs[MAXBITS+1];     /* offsets in table for each length */
+    static const unsigned short lbase[31] = { /* Length codes 257..285 base */
+        3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31,
+        35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0, 0};
+    static const unsigned short lext[31] = { /* Length codes 257..285 extra */
+        16, 16, 16, 16, 16, 16, 16, 16, 17, 17, 17, 17, 18, 18, 18, 18,
+        19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 16, 72, 78};
+    static const unsigned short dbase[32] = { /* Distance codes 0..29 base */
+        1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193,
+        257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145,
+        8193, 12289, 16385, 24577, 0, 0};
+    static const unsigned short dext[32] = { /* Distance codes 0..29 extra */
+        16, 16, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 21, 21, 22, 22,
+        23, 23, 24, 24, 25, 25, 26, 26, 27, 27,
+        28, 28, 29, 29, 64, 64};
+
+    /*
+       Process a set of code lengths to create a canonical Huffman code.  The
+       code lengths are lens[0..codes-1].  Each length corresponds to the
+       symbols 0..codes-1.  The Huffman code is generated by first sorting the
+       symbols by length from short to long, and retaining the symbol order
+       for codes with equal lengths.  Then the code starts with all zero bits
+       for the first code of the shortest length, and the codes are integer
+       increments for the same length, and zeros are appended as the length
+       increases.  For the deflate format, these bits are stored backwards
+       from their more natural integer increment ordering, and so when the
+       decoding tables are built in the large loop below, the integer codes
+       are incremented backwards.
+
+       This routine assumes, but does not check, that all of the entries in
+       lens[] are in the range 0..MAXBITS.  The caller must assure this.
+       1..MAXBITS is interpreted as that code length.  zero means that that
+       symbol does not occur in this code.
+
+       The codes are sorted by computing a count of codes for each length,
+       creating from that a table of starting indices for each length in the
+       sorted table, and then entering the symbols in order in the sorted
+       table.  The sorted table is work[], with that space being provided by
+       the caller.
+
+       The length counts are used for other purposes as well, i.e. finding
+       the minimum and maximum length codes, determining if there are any
+       codes at all, checking for a valid set of lengths, and looking ahead
+       at length counts to determine sub-table sizes when building the
+       decoding tables.
+     */
+
+    /* accumulate lengths for codes (assumes lens[] all in 0..MAXBITS) */
+    for (len = 0; len <= MAXBITS; len++)
+        count[len] = 0;
+    for (sym = 0; sym < codes; sym++)
+        count[lens[sym]]++;
+
+    /* bound code lengths, force root to be within code lengths */
+    root = *bits;
+    for (max = MAXBITS; max >= 1; max--)
+        if (count[max] != 0) break;
+    if (root > max) root = max;
+    if (max == 0) {                     /* no symbols to code at all */
+        here.op = (unsigned char)64;    /* invalid code marker */
+        here.bits = (unsigned char)1;
+        here.val = (unsigned short)0;
+        *(*table)++ = here;             /* make a table to force an error */
+        *(*table)++ = here;
+        *bits = 1;
+        return 0;     /* no symbols, but wait for decoding to report error */
+    }
+    for (min = 1; min < max; min++)
+        if (count[min] != 0) break;
+    if (root < min) root = min;
+
+    /* check for an over-subscribed or incomplete set of lengths */
+    left = 1;
+    for (len = 1; len <= MAXBITS; len++) {
+        left <<= 1;
+        left -= count[len];
+        if (left < 0) return -1;        /* over-subscribed */
+    }
+    if (left > 0 && (type == CODES || max != 1))
+        return -1;                      /* incomplete set */
+
+    /* generate offsets into symbol table for each length for sorting */
+    offs[1] = 0;
+    for (len = 1; len < MAXBITS; len++)
+        offs[len + 1] = offs[len] + count[len];
+
+    /* sort symbols by length, by symbol order within each length */
+    for (sym = 0; sym < codes; sym++)
+        if (lens[sym] != 0) work[offs[lens[sym]]++] = (unsigned short)sym;
+
+    /*
+       Create and fill in decoding tables.  In this loop, the table being
+       filled is at next and has curr index bits.  The code being used is huff
+       with length len.  That code is converted to an index by dropping drop
+       bits off of the bottom.  For codes where len is less than drop + curr,
+       those top drop + curr - len bits are incremented through all values to
+       fill the table with replicated entries.
+
+       root is the number of index bits for the root table.  When len exceeds
+       root, sub-tables are created pointed to by the root entry with an index
+       of the low root bits of huff.  This is saved in low to check for when a
+       new sub-table should be started.  drop is zero when the root table is
+       being filled, and drop is root when sub-tables are being filled.
+
+       When a new sub-table is needed, it is necessary to look ahead in the
+       code lengths to determine what size sub-table is needed.  The length
+       counts are used for this, and so count[] is decremented as codes are
+       entered in the tables.
+
+       used keeps track of how many table entries have been allocated from the
+       provided *table space.  It is checked for LENS and DIST tables against
+       the constants ENOUGH_LENS and ENOUGH_DISTS to guard against changes in
+       the initial root table size constants.  See the comments in inftrees.h
+       for more information.
+
+       sym increments through all symbols, and the loop terminates when
+       all codes of length max, i.e. all codes, have been processed.  This
+       routine permits incomplete codes, so another loop after this one fills
+       in the rest of the decoding tables with invalid code markers.
+     */
+
+    /* set up for code type */
+    switch (type) {
+    case CODES:
+        base = extra = work;    /* dummy value--not used */
+        end = 19;
+        break;
+    case LENS:
+        base = lbase;
+        base -= 257;
+        extra = lext;
+        extra -= 257;
+        end = 256;
+        break;
+    default:            /* DISTS */
+        base = dbase;
+        extra = dext;
+        end = -1;
+    }
+
+    /* initialize state for loop */
+    huff = 0;                   /* starting code */
+    sym = 0;                    /* starting code symbol */
+    len = min;                  /* starting code length */
+    next = *table;              /* current table to fill in */
+    curr = root;                /* current table index bits */
+    drop = 0;                   /* current bits to drop from code for index */
+    low = (unsigned)(-1);       /* trigger new sub-table when len > root */
+    used = 1U << root;          /* use root table entries */
+    mask = used - 1;            /* mask for comparing low */
+
+    /* check available table space */
+    if ((type == LENS && used > ENOUGH_LENS) ||
+        (type == DISTS && used > ENOUGH_DISTS))
+        return 1;
+
+    /* process all codes and make table entries */
+    for (;;) {
+        /* create table entry */
+        here.bits = (unsigned char)(len - drop);
+        if ((int)(work[sym]) < end) {
+            here.op = (unsigned char)0;
+            here.val = work[sym];
+        }
+        else if ((int)(work[sym]) > end) {
+            here.op = (unsigned char)(extra[work[sym]]);
+            here.val = base[work[sym]];
+        }
+        else {
+            here.op = (unsigned char)(32 + 64);         /* end of block */
+            here.val = 0;
+        }
+
+        /* replicate for those indices with low len bits equal to huff */
+        incr = 1U << (len - drop);
+        fill = 1U << curr;
+        min = fill;                 /* save offset to next table */
+        do {
+            fill -= incr;
+            next[(huff >> drop) + fill] = here;
+        } while (fill != 0);
+
+        /* backwards increment the len-bit code huff */
+        incr = 1U << (len - 1);
+        while (huff & incr)
+            incr >>= 1;
+        if (incr != 0) {
+            huff &= incr - 1;
+            huff += incr;
+        }
+        else
+            huff = 0;
+
+        /* go to next symbol, update count, len */
+        sym++;
+        if (--(count[len]) == 0) {
+            if (len == max) break;
+            len = lens[work[sym]];
+        }
+
+        /* create new sub-table if needed */
+        if (len > root && (huff & mask) != low) {
+            /* if first time, transition to sub-tables */
+            if (drop == 0)
+                drop = root;
+
+            /* increment past last table */
+            next += min;            /* here min is 1 << curr */
+
+            /* determine length of next table */
+            curr = len - drop;
+            left = (int)(1 << curr);
+            while (curr + drop < max) {
+                left -= count[curr + drop];
+                if (left <= 0) break;
+                curr++;
+                left <<= 1;
+            }
+
+            /* check for enough space */
+            used += 1U << curr;
+            if ((type == LENS && used > ENOUGH_LENS) ||
+                (type == DISTS && used > ENOUGH_DISTS))
+                return 1;
+
+            /* point entry in root table to sub-table */
+            low = huff & mask;
+            (*table)[low].op = (unsigned char)curr;
+            (*table)[low].bits = (unsigned char)root;
+            (*table)[low].val = (unsigned short)(next - *table);
+        }
+    }
+
+    /* fill in remaining table entry if code is incomplete (guaranteed to have
+       at most one remaining entry, since if the code is incomplete, the
+       maximum code length that was allowed to get this far is one bit) */
+    if (huff != 0) {
+        here.op = (unsigned char)64;            /* invalid code marker */
+        here.bits = (unsigned char)(len - drop);
+        here.val = (unsigned short)0;
+        next[huff] = here;
+    }
+
+    /* set return parameters */
+    *table += used;
+    *bits = root;
+    return 0;
+}
diff --git a/ml/dlib/dlib/external/zlib/inftrees.h b/ml/dlib/dlib/external/zlib/inftrees.h
new file mode 100644
index 000000000..baa53a0b1
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/inftrees.h
@@ -0,0 +1,62 @@
+/* inftrees.h -- header to use inftrees.c
+ * Copyright (C) 1995-2005, 2010 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* WARNING: this file should *not* be used by applications. It is
+   part of the implementation of the compression library and is
+   subject to change. Applications should only use zlib.h.
+ */
+
+/* Structure for decoding tables.  Each entry provides either the
+   information needed to do the operation requested by the code that
+   indexed that table entry, or it provides a pointer to another
+   table that indexes more bits of the code.  op indicates whether
+   the entry is a pointer to another table, a literal, a length or
+   distance, an end-of-block, or an invalid code.  For a table
+   pointer, the low four bits of op is the number of index bits of
+   that table.  For a length or distance, the low four bits of op
+   is the number of extra bits to get after the code.  bits is
+   the number of bits in this code or part of the code to drop off
+   of the bit buffer.  val is the actual byte to output in the case
+   of a literal, the base length or distance, or the offset from
+   the current table to the next table.  Each entry is four bytes. */
+typedef struct {
+    unsigned char op;           /* operation, extra bits, table bits */
+    unsigned char bits;         /* bits in this part of the code */
+    unsigned short val;         /* offset in table or code value */
+} code;
+
+/* op values as set by inflate_table():
+    00000000 - literal
+    0000tttt - table link, tttt != 0 is the number of table index bits
+    0001eeee - length or distance, eeee is the number of extra bits
+    01100000 - end of block
+    01000000 - invalid code
+ */
+
+/* Maximum size of the dynamic table.  The maximum number of code structures is
+   1444, which is the sum of 852 for literal/length codes and 592 for distance
+   codes.  These values were found by exhaustive searches using the program
+   examples/enough.c found in the zlib distribtution.  The arguments to that
+   program are the number of symbols, the initial root table size, and the
+   maximum bit length of a code.  "enough 286 9 15" for literal/length codes
+   returns returns 852, and "enough 30 6 15" for distance codes returns 592.
+   The initial root table size (9 or 6) is found in the fifth argument of the
+   inflate_table() calls in inflate.c and infback.c.  If the root table size is
+   changed, then these maximum sizes would be need to be recalculated and
+   updated. */
+#define ENOUGH_LENS 852
+#define ENOUGH_DISTS 592
+#define ENOUGH (ENOUGH_LENS+ENOUGH_DISTS)
+
+/* Type of code to build for inflate_table() */
+typedef enum {
+    CODES,
+    LENS,
+    DISTS
+} codetype;
+
+int ZLIB_INTERNAL inflate_table OF((codetype type, unsigned short FAR *lens,
+                             unsigned codes, code FAR * FAR *table,
+                             unsigned FAR *bits, unsigned short FAR *work));
diff --git a/ml/dlib/dlib/external/zlib/trees.c b/ml/dlib/dlib/external/zlib/trees.c
new file mode 100644
index 000000000..1fd7759ef
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/trees.c
@@ -0,0 +1,1226 @@
+/* trees.c -- output deflated data using Huffman coding
+ * Copyright (C) 1995-2012 Jean-loup Gailly
+ * detect_data_type() function provided freely by Cosmin Truta, 2006
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/*
+ *  ALGORITHM
+ *
+ *      The "deflation" process uses several Huffman trees. The more
+ *      common source values are represented by shorter bit sequences.
+ *
+ *      Each code tree is stored in a compressed form which is itself
+ * a Huffman encoding of the lengths of all the code strings (in
+ * ascending order by source values).  The actual code strings are
+ * reconstructed from the lengths in the inflate process, as described
+ * in the deflate specification.
+ *
+ *  REFERENCES
+ *
+ *      Deutsch, L.P.,"'Deflate' Compressed Data Format Specification".
+ *      Available in ftp.uu.net:/pub/archiving/zip/doc/deflate-1.1.doc
+ *
+ *      Storer, James A.
+ *          Data Compression:  Methods and Theory, pp. 49-50.
+ *          Computer Science Press, 1988.  ISBN 0-7167-8156-5.
+ *
+ *      Sedgewick, R.
+ *          Algorithms, p290.
+ *          Addison-Wesley, 1983. ISBN 0-201-06672-6.
+ */
+
+/* @(#) $Id$ */
+
+/* #define GEN_TREES_H */
+
+#include "deflate.h"
+
+#ifdef DEBUG
+#  include <ctype.h>
+#endif
+
+/* ===========================================================================
+ * Constants
+ */
+
+#define MAX_BL_BITS 7
+/* Bit length codes must not exceed MAX_BL_BITS bits */
+
+#define END_BLOCK 256
+/* end of block literal code */
+
+#define REP_3_6      16
+/* repeat previous bit length 3-6 times (2 bits of repeat count) */
+
+#define REPZ_3_10    17
+/* repeat a zero length 3-10 times  (3 bits of repeat count) */
+
+#define REPZ_11_138  18
+/* repeat a zero length 11-138 times  (7 bits of repeat count) */
+
+local const int extra_lbits[LENGTH_CODES] /* extra bits for each length code */
+   = {0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0};
+
+local const int extra_dbits[D_CODES] /* extra bits for each distance code */
+   = {0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13};
+
+local const int extra_blbits[BL_CODES]/* extra bits for each bit length code */
+   = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,2,3,7};
+
+local const uch bl_order[BL_CODES]
+   = {16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15};
+/* The lengths of the bit length codes are sent in order of decreasing
+ * probability, to avoid transmitting the lengths for unused bit length codes.
+ */
+
+/* ===========================================================================
+ * Local data. These are initialized only once.
+ */
+
+#define DIST_CODE_LEN  512 /* see definition of array dist_code below */
+
+#if defined(GEN_TREES_H) || !defined(STDC)
+/* non ANSI compilers may not accept trees.h */
+
+local ct_data static_ltree[L_CODES+2];
+/* The static literal tree. Since the bit lengths are imposed, there is no
+ * need for the L_CODES extra codes used during heap construction. However
+ * The codes 286 and 287 are needed to build a canonical tree (see _tr_init
+ * below).
+ */
+
+local ct_data static_dtree[D_CODES];
+/* The static distance tree. (Actually a trivial tree since all codes use
+ * 5 bits.)
+ */
+
+uch _dist_code[DIST_CODE_LEN];
+/* Distance codes. The first 256 values correspond to the distances
+ * 3 .. 258, the last 256 values correspond to the top 8 bits of
+ * the 15 bit distances.
+ */
+
+uch _length_code[MAX_MATCH-MIN_MATCH+1];
+/* length code for each normalized match length (0 == MIN_MATCH) */
+
+local int base_length[LENGTH_CODES];
+/* First normalized length for each code (0 = MIN_MATCH) */
+
+local int base_dist[D_CODES];
+/* First normalized distance for each code (0 = distance of 1) */
+
+#else
+#  include "trees.h"
+#endif /* GEN_TREES_H */
+
+struct static_tree_desc_s {
+    const ct_data *static_tree;  /* static tree or NULL */
+    const intf *extra_bits;      /* extra bits for each code or NULL */
+    int     extra_base;          /* base index for extra_bits */
+    int     elems;               /* max number of elements in the tree */
+    int     max_length;          /* max bit length for the codes */
+};
+
+local static_tree_desc  static_l_desc =
+{static_ltree, extra_lbits, LITERALS+1, L_CODES, MAX_BITS};
+
+local static_tree_desc  static_d_desc =
+{static_dtree, extra_dbits, 0,          D_CODES, MAX_BITS};
+
+local static_tree_desc  static_bl_desc =
+{(const ct_data *)0, extra_blbits, 0,   BL_CODES, MAX_BL_BITS};
+
+/* ===========================================================================
+ * Local (static) routines in this file.
+ */
+
+local void tr_static_init OF((void));
+local void init_block     OF((deflate_state *s));
+local void pqdownheap     OF((deflate_state *s, ct_data *tree, int k));
+local void gen_bitlen     OF((deflate_state *s, tree_desc *desc));
+local void gen_codes      OF((ct_data *tree, int max_code, ushf *bl_count));
+local void build_tree     OF((deflate_state *s, tree_desc *desc));
+local void scan_tree      OF((deflate_state *s, ct_data *tree, int max_code));
+local void send_tree      OF((deflate_state *s, ct_data *tree, int max_code));
+local int  build_bl_tree  OF((deflate_state *s));
+local void send_all_trees OF((deflate_state *s, int lcodes, int dcodes,
+                              int blcodes));
+local void compress_block OF((deflate_state *s, const ct_data *ltree,
+                              const ct_data *dtree));
+local int  detect_data_type OF((deflate_state *s));
+local unsigned bi_reverse OF((unsigned value, int length));
+local void bi_windup      OF((deflate_state *s));
+local void bi_flush       OF((deflate_state *s));
+local void copy_block     OF((deflate_state *s, charf *buf, unsigned len,
+                              int header));
+
+#ifdef GEN_TREES_H
+local void gen_trees_header OF((void));
+#endif
+
+#ifndef DEBUG
+#  define send_code(s, c, tree) send_bits(s, tree[c].Code, tree[c].Len)
+   /* Send a code of the given tree. c and tree must not have side effects */
+
+#else /* DEBUG */
+#  define send_code(s, c, tree) \
+     { if (z_verbose>2) fprintf(stderr,"\ncd %3d ",(c)); \
+       send_bits(s, tree[c].Code, tree[c].Len); }
+#endif
+
+/* ===========================================================================
+ * Output a short LSB first on the stream.
+ * IN assertion: there is enough room in pendingBuf.
+ */
+#define put_short(s, w) { \
+    put_byte(s, (uch)((w) & 0xff)); \
+    put_byte(s, (uch)((ush)(w) >> 8)); \
+}
+
+/* ===========================================================================
+ * Send a value on a given number of bits.
+ * IN assertion: length <= 16 and value fits in length bits.
+ */
+#ifdef DEBUG
+local void send_bits      OF((deflate_state *s, int value, int length));
+
+local void send_bits(s, value, length)
+    deflate_state *s;
+    int value;  /* value to send */
+    int length; /* number of bits */
+{
+    Tracevv((stderr," l %2d v %4x ", length, value));
+    Assert(length > 0 && length <= 15, "invalid length");
+    s->bits_sent += (ulg)length;
+
+    /* If not enough room in bi_buf, use (valid) bits from bi_buf and
+     * (16 - bi_valid) bits from value, leaving (width - (16-bi_valid))
+     * unused bits in value.
+     */
+    if (s->bi_valid > (int)Buf_size - length) {
+        s->bi_buf |= (ush)value << s->bi_valid;
+        put_short(s, s->bi_buf);
+        s->bi_buf = (ush)value >> (Buf_size - s->bi_valid);
+        s->bi_valid += length - Buf_size;
+    } else {
+        s->bi_buf |= (ush)value << s->bi_valid;
+        s->bi_valid += length;
+    }
+}
+#else /* !DEBUG */
+
+#define send_bits(s, value, length) \
+{ int len = length;\
+  if (s->bi_valid > (int)Buf_size - len) {\
+    int val = value;\
+    s->bi_buf |= (ush)val << s->bi_valid;\
+    put_short(s, s->bi_buf);\
+    s->bi_buf = (ush)val >> (Buf_size - s->bi_valid);\
+    s->bi_valid += len - Buf_size;\
+  } else {\
+    s->bi_buf |= (ush)(value) << s->bi_valid;\
+    s->bi_valid += len;\
+  }\
+}
+#endif /* DEBUG */
+
+
+/* the arguments must not have side effects */
+
+/* ===========================================================================
+ * Initialize the various 'constant' tables.
+ */
+local void tr_static_init()
+{
+#if defined(GEN_TREES_H) || !defined(STDC)
+    static int static_init_done = 0;
+    int n;        /* iterates over tree elements */
+    int bits;     /* bit counter */
+    int length;   /* length value */
+    int code;     /* code value */
+    int dist;     /* distance index */
+    ush bl_count[MAX_BITS+1];
+    /* number of codes at each bit length for an optimal tree */
+
+    if (static_init_done) return;
+
+    /* For some embedded targets, global variables are not initialized: */
+#ifdef NO_INIT_GLOBAL_POINTERS
+    static_l_desc.static_tree = static_ltree;
+    static_l_desc.extra_bits = extra_lbits;
+    static_d_desc.static_tree = static_dtree;
+    static_d_desc.extra_bits = extra_dbits;
+    static_bl_desc.extra_bits = extra_blbits;
+#endif
+
+    /* Initialize the mapping length (0..255) -> length code (0..28) */
+    length = 0;
+    for (code = 0; code < LENGTH_CODES-1; code++) {
+        base_length[code] = length;
+        for (n = 0; n < (1<<extra_lbits[code]); n++) {
+            _length_code[length++] = (uch)code;
+        }
+    }
+    Assert (length == 256, "tr_static_init: length != 256");
+    /* Note that the length 255 (match length 258) can be represented
+     * in two different ways: code 284 + 5 bits or code 285, so we
+     * overwrite length_code[255] to use the best encoding:
+     */
+    _length_code[length-1] = (uch)code;
+
+    /* Initialize the mapping dist (0..32K) -> dist code (0..29) */
+    dist = 0;
+    for (code = 0 ; code < 16; code++) {
+        base_dist[code] = dist;
+        for (n = 0; n < (1<<extra_dbits[code]); n++) {
+            _dist_code[dist++] = (uch)code;
+        }
+    }
+    Assert (dist == 256, "tr_static_init: dist != 256");
+    dist >>= 7; /* from now on, all distances are divided by 128 */
+    for ( ; code < D_CODES; code++) {
+        base_dist[code] = dist << 7;
+        for (n = 0; n < (1<<(extra_dbits[code]-7)); n++) {
+            _dist_code[256 + dist++] = (uch)code;
+        }
+    }
+    Assert (dist == 256, "tr_static_init: 256+dist != 512");
+
+    /* Construct the codes of the static literal tree */
+    for (bits = 0; bits <= MAX_BITS; bits++) bl_count[bits] = 0;
+    n = 0;
+    while (n <= 143) static_ltree[n++].Len = 8, bl_count[8]++;
+    while (n <= 255) static_ltree[n++].Len = 9, bl_count[9]++;
+    while (n <= 279) static_ltree[n++].Len = 7, bl_count[7]++;
+    while (n <= 287) static_ltree[n++].Len = 8, bl_count[8]++;
+    /* Codes 286 and 287 do not exist, but we must include them in the
+     * tree construction to get a canonical Huffman tree (longest code
+     * all ones)
+     */
+    gen_codes((ct_data *)static_ltree, L_CODES+1, bl_count);
+
+    /* The static distance tree is trivial: */
+    for (n = 0; n < D_CODES; n++) {
+        static_dtree[n].Len = 5;
+        static_dtree[n].Code = bi_reverse((unsigned)n, 5);
+    }
+    static_init_done = 1;
+
+#  ifdef GEN_TREES_H
+    gen_trees_header();
+#  endif
+#endif /* defined(GEN_TREES_H) || !defined(STDC) */
+}
+
+/* ===========================================================================
+ * Genererate the file trees.h describing the static trees.
+ */
+#ifdef GEN_TREES_H
+#  ifndef DEBUG
+#    include <stdio.h>
+#  endif
+
+#  define SEPARATOR(i, last, width) \
+      ((i) == (last)? "\n};\n\n" :    \
+       ((i) % (width) == (width)-1 ? ",\n" : ", "))
+
+void gen_trees_header()
+{
+    FILE *header = fopen("trees.h", "w");
+    int i;
+
+    Assert (header != NULL, "Can't open trees.h");
+    fprintf(header,
+            "/* header created automatically with -DGEN_TREES_H */\n\n");
+
+    fprintf(header, "local const ct_data static_ltree[L_CODES+2] = {\n");
+    for (i = 0; i < L_CODES+2; i++) {
+        fprintf(header, "{{%3u},{%3u}}%s", static_ltree[i].Code,
+                static_ltree[i].Len, SEPARATOR(i, L_CODES+1, 5));
+    }
+
+    fprintf(header, "local const ct_data static_dtree[D_CODES] = {\n");
+    for (i = 0; i < D_CODES; i++) {
+        fprintf(header, "{{%2u},{%2u}}%s", static_dtree[i].Code,
+                static_dtree[i].Len, SEPARATOR(i, D_CODES-1, 5));
+    }
+
+    fprintf(header, "const uch ZLIB_INTERNAL _dist_code[DIST_CODE_LEN] = {\n");
+    for (i = 0; i < DIST_CODE_LEN; i++) {
+        fprintf(header, "%2u%s", _dist_code[i],
+                SEPARATOR(i, DIST_CODE_LEN-1, 20));
+    }
+
+    fprintf(header,
+        "const uch ZLIB_INTERNAL _length_code[MAX_MATCH-MIN_MATCH+1]= {\n");
+    for (i = 0; i < MAX_MATCH-MIN_MATCH+1; i++) {
+        fprintf(header, "%2u%s", _length_code[i],
+                SEPARATOR(i, MAX_MATCH-MIN_MATCH, 20));
+    }
+
+    fprintf(header, "local const int base_length[LENGTH_CODES] = {\n");
+    for (i = 0; i < LENGTH_CODES; i++) {
+        fprintf(header, "%1u%s", base_length[i],
+                SEPARATOR(i, LENGTH_CODES-1, 20));
+    }
+
+    fprintf(header, "local const int base_dist[D_CODES] = {\n");
+    for (i = 0; i < D_CODES; i++) {
+        fprintf(header, "%5u%s", base_dist[i],
+                SEPARATOR(i, D_CODES-1, 10));
+    }
+
+    fclose(header);
+}
+#endif /* GEN_TREES_H */
+
+/* ===========================================================================
+ * Initialize the tree data structures for a new zlib stream.
+ */
+void ZLIB_INTERNAL _tr_init(s)
+    deflate_state *s;
+{
+    tr_static_init();
+
+    s->l_desc.dyn_tree = s->dyn_ltree;
+    s->l_desc.stat_desc = &static_l_desc;
+
+    s->d_desc.dyn_tree = s->dyn_dtree;
+    s->d_desc.stat_desc = &static_d_desc;
+
+    s->bl_desc.dyn_tree = s->bl_tree;
+    s->bl_desc.stat_desc = &static_bl_desc;
+
+    s->bi_buf = 0;
+    s->bi_valid = 0;
+#ifdef DEBUG
+    s->compressed_len = 0L;
+    s->bits_sent = 0L;
+#endif
+
+    /* Initialize the first block of the first file: */
+    init_block(s);
+}
+
+/* ===========================================================================
+ * Initialize a new block.
+ */
+local void init_block(s)
+    deflate_state *s;
+{
+    int n; /* iterates over tree elements */
+
+    /* Initialize the trees. */
+    for (n = 0; n < L_CODES;  n++) s->dyn_ltree[n].Freq = 0;
+    for (n = 0; n < D_CODES;  n++) s->dyn_dtree[n].Freq = 0;
+    for (n = 0; n < BL_CODES; n++) s->bl_tree[n].Freq = 0;
+
+    s->dyn_ltree[END_BLOCK].Freq = 1;
+    s->opt_len = s->static_len = 0L;
+    s->last_lit = s->matches = 0;
+}
+
+#define SMALLEST 1
+/* Index within the heap array of least frequent node in the Huffman tree */
+
+
+/* ===========================================================================
+ * Remove the smallest element from the heap and recreate the heap with
+ * one less element. Updates heap and heap_len.
+ */
+#define pqremove(s, tree, top) \
+{\
+    top = s->heap[SMALLEST]; \
+    s->heap[SMALLEST] = s->heap[s->heap_len--]; \
+    pqdownheap(s, tree, SMALLEST); \
+}
+
+/* ===========================================================================
+ * Compares to subtrees, using the tree depth as tie breaker when
+ * the subtrees have equal frequency. This minimizes the worst case length.
+ */
+#define smaller(tree, n, m, depth) \
+   (tree[n].Freq < tree[m].Freq || \
+   (tree[n].Freq == tree[m].Freq && depth[n] <= depth[m]))
+
+/* ===========================================================================
+ * Restore the heap property by moving down the tree starting at node k,
+ * exchanging a node with the smallest of its two sons if necessary, stopping
+ * when the heap property is re-established (each father smaller than its
+ * two sons).
+ */
+local void pqdownheap(s, tree, k)
+    deflate_state *s;
+    ct_data *tree;  /* the tree to restore */
+    int k;               /* node to move down */
+{
+    int v = s->heap[k];
+    int j = k << 1;  /* left son of k */
+    while (j <= s->heap_len) {
+        /* Set j to the smallest of the two sons: */
+        if (j < s->heap_len &&
+            smaller(tree, s->heap[j+1], s->heap[j], s->depth)) {
+            j++;
+        }
+        /* Exit if v is smaller than both sons */
+        if (smaller(tree, v, s->heap[j], s->depth)) break;
+
+        /* Exchange v with the smallest son */
+        s->heap[k] = s->heap[j];  k = j;
+
+        /* And continue down the tree, setting j to the left son of k */
+        j <<= 1;
+    }
+    s->heap[k] = v;
+}
+
+/* ===========================================================================
+ * Compute the optimal bit lengths for a tree and update the total bit length
+ * for the current block.
+ * IN assertion: the fields freq and dad are set, heap[heap_max] and
+ *    above are the tree nodes sorted by increasing frequency.
+ * OUT assertions: the field len is set to the optimal bit length, the
+ *     array bl_count contains the frequencies for each bit length.
+ *     The length opt_len is updated; static_len is also updated if stree is
+ *     not null.
+ */
+local void gen_bitlen(s, desc)
+    deflate_state *s;
+    tree_desc *desc;    /* the tree descriptor */
+{
+    ct_data *tree        = desc->dyn_tree;
+    int max_code         = desc->max_code;
+    const ct_data *stree = desc->stat_desc->static_tree;
+    const intf *extra    = desc->stat_desc->extra_bits;
+    int base             = desc->stat_desc->extra_base;
+    int max_length       = desc->stat_desc->max_length;
+    int h;              /* heap index */
+    int n, m;           /* iterate over the tree elements */
+    int bits;           /* bit length */
+    int xbits;          /* extra bits */
+    ush f;              /* frequency */
+    int overflow = 0;   /* number of elements with bit length too large */
+
+    for (bits = 0; bits <= MAX_BITS; bits++) s->bl_count[bits] = 0;
+
+    /* In a first pass, compute the optimal bit lengths (which may
+     * overflow in the case of the bit length tree).
+     */
+    tree[s->heap[s->heap_max]].Len = 0; /* root of the heap */
+
+    for (h = s->heap_max+1; h < HEAP_SIZE; h++) {
+        n = s->heap[h];
+        bits = tree[tree[n].Dad].Len + 1;
+        if (bits > max_length) bits = max_length, overflow++;
+        tree[n].Len = (ush)bits;
+        /* We overwrite tree[n].Dad which is no longer needed */
+
+        if (n > max_code) continue; /* not a leaf node */
+
+        s->bl_count[bits]++;
+        xbits = 0;
+        if (n >= base) xbits = extra[n-base];
+        f = tree[n].Freq;
+        s->opt_len += (ulg)f * (bits + xbits);
+        if (stree) s->static_len += (ulg)f * (stree[n].Len + xbits);
+    }
+    if (overflow == 0) return;
+
+    Trace((stderr,"\nbit length overflow\n"));
+    /* This happens for example on obj2 and pic of the Calgary corpus */
+
+    /* Find the first bit length which could increase: */
+    do {
+        bits = max_length-1;
+        while (s->bl_count[bits] == 0) bits--;
+        s->bl_count[bits]--;      /* move one leaf down the tree */
+        s->bl_count[bits+1] += 2; /* move one overflow item as its brother */
+        s->bl_count[max_length]--;
+        /* The brother of the overflow item also moves one step up,
+         * but this does not affect bl_count[max_length]
+         */
+        overflow -= 2;
+    } while (overflow > 0);
+
+    /* Now recompute all bit lengths, scanning in increasing frequency.
+     * h is still equal to HEAP_SIZE. (It is simpler to reconstruct all
+     * lengths instead of fixing only the wrong ones. This idea is taken
+     * from 'ar' written by Haruhiko Okumura.)
+     */
+    for (bits = max_length; bits != 0; bits--) {
+        n = s->bl_count[bits];
+        while (n != 0) {
+            m = s->heap[--h];
+            if (m > max_code) continue;
+            if ((unsigned) tree[m].Len != (unsigned) bits) {
+                Trace((stderr,"code %d bits %d->%d\n", m, tree[m].Len, bits));
+                s->opt_len += ((long)bits - (long)tree[m].Len)
+                              *(long)tree[m].Freq;
+                tree[m].Len = (ush)bits;
+            }
+            n--;
+        }
+    }
+}
+
+/* ===========================================================================
+ * Generate the codes for a given tree and bit counts (which need not be
+ * optimal).
+ * IN assertion: the array bl_count contains the bit length statistics for
+ * the given tree and the field len is set for all tree elements.
+ * OUT assertion: the field code is set for all tree elements of non
+ *     zero code length.
+ */
+local void gen_codes (tree, max_code, bl_count)
+    ct_data *tree;             /* the tree to decorate */
+    int max_code;              /* largest code with non zero frequency */
+    ushf *bl_count;            /* number of codes at each bit length */
+{
+    ush next_code[MAX_BITS+1]; /* next code value for each bit length */
+    ush code = 0;              /* running code value */
+    int bits;                  /* bit index */
+    int n;                     /* code index */
+
+    /* The distribution counts are first used to generate the code values
+     * without bit reversal.
+     */
+    for (bits = 1; bits <= MAX_BITS; bits++) {
+        next_code[bits] = code = (code + bl_count[bits-1]) << 1;
+    }
+    /* Check that the bit counts in bl_count are consistent. The last code
+     * must be all ones.
+     */
+    Assert (code + bl_count[MAX_BITS]-1 == (1<<MAX_BITS)-1,
+            "inconsistent bit counts");
+    Tracev((stderr,"\ngen_codes: max_code %d ", max_code));
+
+    for (n = 0;  n <= max_code; n++) {
+        int len = tree[n].Len;
+        if (len == 0) continue;
+        /* Now reverse the bits */
+        tree[n].Code = bi_reverse(next_code[len]++, len);
+
+        Tracecv(tree != static_ltree, (stderr,"\nn %3d %c l %2d c %4x (%x) ",
+             n, (isgraph(n) ? n : ' '), len, tree[n].Code, next_code[len]-1));
+    }
+}
+
+/* ===========================================================================
+ * Construct one Huffman tree and assigns the code bit strings and lengths.
+ * Update the total bit length for the current block.
+ * IN assertion: the field freq is set for all tree elements.
+ * OUT assertions: the fields len and code are set to the optimal bit length
+ *     and corresponding code. The length opt_len is updated; static_len is
+ *     also updated if stree is not null. The field max_code is set.
+ */
+local void build_tree(s, desc)
+    deflate_state *s;
+    tree_desc *desc; /* the tree descriptor */
+{
+    ct_data *tree         = desc->dyn_tree;
+    const ct_data *stree  = desc->stat_desc->static_tree;
+    int elems             = desc->stat_desc->elems;
+    int n, m;          /* iterate over heap elements */
+    int max_code = -1; /* largest code with non zero frequency */
+    int node;          /* new node being created */
+
+    /* Construct the initial heap, with least frequent element in
+     * heap[SMALLEST]. The sons of heap[n] are heap[2*n] and heap[2*n+1].
+     * heap[0] is not used.
+     */
+    s->heap_len = 0, s->heap_max = HEAP_SIZE;
+
+    for (n = 0; n < elems; n++) {
+        if (tree[n].Freq != 0) {
+            s->heap[++(s->heap_len)] = max_code = n;
+            s->depth[n] = 0;
+        } else {
+            tree[n].Len = 0;
+        }
+    }
+
+    /* The pkzip format requires that at least one distance code exists,
+     * and that at least one bit should be sent even if there is only one
+     * possible code. So to avoid special checks later on we force at least
+     * two codes of non zero frequency.
+     */
+    while (s->heap_len < 2) {
+        node = s->heap[++(s->heap_len)] = (max_code < 2 ? ++max_code : 0);
+        tree[node].Freq = 1;
+        s->depth[node] = 0;
+        s->opt_len--; if (stree) s->static_len -= stree[node].Len;
+        /* node is 0 or 1 so it does not have extra bits */
+    }
+    desc->max_code = max_code;
+
+    /* The elements heap[heap_len/2+1 .. heap_len] are leaves of the tree,
+     * establish sub-heaps of increasing lengths:
+     */
+    for (n = s->heap_len/2; n >= 1; n--) pqdownheap(s, tree, n);
+
+    /* Construct the Huffman tree by repeatedly combining the least two
+     * frequent nodes.
+     */
+    node = elems;              /* next internal node of the tree */
+    do {
+        pqremove(s, tree, n);  /* n = node of least frequency */
+        m = s->heap[SMALLEST]; /* m = node of next least frequency */
+
+        s->heap[--(s->heap_max)] = n; /* keep the nodes sorted by frequency */
+        s->heap[--(s->heap_max)] = m;
+
+        /* Create a new node father of n and m */
+        tree[node].Freq = tree[n].Freq + tree[m].Freq;
+        s->depth[node] = (uch)((s->depth[n] >= s->depth[m] ?
+                                s->depth[n] : s->depth[m]) + 1);
+        tree[n].Dad = tree[m].Dad = (ush)node;
+#ifdef DUMP_BL_TREE
+        if (tree == s->bl_tree) {
+            fprintf(stderr,"\nnode %d(%d), sons %d(%d) %d(%d)",
+                    node, tree[node].Freq, n, tree[n].Freq, m, tree[m].Freq);
+        }
+#endif
+        /* and insert the new node in the heap */
+        s->heap[SMALLEST] = node++;
+        pqdownheap(s, tree, SMALLEST);
+
+    } while (s->heap_len >= 2);
+
+    s->heap[--(s->heap_max)] = s->heap[SMALLEST];
+
+    /* At this point, the fields freq and dad are set. We can now
+     * generate the bit lengths.
+     */
+    gen_bitlen(s, (tree_desc *)desc);
+
+    /* The field len is now set, we can generate the bit codes */
+    gen_codes ((ct_data *)tree, max_code, s->bl_count);
+}
+
+/* ===========================================================================
+ * Scan a literal or distance tree to determine the frequencies of the codes
+ * in the bit length tree.
+ */
+local void scan_tree (s, tree, max_code)
+    deflate_state *s;
+    ct_data *tree;   /* the tree to be scanned */
+    int max_code;    /* and its largest code of non zero frequency */
+{
+    int n;                     /* iterates over all tree elements */
+    int prevlen = -1;          /* last emitted length */
+    int curlen;                /* length of current code */
+    int nextlen = tree[0].Len; /* length of next code */
+    int count = 0;             /* repeat count of the current code */
+    int max_count = 7;         /* max repeat count */
+    int min_count = 4;         /* min repeat count */
+
+    if (nextlen == 0) max_count = 138, min_count = 3;
+    tree[max_code+1].Len = (ush)0xffff; /* guard */
+
+    for (n = 0; n <= max_code; n++) {
+        curlen = nextlen; nextlen = tree[n+1].Len;
+        if (++count < max_count && curlen == nextlen) {
+            continue;
+        } else if (count < min_count) {
+            s->bl_tree[curlen].Freq += count;
+        } else if (curlen != 0) {
+            if (curlen != prevlen) s->bl_tree[curlen].Freq++;
+            s->bl_tree[REP_3_6].Freq++;
+        } else if (count <= 10) {
+            s->bl_tree[REPZ_3_10].Freq++;
+        } else {
+            s->bl_tree[REPZ_11_138].Freq++;
+        }
+        count = 0; prevlen = curlen;
+        if (nextlen == 0) {
+            max_count = 138, min_count = 3;
+        } else if (curlen == nextlen) {
+            max_count = 6, min_count = 3;
+        } else {
+            max_count = 7, min_count = 4;
+        }
+    }
+}
+
+/* ===========================================================================
+ * Send a literal or distance tree in compressed form, using the codes in
+ * bl_tree.
+ */
+local void send_tree (s, tree, max_code)
+    deflate_state *s;
+    ct_data *tree; /* the tree to be scanned */
+    int max_code;       /* and its largest code of non zero frequency */
+{
+    int n;                     /* iterates over all tree elements */
+    int prevlen = -1;          /* last emitted length */
+    int curlen;                /* length of current code */
+    int nextlen = tree[0].Len; /* length of next code */
+    int count = 0;             /* repeat count of the current code */
+    int max_count = 7;         /* max repeat count */
+    int min_count = 4;         /* min repeat count */
+
+    /* tree[max_code+1].Len = -1; */  /* guard already set */
+    if (nextlen == 0) max_count = 138, min_count = 3;
+
+    for (n = 0; n <= max_code; n++) {
+        curlen = nextlen; nextlen = tree[n+1].Len;
+        if (++count < max_count && curlen == nextlen) {
+            continue;
+        } else if (count < min_count) {
+            do { send_code(s, curlen, s->bl_tree); } while (--count != 0);
+
+        } else if (curlen != 0) {
+            if (curlen != prevlen) {
+                send_code(s, curlen, s->bl_tree); count--;
+            }
+            Assert(count >= 3 && count <= 6, " 3_6?");
+            send_code(s, REP_3_6, s->bl_tree); send_bits(s, count-3, 2);
+
+        } else if (count <= 10) {
+            send_code(s, REPZ_3_10, s->bl_tree); send_bits(s, count-3, 3);
+
+        } else {
+            send_code(s, REPZ_11_138, s->bl_tree); send_bits(s, count-11, 7);
+        }
+        count = 0; prevlen = curlen;
+        if (nextlen == 0) {
+            max_count = 138, min_count = 3;
+        } else if (curlen == nextlen) {
+            max_count = 6, min_count = 3;
+        } else {
+            max_count = 7, min_count = 4;
+        }
+    }
+}
+
+/* ===========================================================================
+ * Construct the Huffman tree for the bit lengths and return the index in
+ * bl_order of the last bit length code to send.
+ */
+local int build_bl_tree(s)
+    deflate_state *s;
+{
+    int max_blindex;  /* index of last bit length code of non zero freq */
+
+    /* Determine the bit length frequencies for literal and distance trees */
+    scan_tree(s, (ct_data *)s->dyn_ltree, s->l_desc.max_code);
+    scan_tree(s, (ct_data *)s->dyn_dtree, s->d_desc.max_code);
+
+    /* Build the bit length tree: */
+    build_tree(s, (tree_desc *)(&(s->bl_desc)));
+    /* opt_len now includes the length of the tree representations, except
+     * the lengths of the bit lengths codes and the 5+5+4 bits for the counts.
+     */
+
+    /* Determine the number of bit length codes to send. The pkzip format
+     * requires that at least 4 bit length codes be sent. (appnote.txt says
+     * 3 but the actual value used is 4.)
+     */
+    for (max_blindex = BL_CODES-1; max_blindex >= 3; max_blindex--) {
+        if (s->bl_tree[bl_order[max_blindex]].Len != 0) break;
+    }
+    /* Update opt_len to include the bit length tree and counts */
+    s->opt_len += 3*(max_blindex+1) + 5+5+4;
+    Tracev((stderr, "\ndyn trees: dyn %ld, stat %ld",
+            s->opt_len, s->static_len));
+
+    return max_blindex;
+}
+
+/* ===========================================================================
+ * Send the header for a block using dynamic Huffman trees: the counts, the
+ * lengths of the bit length codes, the literal tree and the distance tree.
+ * IN assertion: lcodes >= 257, dcodes >= 1, blcodes >= 4.
+ */
+local void send_all_trees(s, lcodes, dcodes, blcodes)
+    deflate_state *s;
+    int lcodes, dcodes, blcodes; /* number of codes for each tree */
+{
+    int rank;                    /* index in bl_order */
+
+    Assert (lcodes >= 257 && dcodes >= 1 && blcodes >= 4, "not enough codes");
+    Assert (lcodes <= L_CODES && dcodes <= D_CODES && blcodes <= BL_CODES,
+            "too many codes");
+    Tracev((stderr, "\nbl counts: "));
+    send_bits(s, lcodes-257, 5); /* not +255 as stated in appnote.txt */
+    send_bits(s, dcodes-1,   5);
+    send_bits(s, blcodes-4,  4); /* not -3 as stated in appnote.txt */
+    for (rank = 0; rank < blcodes; rank++) {
+        Tracev((stderr, "\nbl code %2d ", bl_order[rank]));
+        send_bits(s, s->bl_tree[bl_order[rank]].Len, 3);
+    }
+    Tracev((stderr, "\nbl tree: sent %ld", s->bits_sent));
+
+    send_tree(s, (ct_data *)s->dyn_ltree, lcodes-1); /* literal tree */
+    Tracev((stderr, "\nlit tree: sent %ld", s->bits_sent));
+
+    send_tree(s, (ct_data *)s->dyn_dtree, dcodes-1); /* distance tree */
+    Tracev((stderr, "\ndist tree: sent %ld", s->bits_sent));
+}
+
+/* ===========================================================================
+ * Send a stored block
+ */
+void ZLIB_INTERNAL _tr_stored_block(s, buf, stored_len, last)
+    deflate_state *s;
+    charf *buf;       /* input block */
+    ulg stored_len;   /* length of input block */
+    int last;         /* one if this is the last block for a file */
+{
+    send_bits(s, (STORED_BLOCK<<1)+last, 3);    /* send block type */
+#ifdef DEBUG
+    s->compressed_len = (s->compressed_len + 3 + 7) & (ulg)~7L;
+    s->compressed_len += (stored_len + 4) << 3;
+#endif
+    copy_block(s, buf, (unsigned)stored_len, 1); /* with header */
+}
+
+/* ===========================================================================
+ * Flush the bits in the bit buffer to pending output (leaves at most 7 bits)
+ */
+void ZLIB_INTERNAL _tr_flush_bits(s)
+    deflate_state *s;
+{
+    bi_flush(s);
+}
+
+/* ===========================================================================
+ * Send one empty static block to give enough lookahead for inflate.
+ * This takes 10 bits, of which 7 may remain in the bit buffer.
+ */
+void ZLIB_INTERNAL _tr_align(s)
+    deflate_state *s;
+{
+    send_bits(s, STATIC_TREES<<1, 3);
+    send_code(s, END_BLOCK, static_ltree);
+#ifdef DEBUG
+    s->compressed_len += 10L; /* 3 for block type, 7 for EOB */
+#endif
+    bi_flush(s);
+}
+
+/* ===========================================================================
+ * Determine the best encoding for the current block: dynamic trees, static
+ * trees or store, and output the encoded block to the zip file.
+ */
+void ZLIB_INTERNAL _tr_flush_block(s, buf, stored_len, last)
+    deflate_state *s;
+    charf *buf;       /* input block, or NULL if too old */
+    ulg stored_len;   /* length of input block */
+    int last;         /* one if this is the last block for a file */
+{
+    ulg opt_lenb, static_lenb; /* opt_len and static_len in bytes */
+    int max_blindex = 0;  /* index of last bit length code of non zero freq */
+
+    /* Build the Huffman trees unless a stored block is forced */
+    if (s->level > 0) {
+
+        /* Check if the file is binary or text */
+        if (s->strm->data_type == Z_UNKNOWN)
+            s->strm->data_type = detect_data_type(s);
+
+        /* Construct the literal and distance trees */
+        build_tree(s, (tree_desc *)(&(s->l_desc)));
+        Tracev((stderr, "\nlit data: dyn %ld, stat %ld", s->opt_len,
+                s->static_len));
+
+        build_tree(s, (tree_desc *)(&(s->d_desc)));
+        Tracev((stderr, "\ndist data: dyn %ld, stat %ld", s->opt_len,
+                s->static_len));
+        /* At this point, opt_len and static_len are the total bit lengths of
+         * the compressed block data, excluding the tree representations.
+         */
+
+        /* Build the bit length tree for the above two trees, and get the index
+         * in bl_order of the last bit length code to send.
+         */
+        max_blindex = build_bl_tree(s);
+
+        /* Determine the best encoding. Compute the block lengths in bytes. */
+        opt_lenb = (s->opt_len+3+7)>>3;
+        static_lenb = (s->static_len+3+7)>>3;
+
+        Tracev((stderr, "\nopt %lu(%lu) stat %lu(%lu) stored %lu lit %u ",
+                opt_lenb, s->opt_len, static_lenb, s->static_len, stored_len,
+                s->last_lit));
+
+        if (static_lenb <= opt_lenb) opt_lenb = static_lenb;
+
+    } else {
+        Assert(buf != (char*)0, "lost buf");
+        opt_lenb = static_lenb = stored_len + 5; /* force a stored block */
+    }
+
+#ifdef FORCE_STORED
+    if (buf != (char*)0) { /* force stored block */
+#else
+    if (stored_len+4 <= opt_lenb && buf != (char*)0) {
+                       /* 4: two words for the lengths */
+#endif
+        /* The test buf != NULL is only necessary if LIT_BUFSIZE > WSIZE.
+         * Otherwise we can't have processed more than WSIZE input bytes since
+         * the last block flush, because compression would have been
+         * successful. If LIT_BUFSIZE <= WSIZE, it is never too late to
+         * transform a block into a stored block.
+         */
+        _tr_stored_block(s, buf, stored_len, last);
+
+#ifdef FORCE_STATIC
+    } else if (static_lenb >= 0) { /* force static trees */
+#else
+    } else if (s->strategy == Z_FIXED || static_lenb == opt_lenb) {
+#endif
+        send_bits(s, (STATIC_TREES<<1)+last, 3);
+        compress_block(s, (const ct_data *)static_ltree,
+                       (const ct_data *)static_dtree);
+#ifdef DEBUG
+        s->compressed_len += 3 + s->static_len;
+#endif
+    } else {
+        send_bits(s, (DYN_TREES<<1)+last, 3);
+        send_all_trees(s, s->l_desc.max_code+1, s->d_desc.max_code+1,
+                       max_blindex+1);
+        compress_block(s, (const ct_data *)s->dyn_ltree,
+                       (const ct_data *)s->dyn_dtree);
+#ifdef DEBUG
+        s->compressed_len += 3 + s->opt_len;
+#endif
+    }
+    Assert (s->compressed_len == s->bits_sent, "bad compressed size");
+    /* The above check is made mod 2^32, for files larger than 512 MB
+     * and uLong implemented on 32 bits.
+     */
+    init_block(s);
+
+    if (last) {
+        bi_windup(s);
+#ifdef DEBUG
+        s->compressed_len += 7;  /* align on byte boundary */
+#endif
+    }
+    Tracev((stderr,"\ncomprlen %lu(%lu) ", s->compressed_len>>3,
+           s->compressed_len-7*last));
+}
+
+/* ===========================================================================
+ * Save the match info and tally the frequency counts. Return true if
+ * the current block must be flushed.
+ */
+int ZLIB_INTERNAL _tr_tally (s, dist, lc)
+    deflate_state *s;
+    unsigned dist;  /* distance of matched string */
+    unsigned lc;    /* match length-MIN_MATCH or unmatched char (if dist==0) */
+{
+    s->d_buf[s->last_lit] = (ush)dist;
+    s->l_buf[s->last_lit++] = (uch)lc;
+    if (dist == 0) {
+        /* lc is the unmatched char */
+        s->dyn_ltree[lc].Freq++;
+    } else {
+        s->matches++;
+        /* Here, lc is the match length - MIN_MATCH */
+        dist--;             /* dist = match distance - 1 */
+        Assert((ush)dist < (ush)MAX_DIST(s) &&
+               (ush)lc <= (ush)(MAX_MATCH-MIN_MATCH) &&
+               (ush)d_code(dist) < (ush)D_CODES,  "_tr_tally: bad match");
+
+        s->dyn_ltree[_length_code[lc]+LITERALS+1].Freq++;
+        s->dyn_dtree[d_code(dist)].Freq++;
+    }
+
+#ifdef TRUNCATE_BLOCK
+    /* Try to guess if it is profitable to stop the current block here */
+    if ((s->last_lit & 0x1fff) == 0 && s->level > 2) {
+        /* Compute an upper bound for the compressed length */
+        ulg out_length = (ulg)s->last_lit*8L;
+        ulg in_length = (ulg)((long)s->strstart - s->block_start);
+        int dcode;
+        for (dcode = 0; dcode < D_CODES; dcode++) {
+            out_length += (ulg)s->dyn_dtree[dcode].Freq *
+                (5L+extra_dbits[dcode]);
+        }
+        out_length >>= 3;
+        Tracev((stderr,"\nlast_lit %u, in %ld, out ~%ld(%ld%%) ",
+               s->last_lit, in_length, out_length,
+               100L - out_length*100L/in_length));
+        if (s->matches < s->last_lit/2 && out_length < in_length/2) return 1;
+    }
+#endif
+    return (s->last_lit == s->lit_bufsize-1);
+    /* We avoid equality with lit_bufsize because of wraparound at 64K
+     * on 16 bit machines and because stored blocks are restricted to
+     * 64K-1 bytes.
+     */
+}
+
+/* ===========================================================================
+ * Send the block data compressed using the given Huffman trees
+ */
+local void compress_block(s, ltree, dtree)
+    deflate_state *s;
+    const ct_data *ltree; /* literal tree */
+    const ct_data *dtree; /* distance tree */
+{
+    unsigned dist;      /* distance of matched string */
+    int lc;             /* match length or unmatched char (if dist == 0) */
+    unsigned lx = 0;    /* running index in l_buf */
+    unsigned code;      /* the code to send */
+    int extra;          /* number of extra bits to send */
+
+    if (s->last_lit != 0) do {
+        dist = s->d_buf[lx];
+        lc = s->l_buf[lx++];
+        if (dist == 0) {
+            send_code(s, lc, ltree); /* send a literal byte */
+            Tracecv(isgraph(lc), (stderr," '%c' ", lc));
+        } else {
+            /* Here, lc is the match length - MIN_MATCH */
+            code = _length_code[lc];
+            send_code(s, code+LITERALS+1, ltree); /* send the length code */
+            extra = extra_lbits[code];
+            if (extra != 0) {
+                lc -= base_length[code];
+                send_bits(s, lc, extra);       /* send the extra length bits */
+            }
+            dist--; /* dist is now the match distance - 1 */
+            code = d_code(dist);
+            Assert (code < D_CODES, "bad d_code");
+
+            send_code(s, code, dtree);       /* send the distance code */
+            extra = extra_dbits[code];
+            if (extra != 0) {
+                dist -= base_dist[code];
+                send_bits(s, dist, extra);   /* send the extra distance bits */
+            }
+        } /* literal or match pair ? */
+
+        /* Check that the overlay between pending_buf and d_buf+l_buf is ok: */
+        Assert((uInt)(s->pending) < s->lit_bufsize + 2*lx,
+               "pendingBuf overflow");
+
+    } while (lx < s->last_lit);
+
+    send_code(s, END_BLOCK, ltree);
+}
+
+/* ===========================================================================
+ * Check if the data type is TEXT or BINARY, using the following algorithm:
+ * - TEXT if the two conditions below are satisfied:
+ *    a) There are no non-portable control characters belonging to the
+ *       "black list" (0..6, 14..25, 28..31).
+ *    b) There is at least one printable character belonging to the
+ *       "white list" (9 {TAB}, 10 {LF}, 13 {CR}, 32..255).
+ * - BINARY otherwise.
+ * - The following partially-portable control characters form a
+ *   "gray list" that is ignored in this detection algorithm:
+ *   (7 {BEL}, 8 {BS}, 11 {VT}, 12 {FF}, 26 {SUB}, 27 {ESC}).
+ * IN assertion: the fields Freq of dyn_ltree are set.
+ */
+local int detect_data_type(s)
+    deflate_state *s;
+{
+    /* black_mask is the bit mask of black-listed bytes
+     * set bits 0..6, 14..25, and 28..31
+     * 0xf3ffc07f = binary 11110011111111111100000001111111
+     */
+    unsigned long black_mask = 0xf3ffc07fUL;
+    int n;
+
+    /* Check for non-textual ("black-listed") bytes. */
+    for (n = 0; n <= 31; n++, black_mask >>= 1)
+        if ((black_mask & 1) && (s->dyn_ltree[n].Freq != 0))
+            return Z_BINARY;
+
+    /* Check for textual ("white-listed") bytes. */
+    if (s->dyn_ltree[9].Freq != 0 || s->dyn_ltree[10].Freq != 0
+            || s->dyn_ltree[13].Freq != 0)
+        return Z_TEXT;
+    for (n = 32; n < LITERALS; n++)
+        if (s->dyn_ltree[n].Freq != 0)
+            return Z_TEXT;
+
+    /* There are no "black-listed" or "white-listed" bytes:
+     * this stream either is empty or has tolerated ("gray-listed") bytes only.
+     */
+    return Z_BINARY;
+}
+
+/* ===========================================================================
+ * Reverse the first len bits of a code, using straightforward code (a faster
+ * method would use a table)
+ * IN assertion: 1 <= len <= 15
+ */
+local unsigned bi_reverse(code, len)
+    unsigned code; /* the value to invert */
+    int len;       /* its bit length */
+{
+    register unsigned res = 0;
+    do {
+        res |= code & 1;
+        code >>= 1, res <<= 1;
+    } while (--len > 0);
+    return res >> 1;
+}
+
+/* ===========================================================================
+ * Flush the bit buffer, keeping at most 7 bits in it.
+ */
+local void bi_flush(s)
+    deflate_state *s;
+{
+    if (s->bi_valid == 16) {
+        put_short(s, s->bi_buf);
+        s->bi_buf = 0;
+        s->bi_valid = 0;
+    } else if (s->bi_valid >= 8) {
+        put_byte(s, (Byte)s->bi_buf);
+        s->bi_buf >>= 8;
+        s->bi_valid -= 8;
+    }
+}
+
+/* ===========================================================================
+ * Flush the bit buffer and align the output on a byte boundary
+ */
+local void bi_windup(s)
+    deflate_state *s;
+{
+    if (s->bi_valid > 8) {
+        put_short(s, s->bi_buf);
+    } else if (s->bi_valid > 0) {
+        put_byte(s, (Byte)s->bi_buf);
+    }
+    s->bi_buf = 0;
+    s->bi_valid = 0;
+#ifdef DEBUG
+    s->bits_sent = (s->bits_sent+7) & ~7;
+#endif
+}
+
+/* ===========================================================================
+ * Copy a stored block, storing first the length and its
+ * one's complement if requested.
+ */
+local void copy_block(s, buf, len, header)
+    deflate_state *s;
+    charf    *buf;    /* the input data */
+    unsigned len;     /* its length */
+    int      header;  /* true if block header must be written */
+{
+    bi_windup(s);        /* align on byte boundary */
+
+    if (header) {
+        put_short(s, (ush)len);
+        put_short(s, (ush)~len);
+#ifdef DEBUG
+        s->bits_sent += 2*16;
+#endif
+    }
+#ifdef DEBUG
+    s->bits_sent += (ulg)len<<3;
+#endif
+    while (len--) {
+        put_byte(s, *buf++);
+    }
+}
diff --git a/ml/dlib/dlib/external/zlib/trees.h b/ml/dlib/dlib/external/zlib/trees.h
new file mode 100644
index 000000000..d35639d82
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/trees.h
@@ -0,0 +1,128 @@
+/* header created automatically with -DGEN_TREES_H */
+
+local const ct_data static_ltree[L_CODES+2] = {
+{{ 12},{  8}}, {{140},{  8}}, {{ 76},{  8}}, {{204},{  8}}, {{ 44},{  8}},
+{{172},{  8}}, {{108},{  8}}, {{236},{  8}}, {{ 28},{  8}}, {{156},{  8}},
+{{ 92},{  8}}, {{220},{  8}}, {{ 60},{  8}}, {{188},{  8}}, {{124},{  8}},
+{{252},{  8}}, {{  2},{  8}}, {{130},{  8}}, {{ 66},{  8}}, {{194},{  8}},
+{{ 34},{  8}}, {{162},{  8}}, {{ 98},{  8}}, {{226},{  8}}, {{ 18},{  8}},
+{{146},{  8}}, {{ 82},{  8}}, {{210},{  8}}, {{ 50},{  8}}, {{178},{  8}},
+{{114},{  8}}, {{242},{  8}}, {{ 10},{  8}}, {{138},{  8}}, {{ 74},{  8}},
+{{202},{  8}}, {{ 42},{  8}}, {{170},{  8}}, {{106},{  8}}, {{234},{  8}},
+{{ 26},{  8}}, {{154},{  8}}, {{ 90},{  8}}, {{218},{  8}}, {{ 58},{  8}},
+{{186},{  8}}, {{122},{  8}}, {{250},{  8}}, {{  6},{  8}}, {{134},{  8}},
+{{ 70},{  8}}, {{198},{  8}}, {{ 38},{  8}}, {{166},{  8}}, {{102},{  8}},
+{{230},{  8}}, {{ 22},{  8}}, {{150},{  8}}, {{ 86},{  8}}, {{214},{  8}},
+{{ 54},{  8}}, {{182},{  8}}, {{118},{  8}}, {{246},{  8}}, {{ 14},{  8}},
+{{142},{  8}}, {{ 78},{  8}}, {{206},{  8}}, {{ 46},{  8}}, {{174},{  8}},
+{{110},{  8}}, {{238},{  8}}, {{ 30},{  8}}, {{158},{  8}}, {{ 94},{  8}},
+{{222},{  8}}, {{ 62},{  8}}, {{190},{  8}}, {{126},{  8}}, {{254},{  8}},
+{{  1},{  8}}, {{129},{  8}}, {{ 65},{  8}}, {{193},{  8}}, {{ 33},{  8}},
+{{161},{  8}}, {{ 97},{  8}}, {{225},{  8}}, {{ 17},{  8}}, {{145},{  8}},
+{{ 81},{  8}}, {{209},{  8}}, {{ 49},{  8}}, {{177},{  8}}, {{113},{  8}},
+{{241},{  8}}, {{  9},{  8}}, {{137},{  8}}, {{ 73},{  8}}, {{201},{  8}},
+{{ 41},{  8}}, {{169},{  8}}, {{105},{  8}}, {{233},{  8}}, {{ 25},{  8}},
+{{153},{  8}}, {{ 89},{  8}}, {{217},{  8}}, {{ 57},{  8}}, {{185},{  8}},
+{{121},{  8}}, {{249},{  8}}, {{  5},{  8}}, {{133},{  8}}, {{ 69},{  8}},
+{{197},{  8}}, {{ 37},{  8}}, {{165},{  8}}, {{101},{  8}}, {{229},{  8}},
+{{ 21},{  8}}, {{149},{  8}}, {{ 85},{  8}}, {{213},{  8}}, {{ 53},{  8}},
+{{181},{  8}}, {{117},{  8}}, {{245},{  8}}, {{ 13},{  8}}, {{141},{  8}},
+{{ 77},{  8}}, {{205},{  8}}, {{ 45},{  8}}, {{173},{  8}}, {{109},{  8}},
+{{237},{  8}}, {{ 29},{  8}}, {{157},{  8}}, {{ 93},{  8}}, {{221},{  8}},
+{{ 61},{  8}}, {{189},{  8}}, {{125},{  8}}, {{253},{  8}}, {{ 19},{  9}},
+{{275},{  9}}, {{147},{  9}}, {{403},{  9}}, {{ 83},{  9}}, {{339},{  9}},
+{{211},{  9}}, {{467},{  9}}, {{ 51},{  9}}, {{307},{  9}}, {{179},{  9}},
+{{435},{  9}}, {{115},{  9}}, {{371},{  9}}, {{243},{  9}}, {{499},{  9}},
+{{ 11},{  9}}, {{267},{  9}}, {{139},{  9}}, {{395},{  9}}, {{ 75},{  9}},
+{{331},{  9}}, {{203},{  9}}, {{459},{  9}}, {{ 43},{  9}}, {{299},{  9}},
+{{171},{  9}}, {{427},{  9}}, {{107},{  9}}, {{363},{  9}}, {{235},{  9}},
+{{491},{  9}}, {{ 27},{  9}}, {{283},{  9}}, {{155},{  9}}, {{411},{  9}},
+{{ 91},{  9}}, {{347},{  9}}, {{219},{  9}}, {{475},{  9}}, {{ 59},{  9}},
+{{315},{  9}}, {{187},{  9}}, {{443},{  9}}, {{123},{  9}}, {{379},{  9}},
+{{251},{  9}}, {{507},{  9}}, {{  7},{  9}}, {{263},{  9}}, {{135},{  9}},
+{{391},{  9}}, {{ 71},{  9}}, {{327},{  9}}, {{199},{  9}}, {{455},{  9}},
+{{ 39},{  9}}, {{295},{  9}}, {{167},{  9}}, {{423},{  9}}, {{103},{  9}},
+{{359},{  9}}, {{231},{  9}}, {{487},{  9}}, {{ 23},{  9}}, {{279},{  9}},
+{{151},{  9}}, {{407},{  9}}, {{ 87},{  9}}, {{343},{  9}}, {{215},{  9}},
+{{471},{  9}}, {{ 55},{  9}}, {{311},{  9}}, {{183},{  9}}, {{439},{  9}},
+{{119},{  9}}, {{375},{  9}}, {{247},{  9}}, {{503},{  9}}, {{ 15},{  9}},
+{{271},{  9}}, {{143},{  9}}, {{399},{  9}}, {{ 79},{  9}}, {{335},{  9}},
+{{207},{  9}}, {{463},{  9}}, {{ 47},{  9}}, {{303},{  9}}, {{175},{  9}},
+{{431},{  9}}, {{111},{  9}}, {{367},{  9}}, {{239},{  9}}, {{495},{  9}},
+{{ 31},{  9}}, {{287},{  9}}, {{159},{  9}}, {{415},{  9}}, {{ 95},{  9}},
+{{351},{  9}}, {{223},{  9}}, {{479},{  9}}, {{ 63},{  9}}, {{319},{  9}},
+{{191},{  9}}, {{447},{  9}}, {{127},{  9}}, {{383},{  9}}, {{255},{  9}},
+{{511},{  9}}, {{  0},{  7}}, {{ 64},{  7}}, {{ 32},{  7}}, {{ 96},{  7}},
+{{ 16},{  7}}, {{ 80},{  7}}, {{ 48},{  7}}, {{112},{  7}}, {{  8},{  7}},
+{{ 72},{  7}}, {{ 40},{  7}}, {{104},{  7}}, {{ 24},{  7}}, {{ 88},{  7}},
+{{ 56},{  7}}, {{120},{  7}}, {{  4},{  7}}, {{ 68},{  7}}, {{ 36},{  7}},
+{{100},{  7}}, {{ 20},{  7}}, {{ 84},{  7}}, {{ 52},{  7}}, {{116},{  7}},
+{{  3},{  8}}, {{131},{  8}}, {{ 67},{  8}}, {{195},{  8}}, {{ 35},{  8}},
+{{163},{  8}}, {{ 99},{  8}}, {{227},{  8}}
+};
+
+local const ct_data static_dtree[D_CODES] = {
+{{ 0},{ 5}}, {{16},{ 5}}, {{ 8},{ 5}}, {{24},{ 5}}, {{ 4},{ 5}},
+{{20},{ 5}}, {{12},{ 5}}, {{28},{ 5}}, {{ 2},{ 5}}, {{18},{ 5}},
+{{10},{ 5}}, {{26},{ 5}}, {{ 6},{ 5}}, {{22},{ 5}}, {{14},{ 5}},
+{{30},{ 5}}, {{ 1},{ 5}}, {{17},{ 5}}, {{ 9},{ 5}}, {{25},{ 5}},
+{{ 5},{ 5}}, {{21},{ 5}}, {{13},{ 5}}, {{29},{ 5}}, {{ 3},{ 5}},
+{{19},{ 5}}, {{11},{ 5}}, {{27},{ 5}}, {{ 7},{ 5}}, {{23},{ 5}}
+};
+
+const uch ZLIB_INTERNAL _dist_code[DIST_CODE_LEN] = {
+ 0,  1,  2,  3,  4,  4,  5,  5,  6,  6,  6,  6,  7,  7,  7,  7,  8,  8,  8,  8,
+ 8,  8,  8,  8,  9,  9,  9,  9,  9,  9,  9,  9, 10, 10, 10, 10, 10, 10, 10, 10,
+10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
+12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 13, 13, 13, 13,
+13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
+13, 13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
+14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15,
+15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
+15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,  0,  0, 16, 17,
+18, 18, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22, 22,
+23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+24, 24, 24, 24, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 27, 27, 27, 27, 27, 27, 27, 27,
+27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+27, 27, 27, 27, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
+28, 28, 28, 28, 28, 28, 28, 28, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
+29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29
+};
+
+const uch ZLIB_INTERNAL _length_code[MAX_MATCH-MIN_MATCH+1]= {
+ 0,  1,  2,  3,  4,  5,  6,  7,  8,  8,  9,  9, 10, 10, 11, 11, 12, 12, 12, 12,
+13, 13, 13, 13, 14, 14, 14, 14, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16,
+17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19,
+19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20,
+21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22, 22,
+22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23,
+23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
+25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
+25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 26, 26, 26, 26, 26, 26, 26, 26,
+26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
+26, 26, 26, 26, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
+27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 28
+};
+
+local const int base_length[LENGTH_CODES] = {
+0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56,
+64, 80, 96, 112, 128, 160, 192, 224, 0
+};
+
+local const int base_dist[D_CODES] = {
+    0,     1,     2,     3,     4,     6,     8,    12,    16,    24,
+   32,    48,    64,    96,   128,   192,   256,   384,   512,   768,
+ 1024,  1536,  2048,  3072,  4096,  6144,  8192, 12288, 16384, 24576
+};
+
diff --git a/ml/dlib/dlib/external/zlib/uncompr.c b/ml/dlib/dlib/external/zlib/uncompr.c
new file mode 100644
index 000000000..242e9493d
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/uncompr.c
@@ -0,0 +1,59 @@
+/* uncompr.c -- decompress a memory buffer
+ * Copyright (C) 1995-2003, 2010 Jean-loup Gailly.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* @(#) $Id$ */
+
+#define ZLIB_INTERNAL
+#include "zlib.h"
+
+/* ===========================================================================
+     Decompresses the source buffer into the destination buffer.  sourceLen is
+   the byte length of the source buffer. Upon entry, destLen is the total
+   size of the destination buffer, which must be large enough to hold the
+   entire uncompressed data. (The size of the uncompressed data must have
+   been saved previously by the compressor and transmitted to the decompressor
+   by some mechanism outside the scope of this compression library.)
+   Upon exit, destLen is the actual size of the compressed buffer.
+
+     uncompress returns Z_OK if success, Z_MEM_ERROR if there was not
+   enough memory, Z_BUF_ERROR if there was not enough room in the output
+   buffer, or Z_DATA_ERROR if the input data was corrupted.
+*/
+int ZEXPORT uncompress (dest, destLen, source, sourceLen)
+    Bytef *dest;
+    uLongf *destLen;
+    const Bytef *source;
+    uLong sourceLen;
+{
+    z_stream stream;
+    int err;
+
+    stream.next_in = (z_const Bytef *)source;
+    stream.avail_in = (uInt)sourceLen;
+    /* Check for source > 64K on 16-bit machine: */
+    if ((uLong)stream.avail_in != sourceLen) return Z_BUF_ERROR;
+
+    stream.next_out = dest;
+    stream.avail_out = (uInt)*destLen;
+    if ((uLong)stream.avail_out != *destLen) return Z_BUF_ERROR;
+
+    stream.zalloc = (alloc_func)0;
+    stream.zfree = (free_func)0;
+
+    err = inflateInit(&stream);
+    if (err != Z_OK) return err;
+
+    err = inflate(&stream, Z_FINISH);
+    if (err != Z_STREAM_END) {
+        inflateEnd(&stream);
+        if (err == Z_NEED_DICT || (err == Z_BUF_ERROR && stream.avail_in == 0))
+            return Z_DATA_ERROR;
+        return err;
+    }
+    *destLen = stream.total_out;
+
+    err = inflateEnd(&stream);
+    return err;
+}
diff --git a/ml/dlib/dlib/external/zlib/zconf.h b/ml/dlib/dlib/external/zlib/zconf.h
new file mode 100644
index 000000000..9987a7755
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/zconf.h
@@ -0,0 +1,511 @@
+/* zconf.h -- configuration of the zlib compression library
+ * Copyright (C) 1995-2013 Jean-loup Gailly.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* @(#) $Id$ */
+
+#ifndef ZCONF_H
+#define ZCONF_H
+
+/*
+ * If you *really* need a unique prefix for all types and library functions,
+ * compile with -DZ_PREFIX. The "standard" zlib should be compiled without it.
+ * Even better than compiling with -DZ_PREFIX would be to use configure to set
+ * this permanently in zconf.h using "./configure --zprefix".
+ */
+#ifdef Z_PREFIX     /* may be set to #if 1 by ./configure */
+#  define Z_PREFIX_SET
+
+/* all linked symbols */
+#  define _dist_code            z__dist_code
+#  define _length_code          z__length_code
+#  define _tr_align             z__tr_align
+#  define _tr_flush_bits        z__tr_flush_bits
+#  define _tr_flush_block       z__tr_flush_block
+#  define _tr_init              z__tr_init
+#  define _tr_stored_block      z__tr_stored_block
+#  define _tr_tally             z__tr_tally
+#  define adler32               z_adler32
+#  define adler32_combine       z_adler32_combine
+#  define adler32_combine64     z_adler32_combine64
+#  ifndef Z_SOLO
+#    define compress              z_compress
+#    define compress2             z_compress2
+#    define compressBound         z_compressBound
+#  endif
+#  define crc32                 z_crc32
+#  define crc32_combine         z_crc32_combine
+#  define crc32_combine64       z_crc32_combine64
+#  define deflate               z_deflate
+#  define deflateBound          z_deflateBound
+#  define deflateCopy           z_deflateCopy
+#  define deflateEnd            z_deflateEnd
+#  define deflateInit2_         z_deflateInit2_
+#  define deflateInit_          z_deflateInit_
+#  define deflateParams         z_deflateParams
+#  define deflatePending        z_deflatePending
+#  define deflatePrime          z_deflatePrime
+#  define deflateReset          z_deflateReset
+#  define deflateResetKeep      z_deflateResetKeep
+#  define deflateSetDictionary  z_deflateSetDictionary
+#  define deflateSetHeader      z_deflateSetHeader
+#  define deflateTune           z_deflateTune
+#  define deflate_copyright     z_deflate_copyright
+#  define get_crc_table         z_get_crc_table
+#  ifndef Z_SOLO
+#    define gz_error              z_gz_error
+#    define gz_intmax             z_gz_intmax
+#    define gz_strwinerror        z_gz_strwinerror
+#    define gzbuffer              z_gzbuffer
+#    define gzclearerr            z_gzclearerr
+#    define gzclose               z_gzclose
+#    define gzclose_r             z_gzclose_r
+#    define gzclose_w             z_gzclose_w
+#    define gzdirect              z_gzdirect
+#    define gzdopen               z_gzdopen
+#    define gzeof                 z_gzeof
+#    define gzerror               z_gzerror
+#    define gzflush               z_gzflush
+#    define gzgetc                z_gzgetc
+#    define gzgetc_               z_gzgetc_
+#    define gzgets                z_gzgets
+#    define gzoffset              z_gzoffset
+#    define gzoffset64            z_gzoffset64
+#    define gzopen                z_gzopen
+#    define gzopen64              z_gzopen64
+#    ifdef _WIN32
+#      define gzopen_w              z_gzopen_w
+#    endif
+#    define gzprintf              z_gzprintf
+#    define gzvprintf             z_gzvprintf
+#    define gzputc                z_gzputc
+#    define gzputs                z_gzputs
+#    define gzread                z_gzread
+#    define gzrewind              z_gzrewind
+#    define gzseek                z_gzseek
+#    define gzseek64              z_gzseek64
+#    define gzsetparams           z_gzsetparams
+#    define gztell                z_gztell
+#    define gztell64              z_gztell64
+#    define gzungetc              z_gzungetc
+#    define gzwrite               z_gzwrite
+#  endif
+#  define inflate               z_inflate
+#  define inflateBack           z_inflateBack
+#  define inflateBackEnd        z_inflateBackEnd
+#  define inflateBackInit_      z_inflateBackInit_
+#  define inflateCopy           z_inflateCopy
+#  define inflateEnd            z_inflateEnd
+#  define inflateGetHeader      z_inflateGetHeader
+#  define inflateInit2_         z_inflateInit2_
+#  define inflateInit_          z_inflateInit_
+#  define inflateMark           z_inflateMark
+#  define inflatePrime          z_inflatePrime
+#  define inflateReset          z_inflateReset
+#  define inflateReset2         z_inflateReset2
+#  define inflateSetDictionary  z_inflateSetDictionary
+#  define inflateGetDictionary  z_inflateGetDictionary
+#  define inflateSync           z_inflateSync
+#  define inflateSyncPoint      z_inflateSyncPoint
+#  define inflateUndermine      z_inflateUndermine
+#  define inflateResetKeep      z_inflateResetKeep
+#  define inflate_copyright     z_inflate_copyright
+#  define inflate_fast          z_inflate_fast
+#  define inflate_table         z_inflate_table
+#  ifndef Z_SOLO
+#    define uncompress            z_uncompress
+#  endif
+#  define zError                z_zError
+#  ifndef Z_SOLO
+#    define zcalloc               z_zcalloc
+#    define zcfree                z_zcfree
+#  endif
+#  define zlibCompileFlags      z_zlibCompileFlags
+#  define zlibVersion           z_zlibVersion
+
+/* all zlib typedefs in zlib.h and zconf.h */
+#  define Byte                  z_Byte
+#  define Bytef                 z_Bytef
+#  define alloc_func            z_alloc_func
+#  define charf                 z_charf
+#  define free_func             z_free_func
+#  ifndef Z_SOLO
+#    define gzFile                z_gzFile
+#  endif
+#  define gz_header             z_gz_header
+#  define gz_headerp            z_gz_headerp
+#  define in_func               z_in_func
+#  define intf                  z_intf
+#  define out_func              z_out_func
+#  define uInt                  z_uInt
+#  define uIntf                 z_uIntf
+#  define uLong                 z_uLong
+#  define uLongf                z_uLongf
+#  define voidp                 z_voidp
+#  define voidpc                z_voidpc
+#  define voidpf                z_voidpf
+
+/* all zlib structs in zlib.h and zconf.h */
+#  define gz_header_s           z_gz_header_s
+#  define internal_state        z_internal_state
+
+#endif
+
+#if defined(__MSDOS__) && !defined(MSDOS)
+#  define MSDOS
+#endif
+#if (defined(OS_2) || defined(__OS2__)) && !defined(OS2)
+#  define OS2
+#endif
+#if defined(_WINDOWS) && !defined(WINDOWS)
+#  define WINDOWS
+#endif
+#if defined(_WIN32) || defined(_WIN32_WCE) || defined(__WIN32__)
+#  ifndef WIN32
+#    define WIN32
+#  endif
+#endif
+#if (defined(MSDOS) || defined(OS2) || defined(WINDOWS)) && !defined(WIN32)
+#  if !defined(__GNUC__) && !defined(__FLAT__) && !defined(__386__)
+#    ifndef SYS16BIT
+#      define SYS16BIT
+#    endif
+#  endif
+#endif
+
+/*
+ * Compile with -DMAXSEG_64K if the alloc function cannot allocate more
+ * than 64k bytes at a time (needed on systems with 16-bit int).
+ */
+#ifdef SYS16BIT
+#  define MAXSEG_64K
+#endif
+#ifdef MSDOS
+#  define UNALIGNED_OK
+#endif
+
+#ifdef __STDC_VERSION__
+#  ifndef STDC
+#    define STDC
+#  endif
+#  if __STDC_VERSION__ >= 199901L
+#    ifndef STDC99
+#      define STDC99
+#    endif
+#  endif
+#endif
+#if !defined(STDC) && (defined(__STDC__) || defined(__cplusplus))
+#  define STDC
+#endif
+#if !defined(STDC) && (defined(__GNUC__) || defined(__BORLANDC__))
+#  define STDC
+#endif
+#if !defined(STDC) && (defined(MSDOS) || defined(WINDOWS) || defined(WIN32))
+#  define STDC
+#endif
+#if !defined(STDC) && (defined(OS2) || defined(__HOS_AIX__))
+#  define STDC
+#endif
+
+#if defined(__OS400__) && !defined(STDC)    /* iSeries (formerly AS/400). */
+#  define STDC
+#endif
+
+#ifndef STDC
+#  ifndef const /* cannot use !defined(STDC) && !defined(const) on Mac */
+#    define const       /* note: need a more gentle solution here */
+#  endif
+#endif
+
+#if defined(ZLIB_CONST) && !defined(z_const)
+#  define z_const const
+#else
+#  define z_const
+#endif
+
+/* Some Mac compilers merge all .h files incorrectly: */
+#if defined(__MWERKS__)||defined(applec)||defined(THINK_C)||defined(__SC__)
+#  define NO_DUMMY_DECL
+#endif
+
+/* Maximum value for memLevel in deflateInit2 */
+#ifndef MAX_MEM_LEVEL
+#  ifdef MAXSEG_64K
+#    define MAX_MEM_LEVEL 8
+#  else
+#    define MAX_MEM_LEVEL 9
+#  endif
+#endif
+
+/* Maximum value for windowBits in deflateInit2 and inflateInit2.
+ * WARNING: reducing MAX_WBITS makes minigzip unable to extract .gz files
+ * created by gzip. (Files created by minigzip can still be extracted by
+ * gzip.)
+ */
+#ifndef MAX_WBITS
+#  define MAX_WBITS   15 /* 32K LZ77 window */
+#endif
+
+/* The memory requirements for deflate are (in bytes):
+            (1 << (windowBits+2)) +  (1 << (memLevel+9))
+ that is: 128K for windowBits=15  +  128K for memLevel = 8  (default values)
+ plus a few kilobytes for small objects. For example, if you want to reduce
+ the default memory requirements from 256K to 128K, compile with
+     make CFLAGS="-O -DMAX_WBITS=14 -DMAX_MEM_LEVEL=7"
+ Of course this will generally degrade compression (there's no free lunch).
+
+   The memory requirements for inflate are (in bytes) 1 << windowBits
+ that is, 32K for windowBits=15 (default value) plus a few kilobytes
+ for small objects.
+*/
+
+                        /* Type declarations */
+
+#ifndef OF /* function prototypes */
+#  ifdef STDC
+#    define OF(args)  args
+#  else
+#    define OF(args)  ()
+#  endif
+#endif
+
+#ifndef Z_ARG /* function prototypes for stdarg */
+#  if defined(STDC) || defined(Z_HAVE_STDARG_H)
+#    define Z_ARG(args)  args
+#  else
+#    define Z_ARG(args)  ()
+#  endif
+#endif
+
+/* The following definitions for FAR are needed only for MSDOS mixed
+ * model programming (small or medium model with some far allocations).
+ * This was tested only with MSC; for other MSDOS compilers you may have
+ * to define NO_MEMCPY in zutil.h.  If you don't need the mixed model,
+ * just define FAR to be empty.
+ */
+#ifdef SYS16BIT
+#  if defined(M_I86SM) || defined(M_I86MM)
+     /* MSC small or medium model */
+#    define SMALL_MEDIUM
+#    ifdef _MSC_VER
+#      define FAR _far
+#    else
+#      define FAR far
+#    endif
+#  endif
+#  if (defined(__SMALL__) || defined(__MEDIUM__))
+     /* Turbo C small or medium model */
+#    define SMALL_MEDIUM
+#    ifdef __BORLANDC__
+#      define FAR _far
+#    else
+#      define FAR far
+#    endif
+#  endif
+#endif
+
+#if defined(WINDOWS) || defined(WIN32)
+   /* If building or using zlib as a DLL, define ZLIB_DLL.
+    * This is not mandatory, but it offers a little performance increase.
+    */
+#  ifdef ZLIB_DLL
+#    if defined(WIN32) && (!defined(__BORLANDC__) || (__BORLANDC__ >= 0x500))
+#      ifdef ZLIB_INTERNAL
+#        define ZEXTERN extern __declspec(dllexport)
+#      else
+#        define ZEXTERN extern __declspec(dllimport)
+#      endif
+#    endif
+#  endif  /* ZLIB_DLL */
+   /* If building or using zlib with the WINAPI/WINAPIV calling convention,
+    * define ZLIB_WINAPI.
+    * Caution: the standard ZLIB1.DLL is NOT compiled using ZLIB_WINAPI.
+    */
+#  ifdef ZLIB_WINAPI
+#    ifdef FAR
+#      undef FAR
+#    endif
+#    include <windows.h>
+     /* No need for _export, use ZLIB.DEF instead. */
+     /* For complete Windows compatibility, use WINAPI, not __stdcall. */
+#    define ZEXPORT WINAPI
+#    ifdef WIN32
+#      define ZEXPORTVA WINAPIV
+#    else
+#      define ZEXPORTVA FAR CDECL
+#    endif
+#  endif
+#endif
+
+#if defined (__BEOS__)
+#  ifdef ZLIB_DLL
+#    ifdef ZLIB_INTERNAL
+#      define ZEXPORT   __declspec(dllexport)
+#      define ZEXPORTVA __declspec(dllexport)
+#    else
+#      define ZEXPORT   __declspec(dllimport)
+#      define ZEXPORTVA __declspec(dllimport)
+#    endif
+#  endif
+#endif
+
+#ifndef ZEXTERN
+#  define ZEXTERN extern
+#endif
+#ifndef ZEXPORT
+#  define ZEXPORT
+#endif
+#ifndef ZEXPORTVA
+#  define ZEXPORTVA
+#endif
+
+#ifndef FAR
+#  define FAR
+#endif
+
+#if !defined(__MACTYPES__)
+typedef unsigned char  Byte;  /* 8 bits */
+#endif
+typedef unsigned int   uInt;  /* 16 bits or more */
+typedef unsigned long  uLong; /* 32 bits or more */
+
+#ifdef SMALL_MEDIUM
+   /* Borland C/C++ and some old MSC versions ignore FAR inside typedef */
+#  define Bytef Byte FAR
+#else
+   typedef Byte  FAR Bytef;
+#endif
+typedef char  FAR charf;
+typedef int   FAR intf;
+typedef uInt  FAR uIntf;
+typedef uLong FAR uLongf;
+
+#ifdef STDC
+   typedef void const *voidpc;
+   typedef void FAR   *voidpf;
+   typedef void       *voidp;
+#else
+   typedef Byte const *voidpc;
+   typedef Byte FAR   *voidpf;
+   typedef Byte       *voidp;
+#endif
+
+#if !defined(Z_U4) && !defined(Z_SOLO) && defined(STDC)
+#  include <limits.h>
+#  if (UINT_MAX == 0xffffffffUL)
+#    define Z_U4 unsigned
+#  elif (ULONG_MAX == 0xffffffffUL)
+#    define Z_U4 unsigned long
+#  elif (USHRT_MAX == 0xffffffffUL)
+#    define Z_U4 unsigned short
+#  endif
+#endif
+
+#ifdef Z_U4
+   typedef Z_U4 z_crc_t;
+#else
+   typedef unsigned long z_crc_t;
+#endif
+
+#ifdef HAVE_UNISTD_H    /* may be set to #if 1 by ./configure */
+#  define Z_HAVE_UNISTD_H
+#endif
+
+#ifdef HAVE_STDARG_H    /* may be set to #if 1 by ./configure */
+#  define Z_HAVE_STDARG_H
+#endif
+
+#ifdef STDC
+#  ifndef Z_SOLO
+#    include <sys/types.h>      /* for off_t */
+#  endif
+#endif
+
+#if defined(STDC) || defined(Z_HAVE_STDARG_H)
+#  ifndef Z_SOLO
+#    include <stdarg.h>         /* for va_list */
+#  endif
+#endif
+
+#ifdef _WIN32
+#  ifndef Z_SOLO
+#    include <stddef.h>         /* for wchar_t */
+#  endif
+#endif
+
+/* a little trick to accommodate both "#define _LARGEFILE64_SOURCE" and
+ * "#define _LARGEFILE64_SOURCE 1" as requesting 64-bit operations, (even
+ * though the former does not conform to the LFS document), but considering
+ * both "#undef _LARGEFILE64_SOURCE" and "#define _LARGEFILE64_SOURCE 0" as
+ * equivalently requesting no 64-bit operations
+ */
+#if defined(_LARGEFILE64_SOURCE) && -_LARGEFILE64_SOURCE - -1 == 1
+#  undef _LARGEFILE64_SOURCE
+#endif
+
+#if defined(__WATCOMC__) && !defined(Z_HAVE_UNISTD_H)
+#  define Z_HAVE_UNISTD_H
+#endif
+#ifndef Z_SOLO
+#  if defined(Z_HAVE_UNISTD_H) || defined(_LARGEFILE64_SOURCE)
+#    include <unistd.h>         /* for SEEK_*, off_t, and _LFS64_LARGEFILE */
+#    ifdef VMS
+#      include <unixio.h>       /* for off_t */
+#    endif
+#    ifndef z_off_t
+#      define z_off_t off_t
+#    endif
+#  endif
+#endif
+
+#if defined(_LFS64_LARGEFILE) && _LFS64_LARGEFILE-0
+#  define Z_LFS64
+#endif
+
+#if defined(_LARGEFILE64_SOURCE) && defined(Z_LFS64)
+#  define Z_LARGE64
+#endif
+
+#if defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS-0 == 64 && defined(Z_LFS64)
+#  define Z_WANT64
+#endif
+
+#if !defined(SEEK_SET) && !defined(Z_SOLO)
+#  define SEEK_SET        0       /* Seek from beginning of file.  */
+#  define SEEK_CUR        1       /* Seek from current position.  */
+#  define SEEK_END        2       /* Set file pointer to EOF plus "offset" */
+#endif
+
+#ifndef z_off_t
+#  define z_off_t long
+#endif
+
+#if !defined(_WIN32) && defined(Z_LARGE64)
+#  define z_off64_t off64_t
+#else
+#  if defined(_WIN32) && !defined(__GNUC__) && !defined(Z_SOLO)
+#    define z_off64_t __int64
+#  else
+#    define z_off64_t z_off_t
+#  endif
+#endif
+
+/* MVS linker does not support external names larger than 8 bytes */
+#if defined(__MVS__)
+  #pragma map(deflateInit_,"DEIN")
+  #pragma map(deflateInit2_,"DEIN2")
+  #pragma map(deflateEnd,"DEEND")
+  #pragma map(deflateBound,"DEBND")
+  #pragma map(inflateInit_,"ININ")
+  #pragma map(inflateInit2_,"ININ2")
+  #pragma map(inflateEnd,"INEND")
+  #pragma map(inflateSync,"INSY")
+  #pragma map(inflateSetDictionary,"INSEDI")
+  #pragma map(compressBound,"CMBND")
+  #pragma map(inflate_table,"INTABL")
+  #pragma map(inflate_fast,"INFA")
+  #pragma map(inflate_copyright,"INCOPY")
+#endif
+
+#endif /* ZCONF_H */
diff --git a/ml/dlib/dlib/external/zlib/zlib.h b/ml/dlib/dlib/external/zlib/zlib.h
new file mode 100644
index 000000000..3e0c7672a
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/zlib.h
@@ -0,0 +1,1768 @@
+/* zlib.h -- interface of the 'zlib' general purpose compression library
+  version 1.2.8, April 28th, 2013
+
+  Copyright (C) 1995-2013 Jean-loup Gailly and Mark Adler
+
+  This software is provided 'as-is', without any express or implied
+  warranty.  In no event will the authors be held liable for any damages
+  arising from the use of this software.
+
+  Permission is granted to anyone to use this software for any purpose,
+  including commercial applications, and to alter it and redistribute it
+  freely, subject to the following restrictions:
+
+  1. The origin of this software must not be misrepresented; you must not
+     claim that you wrote the original software. If you use this software
+     in a product, an acknowledgment in the product documentation would be
+     appreciated but is not required.
+  2. Altered source versions must be plainly marked as such, and must not be
+     misrepresented as being the original software.
+  3. This notice may not be removed or altered from any source distribution.
+
+  Jean-loup Gailly        Mark Adler
+  jloup@gzip.org          madler@alumni.caltech.edu
+
+
+  The data format used by the zlib library is described by RFCs (Request for
+  Comments) 1950 to 1952 in the files http://tools.ietf.org/html/rfc1950
+  (zlib format), rfc1951 (deflate format) and rfc1952 (gzip format).
+*/
+
+#ifndef ZLIB_H
+#define ZLIB_H
+
+#include "zconf.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define ZLIB_VERSION "1.2.8"
+#define ZLIB_VERNUM 0x1280
+#define ZLIB_VER_MAJOR 1
+#define ZLIB_VER_MINOR 2
+#define ZLIB_VER_REVISION 8
+#define ZLIB_VER_SUBREVISION 0
+
+/*
+    The 'zlib' compression library provides in-memory compression and
+  decompression functions, including integrity checks of the uncompressed data.
+  This version of the library supports only one compression method (deflation)
+  but other algorithms will be added later and will have the same stream
+  interface.
+
+    Compression can be done in a single step if the buffers are large enough,
+  or can be done by repeated calls of the compression function.  In the latter
+  case, the application must provide more input and/or consume the output
+  (providing more output space) before each call.
+
+    The compressed data format used by default by the in-memory functions is
+  the zlib format, which is a zlib wrapper documented in RFC 1950, wrapped
+  around a deflate stream, which is itself documented in RFC 1951.
+
+    The library also supports reading and writing files in gzip (.gz) format
+  with an interface similar to that of stdio using the functions that start
+  with "gz".  The gzip format is different from the zlib format.  gzip is a
+  gzip wrapper, documented in RFC 1952, wrapped around a deflate stream.
+
+    This library can optionally read and write gzip streams in memory as well.
+
+    The zlib format was designed to be compact and fast for use in memory
+  and on communications channels.  The gzip format was designed for single-
+  file compression on file systems, has a larger header than zlib to maintain
+  directory information, and uses a different, slower check method than zlib.
+
+    The library does not install any signal handler.  The decoder checks
+  the consistency of the compressed data, so the library should never crash
+  even in case of corrupted input.
+*/
+
+typedef voidpf (*alloc_func) OF((voidpf opaque, uInt items, uInt size));
+typedef void   (*free_func)  OF((voidpf opaque, voidpf address));
+
+struct internal_state;
+
+typedef struct z_stream_s {
+    z_const Bytef *next_in;     /* next input byte */
+    uInt     avail_in;  /* number of bytes available at next_in */
+    uLong    total_in;  /* total number of input bytes read so far */
+
+    Bytef    *next_out; /* next output byte should be put there */
+    uInt     avail_out; /* remaining free space at next_out */
+    uLong    total_out; /* total number of bytes output so far */
+
+    z_const char *msg;  /* last error message, NULL if no error */
+    struct internal_state FAR *state; /* not visible by applications */
+
+    alloc_func zalloc;  /* used to allocate the internal state */
+    free_func  zfree;   /* used to free the internal state */
+    voidpf     opaque;  /* private data object passed to zalloc and zfree */
+
+    int     data_type;  /* best guess about the data type: binary or text */
+    uLong   adler;      /* adler32 value of the uncompressed data */
+    uLong   reserved;   /* reserved for future use */
+} z_stream;
+
+typedef z_stream FAR *z_streamp;
+
+/*
+     gzip header information passed to and from zlib routines.  See RFC 1952
+  for more details on the meanings of these fields.
+*/
+typedef struct gz_header_s {
+    int     text;       /* true if compressed data believed to be text */
+    uLong   time;       /* modification time */
+    int     xflags;     /* extra flags (not used when writing a gzip file) */
+    int     os;         /* operating system */
+    Bytef   *extra;     /* pointer to extra field or Z_NULL if none */
+    uInt    extra_len;  /* extra field length (valid if extra != Z_NULL) */
+    uInt    extra_max;  /* space at extra (only when reading header) */
+    Bytef   *name;      /* pointer to zero-terminated file name or Z_NULL */
+    uInt    name_max;   /* space at name (only when reading header) */
+    Bytef   *comment;   /* pointer to zero-terminated comment or Z_NULL */
+    uInt    comm_max;   /* space at comment (only when reading header) */
+    int     hcrc;       /* true if there was or will be a header crc */
+    int     done;       /* true when done reading gzip header (not used
+                           when writing a gzip file) */
+} gz_header;
+
+typedef gz_header FAR *gz_headerp;
+
+/*
+     The application must update next_in and avail_in when avail_in has dropped
+   to zero.  It must update next_out and avail_out when avail_out has dropped
+   to zero.  The application must initialize zalloc, zfree and opaque before
+   calling the init function.  All other fields are set by the compression
+   library and must not be updated by the application.
+
+     The opaque value provided by the application will be passed as the first
+   parameter for calls of zalloc and zfree.  This can be useful for custom
+   memory management.  The compression library attaches no meaning to the
+   opaque value.
+
+     zalloc must return Z_NULL if there is not enough memory for the object.
+   If zlib is used in a multi-threaded application, zalloc and zfree must be
+   thread safe.
+
+     On 16-bit systems, the functions zalloc and zfree must be able to allocate
+   exactly 65536 bytes, but will not be required to allocate more than this if
+   the symbol MAXSEG_64K is defined (see zconf.h).  WARNING: On MSDOS, pointers
+   returned by zalloc for objects of exactly 65536 bytes *must* have their
+   offset normalized to zero.  The default allocation function provided by this
+   library ensures this (see zutil.c).  To reduce memory requirements and avoid
+   any allocation of 64K objects, at the expense of compression ratio, compile
+   the library with -DMAX_WBITS=14 (see zconf.h).
+
+     The fields total_in and total_out can be used for statistics or progress
+   reports.  After compression, total_in holds the total size of the
+   uncompressed data and may be saved for use in the decompressor (particularly
+   if the decompressor wants to decompress everything in a single step).
+*/
+
+                        /* constants */
+
+#define Z_NO_FLUSH      0
+#define Z_PARTIAL_FLUSH 1
+#define Z_SYNC_FLUSH    2
+#define Z_FULL_FLUSH    3
+#define Z_FINISH        4
+#define Z_BLOCK         5
+#define Z_TREES         6
+/* Allowed flush values; see deflate() and inflate() below for details */
+
+#define Z_OK            0
+#define Z_STREAM_END    1
+#define Z_NEED_DICT     2
+#define Z_ERRNO        (-1)
+#define Z_STREAM_ERROR (-2)
+#define Z_DATA_ERROR   (-3)
+#define Z_MEM_ERROR    (-4)
+#define Z_BUF_ERROR    (-5)
+#define Z_VERSION_ERROR (-6)
+/* Return codes for the compression/decompression functions. Negative values
+ * are errors, positive values are used for special but normal events.
+ */
+
+#define Z_NO_COMPRESSION         0
+#define Z_BEST_SPEED             1
+#define Z_BEST_COMPRESSION       9
+#define Z_DEFAULT_COMPRESSION  (-1)
+/* compression levels */
+
+#define Z_FILTERED            1
+#define Z_HUFFMAN_ONLY        2
+#define Z_RLE                 3
+#define Z_FIXED               4
+#define Z_DEFAULT_STRATEGY    0
+/* compression strategy; see deflateInit2() below for details */
+
+#define Z_BINARY   0
+#define Z_TEXT     1
+#define Z_ASCII    Z_TEXT   /* for compatibility with 1.2.2 and earlier */
+#define Z_UNKNOWN  2
+/* Possible values of the data_type field (though see inflate()) */
+
+#define Z_DEFLATED   8
+/* The deflate compression method (the only one supported in this version) */
+
+#define Z_NULL  0  /* for initializing zalloc, zfree, opaque */
+
+#define zlib_version zlibVersion()
+/* for compatibility with versions < 1.0.2 */
+
+
+                        /* basic functions */
+
+ZEXTERN const char * ZEXPORT zlibVersion OF((void));
+/* The application can compare zlibVersion and ZLIB_VERSION for consistency.
+   If the first character differs, the library code actually used is not
+   compatible with the zlib.h header file used by the application.  This check
+   is automatically made by deflateInit and inflateInit.
+ */
+
+/*
+ZEXTERN int ZEXPORT deflateInit OF((z_streamp strm, int level));
+
+     Initializes the internal stream state for compression.  The fields
+   zalloc, zfree and opaque must be initialized before by the caller.  If
+   zalloc and zfree are set to Z_NULL, deflateInit updates them to use default
+   allocation functions.
+
+     The compression level must be Z_DEFAULT_COMPRESSION, or between 0 and 9:
+   1 gives best speed, 9 gives best compression, 0 gives no compression at all
+   (the input data is simply copied a block at a time).  Z_DEFAULT_COMPRESSION
+   requests a default compromise between speed and compression (currently
+   equivalent to level 6).
+
+     deflateInit returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_STREAM_ERROR if level is not a valid compression level, or
+   Z_VERSION_ERROR if the zlib library version (zlib_version) is incompatible
+   with the version assumed by the caller (ZLIB_VERSION).  msg is set to null
+   if there is no error message.  deflateInit does not perform any compression:
+   this will be done by deflate().
+*/
+
+
+ZEXTERN int ZEXPORT deflate OF((z_streamp strm, int flush));
+/*
+    deflate compresses as much data as possible, and stops when the input
+  buffer becomes empty or the output buffer becomes full.  It may introduce
+  some output latency (reading input without producing any output) except when
+  forced to flush.
+
+    The detailed semantics are as follows.  deflate performs one or both of the
+  following actions:
+
+  - Compress more input starting at next_in and update next_in and avail_in
+    accordingly.  If not all input can be processed (because there is not
+    enough room in the output buffer), next_in and avail_in are updated and
+    processing will resume at this point for the next call of deflate().
+
+  - Provide more output starting at next_out and update next_out and avail_out
+    accordingly.  This action is forced if the parameter flush is non zero.
+    Forcing flush frequently degrades the compression ratio, so this parameter
+    should be set only when necessary (in interactive applications).  Some
+    output may be provided even if flush is not set.
+
+    Before the call of deflate(), the application should ensure that at least
+  one of the actions is possible, by providing more input and/or consuming more
+  output, and updating avail_in or avail_out accordingly; avail_out should
+  never be zero before the call.  The application can consume the compressed
+  output when it wants, for example when the output buffer is full (avail_out
+  == 0), or after each call of deflate().  If deflate returns Z_OK and with
+  zero avail_out, it must be called again after making room in the output
+  buffer because there might be more output pending.
+
+    Normally the parameter flush is set to Z_NO_FLUSH, which allows deflate to
+  decide how much data to accumulate before producing output, in order to
+  maximize compression.
+
+    If the parameter flush is set to Z_SYNC_FLUSH, all pending output is
+  flushed to the output buffer and the output is aligned on a byte boundary, so
+  that the decompressor can get all input data available so far.  (In
+  particular avail_in is zero after the call if enough output space has been
+  provided before the call.) Flushing may degrade compression for some
+  compression algorithms and so it should be used only when necessary.  This
+  completes the current deflate block and follows it with an empty stored block
+  that is three bits plus filler bits to the next byte, followed by four bytes
+  (00 00 ff ff).
+
+    If flush is set to Z_PARTIAL_FLUSH, all pending output is flushed to the
+  output buffer, but the output is not aligned to a byte boundary.  All of the
+  input data so far will be available to the decompressor, as for Z_SYNC_FLUSH.
+  This completes the current deflate block and follows it with an empty fixed
+  codes block that is 10 bits long.  This assures that enough bytes are output
+  in order for the decompressor to finish the block before the empty fixed code
+  block.
+
+    If flush is set to Z_BLOCK, a deflate block is completed and emitted, as
+  for Z_SYNC_FLUSH, but the output is not aligned on a byte boundary, and up to
+  seven bits of the current block are held to be written as the next byte after
+  the next deflate block is completed.  In this case, the decompressor may not
+  be provided enough bits at this point in order to complete decompression of
+  the data provided so far to the compressor.  It may need to wait for the next
+  block to be emitted.  This is for advanced applications that need to control
+  the emission of deflate blocks.
+
+    If flush is set to Z_FULL_FLUSH, all output is flushed as with
+  Z_SYNC_FLUSH, and the compression state is reset so that decompression can
+  restart from this point if previous compressed data has been damaged or if
+  random access is desired.  Using Z_FULL_FLUSH too often can seriously degrade
+  compression.
+
+    If deflate returns with avail_out == 0, this function must be called again
+  with the same value of the flush parameter and more output space (updated
+  avail_out), until the flush is complete (deflate returns with non-zero
+  avail_out).  In the case of a Z_FULL_FLUSH or Z_SYNC_FLUSH, make sure that
+  avail_out is greater than six to avoid repeated flush markers due to
+  avail_out == 0 on return.
+
+    If the parameter flush is set to Z_FINISH, pending input is processed,
+  pending output is flushed and deflate returns with Z_STREAM_END if there was
+  enough output space; if deflate returns with Z_OK, this function must be
+  called again with Z_FINISH and more output space (updated avail_out) but no
+  more input data, until it returns with Z_STREAM_END or an error.  After
+  deflate has returned Z_STREAM_END, the only possible operations on the stream
+  are deflateReset or deflateEnd.
+
+    Z_FINISH can be used immediately after deflateInit if all the compression
+  is to be done in a single step.  In this case, avail_out must be at least the
+  value returned by deflateBound (see below).  Then deflate is guaranteed to
+  return Z_STREAM_END.  If not enough output space is provided, deflate will
+  not return Z_STREAM_END, and it must be called again as described above.
+
+    deflate() sets strm->adler to the adler32 checksum of all input read
+  so far (that is, total_in bytes).
+
+    deflate() may update strm->data_type if it can make a good guess about
+  the input data type (Z_BINARY or Z_TEXT).  In doubt, the data is considered
+  binary.  This field is only for information purposes and does not affect the
+  compression algorithm in any manner.
+
+    deflate() returns Z_OK if some progress has been made (more input
+  processed or more output produced), Z_STREAM_END if all input has been
+  consumed and all output has been produced (only when flush is set to
+  Z_FINISH), Z_STREAM_ERROR if the stream state was inconsistent (for example
+  if next_in or next_out was Z_NULL), Z_BUF_ERROR if no progress is possible
+  (for example avail_in or avail_out was zero).  Note that Z_BUF_ERROR is not
+  fatal, and deflate() can be called again with more input and more output
+  space to continue compressing.
+*/
+
+
+ZEXTERN int ZEXPORT deflateEnd OF((z_streamp strm));
+/*
+     All dynamically allocated data structures for this stream are freed.
+   This function discards any unprocessed input and does not flush any pending
+   output.
+
+     deflateEnd returns Z_OK if success, Z_STREAM_ERROR if the
+   stream state was inconsistent, Z_DATA_ERROR if the stream was freed
+   prematurely (some input or output was discarded).  In the error case, msg
+   may be set but then points to a static string (which must not be
+   deallocated).
+*/
+
+
+/*
+ZEXTERN int ZEXPORT inflateInit OF((z_streamp strm));
+
+     Initializes the internal stream state for decompression.  The fields
+   next_in, avail_in, zalloc, zfree and opaque must be initialized before by
+   the caller.  If next_in is not Z_NULL and avail_in is large enough (the
+   exact value depends on the compression method), inflateInit determines the
+   compression method from the zlib header and allocates all data structures
+   accordingly; otherwise the allocation will be deferred to the first call of
+   inflate.  If zalloc and zfree are set to Z_NULL, inflateInit updates them to
+   use default allocation functions.
+
+     inflateInit returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_VERSION_ERROR if the zlib library version is incompatible with the
+   version assumed by the caller, or Z_STREAM_ERROR if the parameters are
+   invalid, such as a null pointer to the structure.  msg is set to null if
+   there is no error message.  inflateInit does not perform any decompression
+   apart from possibly reading the zlib header if present: actual decompression
+   will be done by inflate().  (So next_in and avail_in may be modified, but
+   next_out and avail_out are unused and unchanged.) The current implementation
+   of inflateInit() does not process any header information -- that is deferred
+   until inflate() is called.
+*/
+
+
+ZEXTERN int ZEXPORT inflate OF((z_streamp strm, int flush));
+/*
+    inflate decompresses as much data as possible, and stops when the input
+  buffer becomes empty or the output buffer becomes full.  It may introduce
+  some output latency (reading input without producing any output) except when
+  forced to flush.
+
+  The detailed semantics are as follows.  inflate performs one or both of the
+  following actions:
+
+  - Decompress more input starting at next_in and update next_in and avail_in
+    accordingly.  If not all input can be processed (because there is not
+    enough room in the output buffer), next_in is updated and processing will
+    resume at this point for the next call of inflate().
+
+  - Provide more output starting at next_out and update next_out and avail_out
+    accordingly.  inflate() provides as much output as possible, until there is
+    no more input data or no more space in the output buffer (see below about
+    the flush parameter).
+
+    Before the call of inflate(), the application should ensure that at least
+  one of the actions is possible, by providing more input and/or consuming more
+  output, and updating the next_* and avail_* values accordingly.  The
+  application can consume the uncompressed output when it wants, for example
+  when the output buffer is full (avail_out == 0), or after each call of
+  inflate().  If inflate returns Z_OK and with zero avail_out, it must be
+  called again after making room in the output buffer because there might be
+  more output pending.
+
+    The flush parameter of inflate() can be Z_NO_FLUSH, Z_SYNC_FLUSH, Z_FINISH,
+  Z_BLOCK, or Z_TREES.  Z_SYNC_FLUSH requests that inflate() flush as much
+  output as possible to the output buffer.  Z_BLOCK requests that inflate()
+  stop if and when it gets to the next deflate block boundary.  When decoding
+  the zlib or gzip format, this will cause inflate() to return immediately
+  after the header and before the first block.  When doing a raw inflate,
+  inflate() will go ahead and process the first block, and will return when it
+  gets to the end of that block, or when it runs out of data.
+
+    The Z_BLOCK option assists in appending to or combining deflate streams.
+  Also to assist in this, on return inflate() will set strm->data_type to the
+  number of unused bits in the last byte taken from strm->next_in, plus 64 if
+  inflate() is currently decoding the last block in the deflate stream, plus
+  128 if inflate() returned immediately after decoding an end-of-block code or
+  decoding the complete header up to just before the first byte of the deflate
+  stream.  The end-of-block will not be indicated until all of the uncompressed
+  data from that block has been written to strm->next_out.  The number of
+  unused bits may in general be greater than seven, except when bit 7 of
+  data_type is set, in which case the number of unused bits will be less than
+  eight.  data_type is set as noted here every time inflate() returns for all
+  flush options, and so can be used to determine the amount of currently
+  consumed input in bits.
+
+    The Z_TREES option behaves as Z_BLOCK does, but it also returns when the
+  end of each deflate block header is reached, before any actual data in that
+  block is decoded.  This allows the caller to determine the length of the
+  deflate block header for later use in random access within a deflate block.
+  256 is added to the value of strm->data_type when inflate() returns
+  immediately after reaching the end of the deflate block header.
+
+    inflate() should normally be called until it returns Z_STREAM_END or an
+  error.  However if all decompression is to be performed in a single step (a
+  single call of inflate), the parameter flush should be set to Z_FINISH.  In
+  this case all pending input is processed and all pending output is flushed;
+  avail_out must be large enough to hold all of the uncompressed data for the
+  operation to complete.  (The size of the uncompressed data may have been
+  saved by the compressor for this purpose.) The use of Z_FINISH is not
+  required to perform an inflation in one step.  However it may be used to
+  inform inflate that a faster approach can be used for the single inflate()
+  call.  Z_FINISH also informs inflate to not maintain a sliding window if the
+  stream completes, which reduces inflate's memory footprint.  If the stream
+  does not complete, either because not all of the stream is provided or not
+  enough output space is provided, then a sliding window will be allocated and
+  inflate() can be called again to continue the operation as if Z_NO_FLUSH had
+  been used.
+
+     In this implementation, inflate() always flushes as much output as
+  possible to the output buffer, and always uses the faster approach on the
+  first call.  So the effects of the flush parameter in this implementation are
+  on the return value of inflate() as noted below, when inflate() returns early
+  when Z_BLOCK or Z_TREES is used, and when inflate() avoids the allocation of
+  memory for a sliding window when Z_FINISH is used.
+
+     If a preset dictionary is needed after this call (see inflateSetDictionary
+  below), inflate sets strm->adler to the Adler-32 checksum of the dictionary
+  chosen by the compressor and returns Z_NEED_DICT; otherwise it sets
+  strm->adler to the Adler-32 checksum of all output produced so far (that is,
+  total_out bytes) and returns Z_OK, Z_STREAM_END or an error code as described
+  below.  At the end of the stream, inflate() checks that its computed adler32
+  checksum is equal to that saved by the compressor and returns Z_STREAM_END
+  only if the checksum is correct.
+
+    inflate() can decompress and check either zlib-wrapped or gzip-wrapped
+  deflate data.  The header type is detected automatically, if requested when
+  initializing with inflateInit2().  Any information contained in the gzip
+  header is not retained, so applications that need that information should
+  instead use raw inflate, see inflateInit2() below, or inflateBack() and
+  perform their own processing of the gzip header and trailer.  When processing
+  gzip-wrapped deflate data, strm->adler32 is set to the CRC-32 of the output
+  producted so far.  The CRC-32 is checked against the gzip trailer.
+
+    inflate() returns Z_OK if some progress has been made (more input processed
+  or more output produced), Z_STREAM_END if the end of the compressed data has
+  been reached and all uncompressed output has been produced, Z_NEED_DICT if a
+  preset dictionary is needed at this point, Z_DATA_ERROR if the input data was
+  corrupted (input stream not conforming to the zlib format or incorrect check
+  value), Z_STREAM_ERROR if the stream structure was inconsistent (for example
+  next_in or next_out was Z_NULL), Z_MEM_ERROR if there was not enough memory,
+  Z_BUF_ERROR if no progress is possible or if there was not enough room in the
+  output buffer when Z_FINISH is used.  Note that Z_BUF_ERROR is not fatal, and
+  inflate() can be called again with more input and more output space to
+  continue decompressing.  If Z_DATA_ERROR is returned, the application may
+  then call inflateSync() to look for a good compression block if a partial
+  recovery of the data is desired.
+*/
+
+
+ZEXTERN int ZEXPORT inflateEnd OF((z_streamp strm));
+/*
+     All dynamically allocated data structures for this stream are freed.
+   This function discards any unprocessed input and does not flush any pending
+   output.
+
+     inflateEnd returns Z_OK if success, Z_STREAM_ERROR if the stream state
+   was inconsistent.  In the error case, msg may be set but then points to a
+   static string (which must not be deallocated).
+*/
+
+
+                        /* Advanced functions */
+
+/*
+    The following functions are needed only in some special applications.
+*/
+
+/*
+ZEXTERN int ZEXPORT deflateInit2 OF((z_streamp strm,
+                                     int  level,
+                                     int  method,
+                                     int  windowBits,
+                                     int  memLevel,
+                                     int  strategy));
+
+     This is another version of deflateInit with more compression options.  The
+   fields next_in, zalloc, zfree and opaque must be initialized before by the
+   caller.
+
+     The method parameter is the compression method.  It must be Z_DEFLATED in
+   this version of the library.
+
+     The windowBits parameter is the base two logarithm of the window size
+   (the size of the history buffer).  It should be in the range 8..15 for this
+   version of the library.  Larger values of this parameter result in better
+   compression at the expense of memory usage.  The default value is 15 if
+   deflateInit is used instead.
+
+     windowBits can also be -8..-15 for raw deflate.  In this case, -windowBits
+   determines the window size.  deflate() will then generate raw deflate data
+   with no zlib header or trailer, and will not compute an adler32 check value.
+
+     windowBits can also be greater than 15 for optional gzip encoding.  Add
+   16 to windowBits to write a simple gzip header and trailer around the
+   compressed data instead of a zlib wrapper.  The gzip header will have no
+   file name, no extra data, no comment, no modification time (set to zero), no
+   header crc, and the operating system will be set to 255 (unknown).  If a
+   gzip stream is being written, strm->adler is a crc32 instead of an adler32.
+
+     The memLevel parameter specifies how much memory should be allocated
+   for the internal compression state.  memLevel=1 uses minimum memory but is
+   slow and reduces compression ratio; memLevel=9 uses maximum memory for
+   optimal speed.  The default value is 8.  See zconf.h for total memory usage
+   as a function of windowBits and memLevel.
+
+     The strategy parameter is used to tune the compression algorithm.  Use the
+   value Z_DEFAULT_STRATEGY for normal data, Z_FILTERED for data produced by a
+   filter (or predictor), Z_HUFFMAN_ONLY to force Huffman encoding only (no
+   string match), or Z_RLE to limit match distances to one (run-length
+   encoding).  Filtered data consists mostly of small values with a somewhat
+   random distribution.  In this case, the compression algorithm is tuned to
+   compress them better.  The effect of Z_FILTERED is to force more Huffman
+   coding and less string matching; it is somewhat intermediate between
+   Z_DEFAULT_STRATEGY and Z_HUFFMAN_ONLY.  Z_RLE is designed to be almost as
+   fast as Z_HUFFMAN_ONLY, but give better compression for PNG image data.  The
+   strategy parameter only affects the compression ratio but not the
+   correctness of the compressed output even if it is not set appropriately.
+   Z_FIXED prevents the use of dynamic Huffman codes, allowing for a simpler
+   decoder for special applications.
+
+     deflateInit2 returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_STREAM_ERROR if any parameter is invalid (such as an invalid
+   method), or Z_VERSION_ERROR if the zlib library version (zlib_version) is
+   incompatible with the version assumed by the caller (ZLIB_VERSION).  msg is
+   set to null if there is no error message.  deflateInit2 does not perform any
+   compression: this will be done by deflate().
+*/
+
+ZEXTERN int ZEXPORT deflateSetDictionary OF((z_streamp strm,
+                                             const Bytef *dictionary,
+                                             uInt  dictLength));
+/*
+     Initializes the compression dictionary from the given byte sequence
+   without producing any compressed output.  When using the zlib format, this
+   function must be called immediately after deflateInit, deflateInit2 or
+   deflateReset, and before any call of deflate.  When doing raw deflate, this
+   function must be called either before any call of deflate, or immediately
+   after the completion of a deflate block, i.e. after all input has been
+   consumed and all output has been delivered when using any of the flush
+   options Z_BLOCK, Z_PARTIAL_FLUSH, Z_SYNC_FLUSH, or Z_FULL_FLUSH.  The
+   compressor and decompressor must use exactly the same dictionary (see
+   inflateSetDictionary).
+
+     The dictionary should consist of strings (byte sequences) that are likely
+   to be encountered later in the data to be compressed, with the most commonly
+   used strings preferably put towards the end of the dictionary.  Using a
+   dictionary is most useful when the data to be compressed is short and can be
+   predicted with good accuracy; the data can then be compressed better than
+   with the default empty dictionary.
+
+     Depending on the size of the compression data structures selected by
+   deflateInit or deflateInit2, a part of the dictionary may in effect be
+   discarded, for example if the dictionary is larger than the window size
+   provided in deflateInit or deflateInit2.  Thus the strings most likely to be
+   useful should be put at the end of the dictionary, not at the front.  In
+   addition, the current implementation of deflate will use at most the window
+   size minus 262 bytes of the provided dictionary.
+
+     Upon return of this function, strm->adler is set to the adler32 value
+   of the dictionary; the decompressor may later use this value to determine
+   which dictionary has been used by the compressor.  (The adler32 value
+   applies to the whole dictionary even if only a subset of the dictionary is
+   actually used by the compressor.) If a raw deflate was requested, then the
+   adler32 value is not computed and strm->adler is not set.
+
+     deflateSetDictionary returns Z_OK if success, or Z_STREAM_ERROR if a
+   parameter is invalid (e.g.  dictionary being Z_NULL) or the stream state is
+   inconsistent (for example if deflate has already been called for this stream
+   or if not at a block boundary for raw deflate).  deflateSetDictionary does
+   not perform any compression: this will be done by deflate().
+*/
+
+ZEXTERN int ZEXPORT deflateCopy OF((z_streamp dest,
+                                    z_streamp source));
+/*
+     Sets the destination stream as a complete copy of the source stream.
+
+     This function can be useful when several compression strategies will be
+   tried, for example when there are several ways of pre-processing the input
+   data with a filter.  The streams that will be discarded should then be freed
+   by calling deflateEnd.  Note that deflateCopy duplicates the internal
+   compression state which can be quite large, so this strategy is slow and can
+   consume lots of memory.
+
+     deflateCopy returns Z_OK if success, Z_MEM_ERROR if there was not
+   enough memory, Z_STREAM_ERROR if the source stream state was inconsistent
+   (such as zalloc being Z_NULL).  msg is left unchanged in both source and
+   destination.
+*/
+
+ZEXTERN int ZEXPORT deflateReset OF((z_streamp strm));
+/*
+     This function is equivalent to deflateEnd followed by deflateInit,
+   but does not free and reallocate all the internal compression state.  The
+   stream will keep the same compression level and any other attributes that
+   may have been set by deflateInit2.
+
+     deflateReset returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent (such as zalloc or state being Z_NULL).
+*/
+
+ZEXTERN int ZEXPORT deflateParams OF((z_streamp strm,
+                                      int level,
+                                      int strategy));
+/*
+     Dynamically update the compression level and compression strategy.  The
+   interpretation of level and strategy is as in deflateInit2.  This can be
+   used to switch between compression and straight copy of the input data, or
+   to switch to a different kind of input data requiring a different strategy.
+   If the compression level is changed, the input available so far is
+   compressed with the old level (and may be flushed); the new level will take
+   effect only at the next call of deflate().
+
+     Before the call of deflateParams, the stream state must be set as for
+   a call of deflate(), since the currently available input may have to be
+   compressed and flushed.  In particular, strm->avail_out must be non-zero.
+
+     deflateParams returns Z_OK if success, Z_STREAM_ERROR if the source
+   stream state was inconsistent or if a parameter was invalid, Z_BUF_ERROR if
+   strm->avail_out was zero.
+*/
+
+ZEXTERN int ZEXPORT deflateTune OF((z_streamp strm,
+                                    int good_length,
+                                    int max_lazy,
+                                    int nice_length,
+                                    int max_chain));
+/*
+     Fine tune deflate's internal compression parameters.  This should only be
+   used by someone who understands the algorithm used by zlib's deflate for
+   searching for the best matching string, and even then only by the most
+   fanatic optimizer trying to squeeze out the last compressed bit for their
+   specific input data.  Read the deflate.c source code for the meaning of the
+   max_lazy, good_length, nice_length, and max_chain parameters.
+
+     deflateTune() can be called after deflateInit() or deflateInit2(), and
+   returns Z_OK on success, or Z_STREAM_ERROR for an invalid deflate stream.
+ */
+
+ZEXTERN uLong ZEXPORT deflateBound OF((z_streamp strm,
+                                       uLong sourceLen));
+/*
+     deflateBound() returns an upper bound on the compressed size after
+   deflation of sourceLen bytes.  It must be called after deflateInit() or
+   deflateInit2(), and after deflateSetHeader(), if used.  This would be used
+   to allocate an output buffer for deflation in a single pass, and so would be
+   called before deflate().  If that first deflate() call is provided the
+   sourceLen input bytes, an output buffer allocated to the size returned by
+   deflateBound(), and the flush value Z_FINISH, then deflate() is guaranteed
+   to return Z_STREAM_END.  Note that it is possible for the compressed size to
+   be larger than the value returned by deflateBound() if flush options other
+   than Z_FINISH or Z_NO_FLUSH are used.
+*/
+
+ZEXTERN int ZEXPORT deflatePending OF((z_streamp strm,
+                                       unsigned *pending,
+                                       int *bits));
+/*
+     deflatePending() returns the number of bytes and bits of output that have
+   been generated, but not yet provided in the available output.  The bytes not
+   provided would be due to the available output space having being consumed.
+   The number of bits of output not provided are between 0 and 7, where they
+   await more bits to join them in order to fill out a full byte.  If pending
+   or bits are Z_NULL, then those values are not set.
+
+     deflatePending returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent.
+ */
+
+ZEXTERN int ZEXPORT deflatePrime OF((z_streamp strm,
+                                     int bits,
+                                     int value));
+/*
+     deflatePrime() inserts bits in the deflate output stream.  The intent
+   is that this function is used to start off the deflate output with the bits
+   leftover from a previous deflate stream when appending to it.  As such, this
+   function can only be used for raw deflate, and must be used before the first
+   deflate() call after a deflateInit2() or deflateReset().  bits must be less
+   than or equal to 16, and that many of the least significant bits of value
+   will be inserted in the output.
+
+     deflatePrime returns Z_OK if success, Z_BUF_ERROR if there was not enough
+   room in the internal buffer to insert the bits, or Z_STREAM_ERROR if the
+   source stream state was inconsistent.
+*/
+
+ZEXTERN int ZEXPORT deflateSetHeader OF((z_streamp strm,
+                                         gz_headerp head));
+/*
+     deflateSetHeader() provides gzip header information for when a gzip
+   stream is requested by deflateInit2().  deflateSetHeader() may be called
+   after deflateInit2() or deflateReset() and before the first call of
+   deflate().  The text, time, os, extra field, name, and comment information
+   in the provided gz_header structure are written to the gzip header (xflag is
+   ignored -- the extra flags are set according to the compression level).  The
+   caller must assure that, if not Z_NULL, name and comment are terminated with
+   a zero byte, and that if extra is not Z_NULL, that extra_len bytes are
+   available there.  If hcrc is true, a gzip header crc is included.  Note that
+   the current versions of the command-line version of gzip (up through version
+   1.3.x) do not support header crc's, and will report that it is a "multi-part
+   gzip file" and give up.
+
+     If deflateSetHeader is not used, the default gzip header has text false,
+   the time set to zero, and os set to 255, with no extra, name, or comment
+   fields.  The gzip header is returned to the default state by deflateReset().
+
+     deflateSetHeader returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent.
+*/
+
+/*
+ZEXTERN int ZEXPORT inflateInit2 OF((z_streamp strm,
+                                     int  windowBits));
+
+     This is another version of inflateInit with an extra parameter.  The
+   fields next_in, avail_in, zalloc, zfree and opaque must be initialized
+   before by the caller.
+
+     The windowBits parameter is the base two logarithm of the maximum window
+   size (the size of the history buffer).  It should be in the range 8..15 for
+   this version of the library.  The default value is 15 if inflateInit is used
+   instead.  windowBits must be greater than or equal to the windowBits value
+   provided to deflateInit2() while compressing, or it must be equal to 15 if
+   deflateInit2() was not used.  If a compressed stream with a larger window
+   size is given as input, inflate() will return with the error code
+   Z_DATA_ERROR instead of trying to allocate a larger window.
+
+     windowBits can also be zero to request that inflate use the window size in
+   the zlib header of the compressed stream.
+
+     windowBits can also be -8..-15 for raw inflate.  In this case, -windowBits
+   determines the window size.  inflate() will then process raw deflate data,
+   not looking for a zlib or gzip header, not generating a check value, and not
+   looking for any check values for comparison at the end of the stream.  This
+   is for use with other formats that use the deflate compressed data format
+   such as zip.  Those formats provide their own check values.  If a custom
+   format is developed using the raw deflate format for compressed data, it is
+   recommended that a check value such as an adler32 or a crc32 be applied to
+   the uncompressed data as is done in the zlib, gzip, and zip formats.  For
+   most applications, the zlib format should be used as is.  Note that comments
+   above on the use in deflateInit2() applies to the magnitude of windowBits.
+
+     windowBits can also be greater than 15 for optional gzip decoding.  Add
+   32 to windowBits to enable zlib and gzip decoding with automatic header
+   detection, or add 16 to decode only the gzip format (the zlib format will
+   return a Z_DATA_ERROR).  If a gzip stream is being decoded, strm->adler is a
+   crc32 instead of an adler32.
+
+     inflateInit2 returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_VERSION_ERROR if the zlib library version is incompatible with the
+   version assumed by the caller, or Z_STREAM_ERROR if the parameters are
+   invalid, such as a null pointer to the structure.  msg is set to null if
+   there is no error message.  inflateInit2 does not perform any decompression
+   apart from possibly reading the zlib header if present: actual decompression
+   will be done by inflate().  (So next_in and avail_in may be modified, but
+   next_out and avail_out are unused and unchanged.) The current implementation
+   of inflateInit2() does not process any header information -- that is
+   deferred until inflate() is called.
+*/
+
+ZEXTERN int ZEXPORT inflateSetDictionary OF((z_streamp strm,
+                                             const Bytef *dictionary,
+                                             uInt  dictLength));
+/*
+     Initializes the decompression dictionary from the given uncompressed byte
+   sequence.  This function must be called immediately after a call of inflate,
+   if that call returned Z_NEED_DICT.  The dictionary chosen by the compressor
+   can be determined from the adler32 value returned by that call of inflate.
+   The compressor and decompressor must use exactly the same dictionary (see
+   deflateSetDictionary).  For raw inflate, this function can be called at any
+   time to set the dictionary.  If the provided dictionary is smaller than the
+   window and there is already data in the window, then the provided dictionary
+   will amend what's there.  The application must insure that the dictionary
+   that was used for compression is provided.
+
+     inflateSetDictionary returns Z_OK if success, Z_STREAM_ERROR if a
+   parameter is invalid (e.g.  dictionary being Z_NULL) or the stream state is
+   inconsistent, Z_DATA_ERROR if the given dictionary doesn't match the
+   expected one (incorrect adler32 value).  inflateSetDictionary does not
+   perform any decompression: this will be done by subsequent calls of
+   inflate().
+*/
+
+ZEXTERN int ZEXPORT inflateGetDictionary OF((z_streamp strm,
+                                             Bytef *dictionary,
+                                             uInt  *dictLength));
+/*
+     Returns the sliding dictionary being maintained by inflate.  dictLength is
+   set to the number of bytes in the dictionary, and that many bytes are copied
+   to dictionary.  dictionary must have enough space, where 32768 bytes is
+   always enough.  If inflateGetDictionary() is called with dictionary equal to
+   Z_NULL, then only the dictionary length is returned, and nothing is copied.
+   Similary, if dictLength is Z_NULL, then it is not set.
+
+     inflateGetDictionary returns Z_OK on success, or Z_STREAM_ERROR if the
+   stream state is inconsistent.
+*/
+
+ZEXTERN int ZEXPORT inflateSync OF((z_streamp strm));
+/*
+     Skips invalid compressed data until a possible full flush point (see above
+   for the description of deflate with Z_FULL_FLUSH) can be found, or until all
+   available input is skipped.  No output is provided.
+
+     inflateSync searches for a 00 00 FF FF pattern in the compressed data.
+   All full flush points have this pattern, but not all occurrences of this
+   pattern are full flush points.
+
+     inflateSync returns Z_OK if a possible full flush point has been found,
+   Z_BUF_ERROR if no more input was provided, Z_DATA_ERROR if no flush point
+   has been found, or Z_STREAM_ERROR if the stream structure was inconsistent.
+   In the success case, the application may save the current current value of
+   total_in which indicates where valid compressed data was found.  In the
+   error case, the application may repeatedly call inflateSync, providing more
+   input each time, until success or end of the input data.
+*/
+
+ZEXTERN int ZEXPORT inflateCopy OF((z_streamp dest,
+                                    z_streamp source));
+/*
+     Sets the destination stream as a complete copy of the source stream.
+
+     This function can be useful when randomly accessing a large stream.  The
+   first pass through the stream can periodically record the inflate state,
+   allowing restarting inflate at those points when randomly accessing the
+   stream.
+
+     inflateCopy returns Z_OK if success, Z_MEM_ERROR if there was not
+   enough memory, Z_STREAM_ERROR if the source stream state was inconsistent
+   (such as zalloc being Z_NULL).  msg is left unchanged in both source and
+   destination.
+*/
+
+ZEXTERN int ZEXPORT inflateReset OF((z_streamp strm));
+/*
+     This function is equivalent to inflateEnd followed by inflateInit,
+   but does not free and reallocate all the internal decompression state.  The
+   stream will keep attributes that may have been set by inflateInit2.
+
+     inflateReset returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent (such as zalloc or state being Z_NULL).
+*/
+
+ZEXTERN int ZEXPORT inflateReset2 OF((z_streamp strm,
+                                      int windowBits));
+/*
+     This function is the same as inflateReset, but it also permits changing
+   the wrap and window size requests.  The windowBits parameter is interpreted
+   the same as it is for inflateInit2.
+
+     inflateReset2 returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent (such as zalloc or state being Z_NULL), or if
+   the windowBits parameter is invalid.
+*/
+
+ZEXTERN int ZEXPORT inflatePrime OF((z_streamp strm,
+                                     int bits,
+                                     int value));
+/*
+     This function inserts bits in the inflate input stream.  The intent is
+   that this function is used to start inflating at a bit position in the
+   middle of a byte.  The provided bits will be used before any bytes are used
+   from next_in.  This function should only be used with raw inflate, and
+   should be used before the first inflate() call after inflateInit2() or
+   inflateReset().  bits must be less than or equal to 16, and that many of the
+   least significant bits of value will be inserted in the input.
+
+     If bits is negative, then the input stream bit buffer is emptied.  Then
+   inflatePrime() can be called again to put bits in the buffer.  This is used
+   to clear out bits leftover after feeding inflate a block description prior
+   to feeding inflate codes.
+
+     inflatePrime returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent.
+*/
+
+ZEXTERN long ZEXPORT inflateMark OF((z_streamp strm));
+/*
+     This function returns two values, one in the lower 16 bits of the return
+   value, and the other in the remaining upper bits, obtained by shifting the
+   return value down 16 bits.  If the upper value is -1 and the lower value is
+   zero, then inflate() is currently decoding information outside of a block.
+   If the upper value is -1 and the lower value is non-zero, then inflate is in
+   the middle of a stored block, with the lower value equaling the number of
+   bytes from the input remaining to copy.  If the upper value is not -1, then
+   it is the number of bits back from the current bit position in the input of
+   the code (literal or length/distance pair) currently being processed.  In
+   that case the lower value is the number of bytes already emitted for that
+   code.
+
+     A code is being processed if inflate is waiting for more input to complete
+   decoding of the code, or if it has completed decoding but is waiting for
+   more output space to write the literal or match data.
+
+     inflateMark() is used to mark locations in the input data for random
+   access, which may be at bit positions, and to note those cases where the
+   output of a code may span boundaries of random access blocks.  The current
+   location in the input stream can be determined from avail_in and data_type
+   as noted in the description for the Z_BLOCK flush parameter for inflate.
+
+     inflateMark returns the value noted above or -1 << 16 if the provided
+   source stream state was inconsistent.
+*/
+
+ZEXTERN int ZEXPORT inflateGetHeader OF((z_streamp strm,
+                                         gz_headerp head));
+/*
+     inflateGetHeader() requests that gzip header information be stored in the
+   provided gz_header structure.  inflateGetHeader() may be called after
+   inflateInit2() or inflateReset(), and before the first call of inflate().
+   As inflate() processes the gzip stream, head->done is zero until the header
+   is completed, at which time head->done is set to one.  If a zlib stream is
+   being decoded, then head->done is set to -1 to indicate that there will be
+   no gzip header information forthcoming.  Note that Z_BLOCK or Z_TREES can be
+   used to force inflate() to return immediately after header processing is
+   complete and before any actual data is decompressed.
+
+     The text, time, xflags, and os fields are filled in with the gzip header
+   contents.  hcrc is set to true if there is a header CRC.  (The header CRC
+   was valid if done is set to one.) If extra is not Z_NULL, then extra_max
+   contains the maximum number of bytes to write to extra.  Once done is true,
+   extra_len contains the actual extra field length, and extra contains the
+   extra field, or that field truncated if extra_max is less than extra_len.
+   If name is not Z_NULL, then up to name_max characters are written there,
+   terminated with a zero unless the length is greater than name_max.  If
+   comment is not Z_NULL, then up to comm_max characters are written there,
+   terminated with a zero unless the length is greater than comm_max.  When any
+   of extra, name, or comment are not Z_NULL and the respective field is not
+   present in the header, then that field is set to Z_NULL to signal its
+   absence.  This allows the use of deflateSetHeader() with the returned
+   structure to duplicate the header.  However if those fields are set to
+   allocated memory, then the application will need to save those pointers
+   elsewhere so that they can be eventually freed.
+
+     If inflateGetHeader is not used, then the header information is simply
+   discarded.  The header is always checked for validity, including the header
+   CRC if present.  inflateReset() will reset the process to discard the header
+   information.  The application would need to call inflateGetHeader() again to
+   retrieve the header from the next gzip stream.
+
+     inflateGetHeader returns Z_OK if success, or Z_STREAM_ERROR if the source
+   stream state was inconsistent.
+*/
+
+/*
+ZEXTERN int ZEXPORT inflateBackInit OF((z_streamp strm, int windowBits,
+                                        unsigned char FAR *window));
+
+     Initialize the internal stream state for decompression using inflateBack()
+   calls.  The fields zalloc, zfree and opaque in strm must be initialized
+   before the call.  If zalloc and zfree are Z_NULL, then the default library-
+   derived memory allocation routines are used.  windowBits is the base two
+   logarithm of the window size, in the range 8..15.  window is a caller
+   supplied buffer of that size.  Except for special applications where it is
+   assured that deflate was used with small window sizes, windowBits must be 15
+   and a 32K byte window must be supplied to be able to decompress general
+   deflate streams.
+
+     See inflateBack() for the usage of these routines.
+
+     inflateBackInit will return Z_OK on success, Z_STREAM_ERROR if any of
+   the parameters are invalid, Z_MEM_ERROR if the internal state could not be
+   allocated, or Z_VERSION_ERROR if the version of the library does not match
+   the version of the header file.
+*/
+
+typedef unsigned (*in_func) OF((void FAR *,
+                                z_const unsigned char FAR * FAR *));
+typedef int (*out_func) OF((void FAR *, unsigned char FAR *, unsigned));
+
+ZEXTERN int ZEXPORT inflateBack OF((z_streamp strm,
+                                    in_func in, void FAR *in_desc,
+                                    out_func out, void FAR *out_desc));
+/*
+     inflateBack() does a raw inflate with a single call using a call-back
+   interface for input and output.  This is potentially more efficient than
+   inflate() for file i/o applications, in that it avoids copying between the
+   output and the sliding window by simply making the window itself the output
+   buffer.  inflate() can be faster on modern CPUs when used with large
+   buffers.  inflateBack() trusts the application to not change the output
+   buffer passed by the output function, at least until inflateBack() returns.
+
+     inflateBackInit() must be called first to allocate the internal state
+   and to initialize the state with the user-provided window buffer.
+   inflateBack() may then be used multiple times to inflate a complete, raw
+   deflate stream with each call.  inflateBackEnd() is then called to free the
+   allocated state.
+
+     A raw deflate stream is one with no zlib or gzip header or trailer.
+   This routine would normally be used in a utility that reads zip or gzip
+   files and writes out uncompressed files.  The utility would decode the
+   header and process the trailer on its own, hence this routine expects only
+   the raw deflate stream to decompress.  This is different from the normal
+   behavior of inflate(), which expects either a zlib or gzip header and
+   trailer around the deflate stream.
+
+     inflateBack() uses two subroutines supplied by the caller that are then
+   called by inflateBack() for input and output.  inflateBack() calls those
+   routines until it reads a complete deflate stream and writes out all of the
+   uncompressed data, or until it encounters an error.  The function's
+   parameters and return types are defined above in the in_func and out_func
+   typedefs.  inflateBack() will call in(in_desc, &buf) which should return the
+   number of bytes of provided input, and a pointer to that input in buf.  If
+   there is no input available, in() must return zero--buf is ignored in that
+   case--and inflateBack() will return a buffer error.  inflateBack() will call
+   out(out_desc, buf, len) to write the uncompressed data buf[0..len-1].  out()
+   should return zero on success, or non-zero on failure.  If out() returns
+   non-zero, inflateBack() will return with an error.  Neither in() nor out()
+   are permitted to change the contents of the window provided to
+   inflateBackInit(), which is also the buffer that out() uses to write from.
+   The length written by out() will be at most the window size.  Any non-zero
+   amount of input may be provided by in().
+
+     For convenience, inflateBack() can be provided input on the first call by
+   setting strm->next_in and strm->avail_in.  If that input is exhausted, then
+   in() will be called.  Therefore strm->next_in must be initialized before
+   calling inflateBack().  If strm->next_in is Z_NULL, then in() will be called
+   immediately for input.  If strm->next_in is not Z_NULL, then strm->avail_in
+   must also be initialized, and then if strm->avail_in is not zero, input will
+   initially be taken from strm->next_in[0 ..  strm->avail_in - 1].
+
+     The in_desc and out_desc parameters of inflateBack() is passed as the
+   first parameter of in() and out() respectively when they are called.  These
+   descriptors can be optionally used to pass any information that the caller-
+   supplied in() and out() functions need to do their job.
+
+     On return, inflateBack() will set strm->next_in and strm->avail_in to
+   pass back any unused input that was provided by the last in() call.  The
+   return values of inflateBack() can be Z_STREAM_END on success, Z_BUF_ERROR
+   if in() or out() returned an error, Z_DATA_ERROR if there was a format error
+   in the deflate stream (in which case strm->msg is set to indicate the nature
+   of the error), or Z_STREAM_ERROR if the stream was not properly initialized.
+   In the case of Z_BUF_ERROR, an input or output error can be distinguished
+   using strm->next_in which will be Z_NULL only if in() returned an error.  If
+   strm->next_in is not Z_NULL, then the Z_BUF_ERROR was due to out() returning
+   non-zero.  (in() will always be called before out(), so strm->next_in is
+   assured to be defined if out() returns non-zero.) Note that inflateBack()
+   cannot return Z_OK.
+*/
+
+ZEXTERN int ZEXPORT inflateBackEnd OF((z_streamp strm));
+/*
+     All memory allocated by inflateBackInit() is freed.
+
+     inflateBackEnd() returns Z_OK on success, or Z_STREAM_ERROR if the stream
+   state was inconsistent.
+*/
+
+ZEXTERN uLong ZEXPORT zlibCompileFlags OF((void));
+/* Return flags indicating compile-time options.
+
+    Type sizes, two bits each, 00 = 16 bits, 01 = 32, 10 = 64, 11 = other:
+     1.0: size of uInt
+     3.2: size of uLong
+     5.4: size of voidpf (pointer)
+     7.6: size of z_off_t
+
+    Compiler, assembler, and debug options:
+     8: DEBUG
+     9: ASMV or ASMINF -- use ASM code
+     10: ZLIB_WINAPI -- exported functions use the WINAPI calling convention
+     11: 0 (reserved)
+
+    One-time table building (smaller code, but not thread-safe if true):
+     12: BUILDFIXED -- build static block decoding tables when needed
+     13: DYNAMIC_CRC_TABLE -- build CRC calculation tables when needed
+     14,15: 0 (reserved)
+
+    Library content (indicates missing functionality):
+     16: NO_GZCOMPRESS -- gz* functions cannot compress (to avoid linking
+                          deflate code when not needed)
+     17: NO_GZIP -- deflate can't write gzip streams, and inflate can't detect
+                    and decode gzip streams (to avoid linking crc code)
+     18-19: 0 (reserved)
+
+    Operation variations (changes in library functionality):
+     20: PKZIP_BUG_WORKAROUND -- slightly more permissive inflate
+     21: FASTEST -- deflate algorithm with only one, lowest compression level
+     22,23: 0 (reserved)
+
+    The sprintf variant used by gzprintf (zero is best):
+     24: 0 = vs*, 1 = s* -- 1 means limited to 20 arguments after the format
+     25: 0 = *nprintf, 1 = *printf -- 1 means gzprintf() not secure!
+     26: 0 = returns value, 1 = void -- 1 means inferred string length returned
+
+    Remainder:
+     27-31: 0 (reserved)
+ */
+
+#ifndef Z_SOLO
+
+                        /* utility functions */
+
+/*
+     The following utility functions are implemented on top of the basic
+   stream-oriented functions.  To simplify the interface, some default options
+   are assumed (compression level and memory usage, standard memory allocation
+   functions).  The source code of these utility functions can be modified if
+   you need special options.
+*/
+
+ZEXTERN int ZEXPORT compress OF((Bytef *dest,   uLongf *destLen,
+                                 const Bytef *source, uLong sourceLen));
+/*
+     Compresses the source buffer into the destination buffer.  sourceLen is
+   the byte length of the source buffer.  Upon entry, destLen is the total size
+   of the destination buffer, which must be at least the value returned by
+   compressBound(sourceLen).  Upon exit, destLen is the actual size of the
+   compressed buffer.
+
+     compress returns Z_OK if success, Z_MEM_ERROR if there was not
+   enough memory, Z_BUF_ERROR if there was not enough room in the output
+   buffer.
+*/
+
+ZEXTERN int ZEXPORT compress2 OF((Bytef *dest,   uLongf *destLen,
+                                  const Bytef *source, uLong sourceLen,
+                                  int level));
+/*
+     Compresses the source buffer into the destination buffer.  The level
+   parameter has the same meaning as in deflateInit.  sourceLen is the byte
+   length of the source buffer.  Upon entry, destLen is the total size of the
+   destination buffer, which must be at least the value returned by
+   compressBound(sourceLen).  Upon exit, destLen is the actual size of the
+   compressed buffer.
+
+     compress2 returns Z_OK if success, Z_MEM_ERROR if there was not enough
+   memory, Z_BUF_ERROR if there was not enough room in the output buffer,
+   Z_STREAM_ERROR if the level parameter is invalid.
+*/
+
+ZEXTERN uLong ZEXPORT compressBound OF((uLong sourceLen));
+/*
+     compressBound() returns an upper bound on the compressed size after
+   compress() or compress2() on sourceLen bytes.  It would be used before a
+   compress() or compress2() call to allocate the destination buffer.
+*/
+
+ZEXTERN int ZEXPORT uncompress OF((Bytef *dest,   uLongf *destLen,
+                                   const Bytef *source, uLong sourceLen));
+/*
+     Decompresses the source buffer into the destination buffer.  sourceLen is
+   the byte length of the source buffer.  Upon entry, destLen is the total size
+   of the destination buffer, which must be large enough to hold the entire
+   uncompressed data.  (The size of the uncompressed data must have been saved
+   previously by the compressor and transmitted to the decompressor by some
+   mechanism outside the scope of this compression library.) Upon exit, destLen
+   is the actual size of the uncompressed buffer.
+
+     uncompress returns Z_OK if success, Z_MEM_ERROR if there was not
+   enough memory, Z_BUF_ERROR if there was not enough room in the output
+   buffer, or Z_DATA_ERROR if the input data was corrupted or incomplete.  In
+   the case where there is not enough room, uncompress() will fill the output
+   buffer with the uncompressed data up to that point.
+*/
+
+                        /* gzip file access functions */
+
+/*
+     This library supports reading and writing files in gzip (.gz) format with
+   an interface similar to that of stdio, using the functions that start with
+   "gz".  The gzip format is different from the zlib format.  gzip is a gzip
+   wrapper, documented in RFC 1952, wrapped around a deflate stream.
+*/
+
+typedef struct gzFile_s *gzFile;    /* semi-opaque gzip file descriptor */
+
+/*
+ZEXTERN gzFile ZEXPORT gzopen OF((const char *path, const char *mode));
+
+     Opens a gzip (.gz) file for reading or writing.  The mode parameter is as
+   in fopen ("rb" or "wb") but can also include a compression level ("wb9") or
+   a strategy: 'f' for filtered data as in "wb6f", 'h' for Huffman-only
+   compression as in "wb1h", 'R' for run-length encoding as in "wb1R", or 'F'
+   for fixed code compression as in "wb9F".  (See the description of
+   deflateInit2 for more information about the strategy parameter.)  'T' will
+   request transparent writing or appending with no compression and not using
+   the gzip format.
+
+     "a" can be used instead of "w" to request that the gzip stream that will
+   be written be appended to the file.  "+" will result in an error, since
+   reading and writing to the same gzip file is not supported.  The addition of
+   "x" when writing will create the file exclusively, which fails if the file
+   already exists.  On systems that support it, the addition of "e" when
+   reading or writing will set the flag to close the file on an execve() call.
+
+     These functions, as well as gzip, will read and decode a sequence of gzip
+   streams in a file.  The append function of gzopen() can be used to create
+   such a file.  (Also see gzflush() for another way to do this.)  When
+   appending, gzopen does not test whether the file begins with a gzip stream,
+   nor does it look for the end of the gzip streams to begin appending.  gzopen
+   will simply append a gzip stream to the existing file.
+
+     gzopen can be used to read a file which is not in gzip format; in this
+   case gzread will directly read from the file without decompression.  When
+   reading, this will be detected automatically by looking for the magic two-
+   byte gzip header.
+
+     gzopen returns NULL if the file could not be opened, if there was
+   insufficient memory to allocate the gzFile state, or if an invalid mode was
+   specified (an 'r', 'w', or 'a' was not provided, or '+' was provided).
+   errno can be checked to determine if the reason gzopen failed was that the
+   file could not be opened.
+*/
+
+ZEXTERN gzFile ZEXPORT gzdopen OF((int fd, const char *mode));
+/*
+     gzdopen associates a gzFile with the file descriptor fd.  File descriptors
+   are obtained from calls like open, dup, creat, pipe or fileno (if the file
+   has been previously opened with fopen).  The mode parameter is as in gzopen.
+
+     The next call of gzclose on the returned gzFile will also close the file
+   descriptor fd, just like fclose(fdopen(fd, mode)) closes the file descriptor
+   fd.  If you want to keep fd open, use fd = dup(fd_keep); gz = gzdopen(fd,
+   mode);.  The duplicated descriptor should be saved to avoid a leak, since
+   gzdopen does not close fd if it fails.  If you are using fileno() to get the
+   file descriptor from a FILE *, then you will have to use dup() to avoid
+   double-close()ing the file descriptor.  Both gzclose() and fclose() will
+   close the associated file descriptor, so they need to have different file
+   descriptors.
+
+     gzdopen returns NULL if there was insufficient memory to allocate the
+   gzFile state, if an invalid mode was specified (an 'r', 'w', or 'a' was not
+   provided, or '+' was provided), or if fd is -1.  The file descriptor is not
+   used until the next gz* read, write, seek, or close operation, so gzdopen
+   will not detect if fd is invalid (unless fd is -1).
+*/
+
+ZEXTERN int ZEXPORT gzbuffer OF((gzFile file, unsigned size));
+/*
+     Set the internal buffer size used by this library's functions.  The
+   default buffer size is 8192 bytes.  This function must be called after
+   gzopen() or gzdopen(), and before any other calls that read or write the
+   file.  The buffer memory allocation is always deferred to the first read or
+   write.  Two buffers are allocated, either both of the specified size when
+   writing, or one of the specified size and the other twice that size when
+   reading.  A larger buffer size of, for example, 64K or 128K bytes will
+   noticeably increase the speed of decompression (reading).
+
+     The new buffer size also affects the maximum length for gzprintf().
+
+     gzbuffer() returns 0 on success, or -1 on failure, such as being called
+   too late.
+*/
+
+ZEXTERN int ZEXPORT gzsetparams OF((gzFile file, int level, int strategy));
+/*
+     Dynamically update the compression level or strategy.  See the description
+   of deflateInit2 for the meaning of these parameters.
+
+     gzsetparams returns Z_OK if success, or Z_STREAM_ERROR if the file was not
+   opened for writing.
+*/
+
+ZEXTERN int ZEXPORT gzread OF((gzFile file, voidp buf, unsigned len));
+/*
+     Reads the given number of uncompressed bytes from the compressed file.  If
+   the input file is not in gzip format, gzread copies the given number of
+   bytes into the buffer directly from the file.
+
+     After reaching the end of a gzip stream in the input, gzread will continue
+   to read, looking for another gzip stream.  Any number of gzip streams may be
+   concatenated in the input file, and will all be decompressed by gzread().
+   If something other than a gzip stream is encountered after a gzip stream,
+   that remaining trailing garbage is ignored (and no error is returned).
+
+     gzread can be used to read a gzip file that is being concurrently written.
+   Upon reaching the end of the input, gzread will return with the available
+   data.  If the error code returned by gzerror is Z_OK or Z_BUF_ERROR, then
+   gzclearerr can be used to clear the end of file indicator in order to permit
+   gzread to be tried again.  Z_OK indicates that a gzip stream was completed
+   on the last gzread.  Z_BUF_ERROR indicates that the input file ended in the
+   middle of a gzip stream.  Note that gzread does not return -1 in the event
+   of an incomplete gzip stream.  This error is deferred until gzclose(), which
+   will return Z_BUF_ERROR if the last gzread ended in the middle of a gzip
+   stream.  Alternatively, gzerror can be used before gzclose to detect this
+   case.
+
+     gzread returns the number of uncompressed bytes actually read, less than
+   len for end of file, or -1 for error.
+*/
+
+ZEXTERN int ZEXPORT gzwrite OF((gzFile file,
+                                voidpc buf, unsigned len));
+/*
+     Writes the given number of uncompressed bytes into the compressed file.
+   gzwrite returns the number of uncompressed bytes written or 0 in case of
+   error.
+*/
+
+ZEXTERN int ZEXPORTVA gzprintf Z_ARG((gzFile file, const char *format, ...));
+/*
+     Converts, formats, and writes the arguments to the compressed file under
+   control of the format string, as in fprintf.  gzprintf returns the number of
+   uncompressed bytes actually written, or 0 in case of error.  The number of
+   uncompressed bytes written is limited to 8191, or one less than the buffer
+   size given to gzbuffer().  The caller should assure that this limit is not
+   exceeded.  If it is exceeded, then gzprintf() will return an error (0) with
+   nothing written.  In this case, there may also be a buffer overflow with
+   unpredictable consequences, which is possible only if zlib was compiled with
+   the insecure functions sprintf() or vsprintf() because the secure snprintf()
+   or vsnprintf() functions were not available.  This can be determined using
+   zlibCompileFlags().
+*/
+
+ZEXTERN int ZEXPORT gzputs OF((gzFile file, const char *s));
+/*
+     Writes the given null-terminated string to the compressed file, excluding
+   the terminating null character.
+
+     gzputs returns the number of characters written, or -1 in case of error.
+*/
+
+ZEXTERN char * ZEXPORT gzgets OF((gzFile file, char *buf, int len));
+/*
+     Reads bytes from the compressed file until len-1 characters are read, or a
+   newline character is read and transferred to buf, or an end-of-file
+   condition is encountered.  If any characters are read or if len == 1, the
+   string is terminated with a null character.  If no characters are read due
+   to an end-of-file or len < 1, then the buffer is left untouched.
+
+     gzgets returns buf which is a null-terminated string, or it returns NULL
+   for end-of-file or in case of error.  If there was an error, the contents at
+   buf are indeterminate.
+*/
+
+ZEXTERN int ZEXPORT gzputc OF((gzFile file, int c));
+/*
+     Writes c, converted to an unsigned char, into the compressed file.  gzputc
+   returns the value that was written, or -1 in case of error.
+*/
+
+ZEXTERN int ZEXPORT gzgetc OF((gzFile file));
+/*
+     Reads one byte from the compressed file.  gzgetc returns this byte or -1
+   in case of end of file or error.  This is implemented as a macro for speed.
+   As such, it does not do all of the checking the other functions do.  I.e.
+   it does not check to see if file is NULL, nor whether the structure file
+   points to has been clobbered or not.
+*/
+
+ZEXTERN int ZEXPORT gzungetc OF((int c, gzFile file));
+/*
+     Push one character back onto the stream to be read as the first character
+   on the next read.  At least one character of push-back is allowed.
+   gzungetc() returns the character pushed, or -1 on failure.  gzungetc() will
+   fail if c is -1, and may fail if a character has been pushed but not read
+   yet.  If gzungetc is used immediately after gzopen or gzdopen, at least the
+   output buffer size of pushed characters is allowed.  (See gzbuffer above.)
+   The pushed character will be discarded if the stream is repositioned with
+   gzseek() or gzrewind().
+*/
+
+ZEXTERN int ZEXPORT gzflush OF((gzFile file, int flush));
+/*
+     Flushes all pending output into the compressed file.  The parameter flush
+   is as in the deflate() function.  The return value is the zlib error number
+   (see function gzerror below).  gzflush is only permitted when writing.
+
+     If the flush parameter is Z_FINISH, the remaining data is written and the
+   gzip stream is completed in the output.  If gzwrite() is called again, a new
+   gzip stream will be started in the output.  gzread() is able to read such
+   concatented gzip streams.
+
+     gzflush should be called only when strictly necessary because it will
+   degrade compression if called too often.
+*/
+
+/*
+ZEXTERN z_off_t ZEXPORT gzseek OF((gzFile file,
+                                   z_off_t offset, int whence));
+
+     Sets the starting position for the next gzread or gzwrite on the given
+   compressed file.  The offset represents a number of bytes in the
+   uncompressed data stream.  The whence parameter is defined as in lseek(2);
+   the value SEEK_END is not supported.
+
+     If the file is opened for reading, this function is emulated but can be
+   extremely slow.  If the file is opened for writing, only forward seeks are
+   supported; gzseek then compresses a sequence of zeroes up to the new
+   starting position.
+
+     gzseek returns the resulting offset location as measured in bytes from
+   the beginning of the uncompressed stream, or -1 in case of error, in
+   particular if the file is opened for writing and the new starting position
+   would be before the current position.
+*/
+
+ZEXTERN int ZEXPORT    gzrewind OF((gzFile file));
+/*
+     Rewinds the given file. This function is supported only for reading.
+
+     gzrewind(file) is equivalent to (int)gzseek(file, 0L, SEEK_SET)
+*/
+
+/*
+ZEXTERN z_off_t ZEXPORT    gztell OF((gzFile file));
+
+     Returns the starting position for the next gzread or gzwrite on the given
+   compressed file.  This position represents a number of bytes in the
+   uncompressed data stream, and is zero when starting, even if appending or
+   reading a gzip stream from the middle of a file using gzdopen().
+
+     gztell(file) is equivalent to gzseek(file, 0L, SEEK_CUR)
+*/
+
+/*
+ZEXTERN z_off_t ZEXPORT gzoffset OF((gzFile file));
+
+     Returns the current offset in the file being read or written.  This offset
+   includes the count of bytes that precede the gzip stream, for example when
+   appending or when using gzdopen() for reading.  When reading, the offset
+   does not include as yet unused buffered input.  This information can be used
+   for a progress indicator.  On error, gzoffset() returns -1.
+*/
+
+ZEXTERN int ZEXPORT gzeof OF((gzFile file));
+/*
+     Returns true (1) if the end-of-file indicator has been set while reading,
+   false (0) otherwise.  Note that the end-of-file indicator is set only if the
+   read tried to go past the end of the input, but came up short.  Therefore,
+   just like feof(), gzeof() may return false even if there is no more data to
+   read, in the event that the last read request was for the exact number of
+   bytes remaining in the input file.  This will happen if the input file size
+   is an exact multiple of the buffer size.
+
+     If gzeof() returns true, then the read functions will return no more data,
+   unless the end-of-file indicator is reset by gzclearerr() and the input file
+   has grown since the previous end of file was detected.
+*/
+
+ZEXTERN int ZEXPORT gzdirect OF((gzFile file));
+/*
+     Returns true (1) if file is being copied directly while reading, or false
+   (0) if file is a gzip stream being decompressed.
+
+     If the input file is empty, gzdirect() will return true, since the input
+   does not contain a gzip stream.
+
+     If gzdirect() is used immediately after gzopen() or gzdopen() it will
+   cause buffers to be allocated to allow reading the file to determine if it
+   is a gzip file.  Therefore if gzbuffer() is used, it should be called before
+   gzdirect().
+
+     When writing, gzdirect() returns true (1) if transparent writing was
+   requested ("wT" for the gzopen() mode), or false (0) otherwise.  (Note:
+   gzdirect() is not needed when writing.  Transparent writing must be
+   explicitly requested, so the application already knows the answer.  When
+   linking statically, using gzdirect() will include all of the zlib code for
+   gzip file reading and decompression, which may not be desired.)
+*/
+
+ZEXTERN int ZEXPORT    gzclose OF((gzFile file));
+/*
+     Flushes all pending output if necessary, closes the compressed file and
+   deallocates the (de)compression state.  Note that once file is closed, you
+   cannot call gzerror with file, since its structures have been deallocated.
+   gzclose must not be called more than once on the same file, just as free
+   must not be called more than once on the same allocation.
+
+     gzclose will return Z_STREAM_ERROR if file is not valid, Z_ERRNO on a
+   file operation error, Z_MEM_ERROR if out of memory, Z_BUF_ERROR if the
+   last read ended in the middle of a gzip stream, or Z_OK on success.
+*/
+
+ZEXTERN int ZEXPORT gzclose_r OF((gzFile file));
+ZEXTERN int ZEXPORT gzclose_w OF((gzFile file));
+/*
+     Same as gzclose(), but gzclose_r() is only for use when reading, and
+   gzclose_w() is only for use when writing or appending.  The advantage to
+   using these instead of gzclose() is that they avoid linking in zlib
+   compression or decompression code that is not used when only reading or only
+   writing respectively.  If gzclose() is used, then both compression and
+   decompression code will be included the application when linking to a static
+   zlib library.
+*/
+
+ZEXTERN const char * ZEXPORT gzerror OF((gzFile file, int *errnum));
+/*
+     Returns the error message for the last error which occurred on the given
+   compressed file.  errnum is set to zlib error number.  If an error occurred
+   in the file system and not in the compression library, errnum is set to
+   Z_ERRNO and the application may consult errno to get the exact error code.
+
+     The application must not modify the returned string.  Future calls to
+   this function may invalidate the previously returned string.  If file is
+   closed, then the string previously returned by gzerror will no longer be
+   available.
+
+     gzerror() should be used to distinguish errors from end-of-file for those
+   functions above that do not distinguish those cases in their return values.
+*/
+
+ZEXTERN void ZEXPORT gzclearerr OF((gzFile file));
+/*
+     Clears the error and end-of-file flags for file.  This is analogous to the
+   clearerr() function in stdio.  This is useful for continuing to read a gzip
+   file that is being written concurrently.
+*/
+
+#endif /* !Z_SOLO */
+
+                        /* checksum functions */
+
+/*
+     These functions are not related to compression but are exported
+   anyway because they might be useful in applications using the compression
+   library.
+*/
+
+ZEXTERN uLong ZEXPORT adler32 OF((uLong adler, const Bytef *buf, uInt len));
+/*
+     Update a running Adler-32 checksum with the bytes buf[0..len-1] and
+   return the updated checksum.  If buf is Z_NULL, this function returns the
+   required initial value for the checksum.
+
+     An Adler-32 checksum is almost as reliable as a CRC32 but can be computed
+   much faster.
+
+   Usage example:
+
+     uLong adler = adler32(0L, Z_NULL, 0);
+
+     while (read_buffer(buffer, length) != EOF) {
+       adler = adler32(adler, buffer, length);
+     }
+     if (adler != original_adler) error();
+*/
+
+/*
+ZEXTERN uLong ZEXPORT adler32_combine OF((uLong adler1, uLong adler2,
+                                          z_off_t len2));
+
+     Combine two Adler-32 checksums into one.  For two sequences of bytes, seq1
+   and seq2 with lengths len1 and len2, Adler-32 checksums were calculated for
+   each, adler1 and adler2.  adler32_combine() returns the Adler-32 checksum of
+   seq1 and seq2 concatenated, requiring only adler1, adler2, and len2.  Note
+   that the z_off_t type (like off_t) is a signed integer.  If len2 is
+   negative, the result has no meaning or utility.
+*/
+
+ZEXTERN uLong ZEXPORT crc32   OF((uLong crc, const Bytef *buf, uInt len));
+/*
+     Update a running CRC-32 with the bytes buf[0..len-1] and return the
+   updated CRC-32.  If buf is Z_NULL, this function returns the required
+   initial value for the crc.  Pre- and post-conditioning (one's complement) is
+   performed within this function so it shouldn't be done by the application.
+
+   Usage example:
+
+     uLong crc = crc32(0L, Z_NULL, 0);
+
+     while (read_buffer(buffer, length) != EOF) {
+       crc = crc32(crc, buffer, length);
+     }
+     if (crc != original_crc) error();
+*/
+
+/*
+ZEXTERN uLong ZEXPORT crc32_combine OF((uLong crc1, uLong crc2, z_off_t len2));
+
+     Combine two CRC-32 check values into one.  For two sequences of bytes,
+   seq1 and seq2 with lengths len1 and len2, CRC-32 check values were
+   calculated for each, crc1 and crc2.  crc32_combine() returns the CRC-32
+   check value of seq1 and seq2 concatenated, requiring only crc1, crc2, and
+   len2.
+*/
+
+
+                        /* various hacks, don't look :) */
+
+/* deflateInit and inflateInit are macros to allow checking the zlib version
+ * and the compiler's view of z_stream:
+ */
+ZEXTERN int ZEXPORT deflateInit_ OF((z_streamp strm, int level,
+                                     const char *version, int stream_size));
+ZEXTERN int ZEXPORT inflateInit_ OF((z_streamp strm,
+                                     const char *version, int stream_size));
+ZEXTERN int ZEXPORT deflateInit2_ OF((z_streamp strm, int  level, int  method,
+                                      int windowBits, int memLevel,
+                                      int strategy, const char *version,
+                                      int stream_size));
+ZEXTERN int ZEXPORT inflateInit2_ OF((z_streamp strm, int  windowBits,
+                                      const char *version, int stream_size));
+ZEXTERN int ZEXPORT inflateBackInit_ OF((z_streamp strm, int windowBits,
+                                         unsigned char FAR *window,
+                                         const char *version,
+                                         int stream_size));
+#define deflateInit(strm, level) \
+        deflateInit_((strm), (level), ZLIB_VERSION, (int)sizeof(z_stream))
+#define inflateInit(strm) \
+        inflateInit_((strm), ZLIB_VERSION, (int)sizeof(z_stream))
+#define deflateInit2(strm, level, method, windowBits, memLevel, strategy) \
+        deflateInit2_((strm),(level),(method),(windowBits),(memLevel),\
+                      (strategy), ZLIB_VERSION, (int)sizeof(z_stream))
+#define inflateInit2(strm, windowBits) \
+        inflateInit2_((strm), (windowBits), ZLIB_VERSION, \
+                      (int)sizeof(z_stream))
+#define inflateBackInit(strm, windowBits, window) \
+        inflateBackInit_((strm), (windowBits), (window), \
+                      ZLIB_VERSION, (int)sizeof(z_stream))
+
+#ifndef Z_SOLO
+
+/* gzgetc() macro and its supporting function and exposed data structure.  Note
+ * that the real internal state is much larger than the exposed structure.
+ * This abbreviated structure exposes just enough for the gzgetc() macro.  The
+ * user should not mess with these exposed elements, since their names or
+ * behavior could change in the future, perhaps even capriciously.  They can
+ * only be used by the gzgetc() macro.  You have been warned.
+ */
+struct gzFile_s {
+    unsigned have;
+    unsigned char *next;
+    z_off64_t pos;
+};
+ZEXTERN int ZEXPORT gzgetc_ OF((gzFile file));  /* backward compatibility */
+#ifdef Z_PREFIX_SET
+#  undef z_gzgetc
+#  define z_gzgetc(g) \
+          ((g)->have ? ((g)->have--, (g)->pos++, *((g)->next)++) : gzgetc(g))
+#else
+#  define gzgetc(g) \
+          ((g)->have ? ((g)->have--, (g)->pos++, *((g)->next)++) : gzgetc(g))
+#endif
+
+/* provide 64-bit offset functions if _LARGEFILE64_SOURCE defined, and/or
+ * change the regular functions to 64 bits if _FILE_OFFSET_BITS is 64 (if
+ * both are true, the application gets the *64 functions, and the regular
+ * functions are changed to 64 bits) -- in case these are set on systems
+ * without large file support, _LFS64_LARGEFILE must also be true
+ */
+#ifdef Z_LARGE64
+   ZEXTERN gzFile ZEXPORT gzopen64 OF((const char *, const char *));
+   ZEXTERN z_off64_t ZEXPORT gzseek64 OF((gzFile, z_off64_t, int));
+   ZEXTERN z_off64_t ZEXPORT gztell64 OF((gzFile));
+   ZEXTERN z_off64_t ZEXPORT gzoffset64 OF((gzFile));
+   ZEXTERN uLong ZEXPORT adler32_combine64 OF((uLong, uLong, z_off64_t));
+   ZEXTERN uLong ZEXPORT crc32_combine64 OF((uLong, uLong, z_off64_t));
+#endif
+
+#if !defined(ZLIB_INTERNAL) && defined(Z_WANT64)
+#  ifdef Z_PREFIX_SET
+#    define z_gzopen z_gzopen64
+#    define z_gzseek z_gzseek64
+#    define z_gztell z_gztell64
+#    define z_gzoffset z_gzoffset64
+#    define z_adler32_combine z_adler32_combine64
+#    define z_crc32_combine z_crc32_combine64
+#  else
+#    define gzopen gzopen64
+#    define gzseek gzseek64
+#    define gztell gztell64
+#    define gzoffset gzoffset64
+#    define adler32_combine adler32_combine64
+#    define crc32_combine crc32_combine64
+#  endif
+#  ifndef Z_LARGE64
+     ZEXTERN gzFile ZEXPORT gzopen64 OF((const char *, const char *));
+     ZEXTERN z_off_t ZEXPORT gzseek64 OF((gzFile, z_off_t, int));
+     ZEXTERN z_off_t ZEXPORT gztell64 OF((gzFile));
+     ZEXTERN z_off_t ZEXPORT gzoffset64 OF((gzFile));
+     ZEXTERN uLong ZEXPORT adler32_combine64 OF((uLong, uLong, z_off_t));
+     ZEXTERN uLong ZEXPORT crc32_combine64 OF((uLong, uLong, z_off_t));
+#  endif
+#else
+   ZEXTERN gzFile ZEXPORT gzopen OF((const char *, const char *));
+   ZEXTERN z_off_t ZEXPORT gzseek OF((gzFile, z_off_t, int));
+   ZEXTERN z_off_t ZEXPORT gztell OF((gzFile));
+   ZEXTERN z_off_t ZEXPORT gzoffset OF((gzFile));
+   ZEXTERN uLong ZEXPORT adler32_combine OF((uLong, uLong, z_off_t));
+   ZEXTERN uLong ZEXPORT crc32_combine OF((uLong, uLong, z_off_t));
+#endif
+
+#else /* Z_SOLO */
+
+   ZEXTERN uLong ZEXPORT adler32_combine OF((uLong, uLong, z_off_t));
+   ZEXTERN uLong ZEXPORT crc32_combine OF((uLong, uLong, z_off_t));
+
+#endif /* !Z_SOLO */
+
+/* hack for buggy compilers */
+#if !defined(ZUTIL_H) && !defined(NO_DUMMY_DECL)
+    struct internal_state {int dummy;};
+#endif
+
+/* undocumented functions */
+ZEXTERN const char   * ZEXPORT zError           OF((int));
+ZEXTERN int            ZEXPORT inflateSyncPoint OF((z_streamp));
+ZEXTERN const z_crc_t FAR * ZEXPORT get_crc_table    OF((void));
+ZEXTERN int            ZEXPORT inflateUndermine OF((z_streamp, int));
+ZEXTERN int            ZEXPORT inflateResetKeep OF((z_streamp));
+ZEXTERN int            ZEXPORT deflateResetKeep OF((z_streamp));
+#if defined(_WIN32) && !defined(Z_SOLO)
+ZEXTERN gzFile         ZEXPORT gzopen_w OF((const wchar_t *path,
+                                            const char *mode));
+#endif
+#if defined(STDC) || defined(Z_HAVE_STDARG_H)
+#  ifndef Z_SOLO
+ZEXTERN int            ZEXPORTVA gzvprintf Z_ARG((gzFile file,
+                                                  const char *format,
+                                                  va_list va));
+#  endif
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* ZLIB_H */
diff --git a/ml/dlib/dlib/external/zlib/zutil.c b/ml/dlib/dlib/external/zlib/zutil.c
new file mode 100644
index 000000000..23d2ebef0
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/zutil.c
@@ -0,0 +1,324 @@
+/* zutil.c -- target dependent utility functions for the compression library
+ * Copyright (C) 1995-2005, 2010, 2011, 2012 Jean-loup Gailly.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* @(#) $Id$ */
+
+#include "zutil.h"
+#ifndef Z_SOLO
+#  include "gzguts.h"
+#endif
+
+#ifndef NO_DUMMY_DECL
+struct internal_state      {int dummy;}; /* for buggy compilers */
+#endif
+
+z_const char * const z_errmsg[10] = {
+"need dictionary",     /* Z_NEED_DICT       2  */
+"stream end",          /* Z_STREAM_END      1  */
+"",                    /* Z_OK              0  */
+"file error",          /* Z_ERRNO         (-1) */
+"stream error",        /* Z_STREAM_ERROR  (-2) */
+"data error",          /* Z_DATA_ERROR    (-3) */
+"insufficient memory", /* Z_MEM_ERROR     (-4) */
+"buffer error",        /* Z_BUF_ERROR     (-5) */
+"incompatible version",/* Z_VERSION_ERROR (-6) */
+""};
+
+
+const char * ZEXPORT zlibVersion()
+{
+    return ZLIB_VERSION;
+}
+
+uLong ZEXPORT zlibCompileFlags()
+{
+    uLong flags;
+
+    flags = 0;
+    switch ((int)(sizeof(uInt))) {
+    case 2:     break;
+    case 4:     flags += 1;     break;
+    case 8:     flags += 2;     break;
+    default:    flags += 3;
+    }
+    switch ((int)(sizeof(uLong))) {
+    case 2:     break;
+    case 4:     flags += 1 << 2;        break;
+    case 8:     flags += 2 << 2;        break;
+    default:    flags += 3 << 2;
+    }
+    switch ((int)(sizeof(voidpf))) {
+    case 2:     break;
+    case 4:     flags += 1 << 4;        break;
+    case 8:     flags += 2 << 4;        break;
+    default:    flags += 3 << 4;
+    }
+    switch ((int)(sizeof(z_off_t))) {
+    case 2:     break;
+    case 4:     flags += 1 << 6;        break;
+    case 8:     flags += 2 << 6;        break;
+    default:    flags += 3 << 6;
+    }
+#ifdef DEBUG
+    flags += 1 << 8;
+#endif
+#if defined(ASMV) || defined(ASMINF)
+    flags += 1 << 9;
+#endif
+#ifdef ZLIB_WINAPI
+    flags += 1 << 10;
+#endif
+#ifdef BUILDFIXED
+    flags += 1 << 12;
+#endif
+#ifdef DYNAMIC_CRC_TABLE
+    flags += 1 << 13;
+#endif
+#ifdef NO_GZCOMPRESS
+    flags += 1L << 16;
+#endif
+#ifdef NO_GZIP
+    flags += 1L << 17;
+#endif
+#ifdef PKZIP_BUG_WORKAROUND
+    flags += 1L << 20;
+#endif
+#ifdef FASTEST
+    flags += 1L << 21;
+#endif
+#if defined(STDC) || defined(Z_HAVE_STDARG_H)
+#  ifdef NO_vsnprintf
+    flags += 1L << 25;
+#    ifdef HAS_vsprintf_void
+    flags += 1L << 26;
+#    endif
+#  else
+#    ifdef HAS_vsnprintf_void
+    flags += 1L << 26;
+#    endif
+#  endif
+#else
+    flags += 1L << 24;
+#  ifdef NO_snprintf
+    flags += 1L << 25;
+#    ifdef HAS_sprintf_void
+    flags += 1L << 26;
+#    endif
+#  else
+#    ifdef HAS_snprintf_void
+    flags += 1L << 26;
+#    endif
+#  endif
+#endif
+    return flags;
+}
+
+#ifdef DEBUG
+
+#  ifndef verbose
+#    define verbose 0
+#  endif
+int ZLIB_INTERNAL z_verbose = verbose;
+
+void ZLIB_INTERNAL z_error (m)
+    char *m;
+{
+    fprintf(stderr, "%s\n", m);
+    exit(1);
+}
+#endif
+
+/* exported to allow conversion of error code to string for compress() and
+ * uncompress()
+ */
+const char * ZEXPORT zError(err)
+    int err;
+{
+    return ERR_MSG(err);
+}
+
+#if defined(_WIN32_WCE)
+    /* The Microsoft C Run-Time Library for Windows CE doesn't have
+     * errno.  We define it as a global variable to simplify porting.
+     * Its value is always 0 and should not be used.
+     */
+    int errno = 0;
+#endif
+
+#ifndef HAVE_MEMCPY
+
+void ZLIB_INTERNAL zmemcpy(dest, source, len)
+    Bytef* dest;
+    const Bytef* source;
+    uInt  len;
+{
+    if (len == 0) return;
+    do {
+        *dest++ = *source++; /* ??? to be unrolled */
+    } while (--len != 0);
+}
+
+int ZLIB_INTERNAL zmemcmp(s1, s2, len)
+    const Bytef* s1;
+    const Bytef* s2;
+    uInt  len;
+{
+    uInt j;
+
+    for (j = 0; j < len; j++) {
+        if (s1[j] != s2[j]) return 2*(s1[j] > s2[j])-1;
+    }
+    return 0;
+}
+
+void ZLIB_INTERNAL zmemzero(dest, len)
+    Bytef* dest;
+    uInt  len;
+{
+    if (len == 0) return;
+    do {
+        *dest++ = 0;  /* ??? to be unrolled */
+    } while (--len != 0);
+}
+#endif
+
+#ifndef Z_SOLO
+
+#ifdef SYS16BIT
+
+#ifdef __TURBOC__
+/* Turbo C in 16-bit mode */
+
+#  define MY_ZCALLOC
+
+/* Turbo C malloc() does not allow dynamic allocation of 64K bytes
+ * and farmalloc(64K) returns a pointer with an offset of 8, so we
+ * must fix the pointer. Warning: the pointer must be put back to its
+ * original form in order to free it, use zcfree().
+ */
+
+#define MAX_PTR 10
+/* 10*64K = 640K */
+
+local int next_ptr = 0;
+
+typedef struct ptr_table_s {
+    voidpf org_ptr;
+    voidpf new_ptr;
+} ptr_table;
+
+local ptr_table table[MAX_PTR];
+/* This table is used to remember the original form of pointers
+ * to large buffers (64K). Such pointers are normalized with a zero offset.
+ * Since MSDOS is not a preemptive multitasking OS, this table is not
+ * protected from concurrent access. This hack doesn't work anyway on
+ * a protected system like OS/2. Use Microsoft C instead.
+ */
+
+voidpf ZLIB_INTERNAL zcalloc (voidpf opaque, unsigned items, unsigned size)
+{
+    voidpf buf = opaque; /* just to make some compilers happy */
+    ulg bsize = (ulg)items*size;
+
+    /* If we allocate less than 65520 bytes, we assume that farmalloc
+     * will return a usable pointer which doesn't have to be normalized.
+     */
+    if (bsize < 65520L) {
+        buf = farmalloc(bsize);
+        if (*(ush*)&buf != 0) return buf;
+    } else {
+        buf = farmalloc(bsize + 16L);
+    }
+    if (buf == NULL || next_ptr >= MAX_PTR) return NULL;
+    table[next_ptr].org_ptr = buf;
+
+    /* Normalize the pointer to seg:0 */
+    *((ush*)&buf+1) += ((ush)((uch*)buf-0) + 15) >> 4;
+    *(ush*)&buf = 0;
+    table[next_ptr++].new_ptr = buf;
+    return buf;
+}
+
+void ZLIB_INTERNAL zcfree (voidpf opaque, voidpf ptr)
+{
+    int n;
+    if (*(ush*)&ptr != 0) { /* object < 64K */
+        farfree(ptr);
+        return;
+    }
+    /* Find the original pointer */
+    for (n = 0; n < next_ptr; n++) {
+        if (ptr != table[n].new_ptr) continue;
+
+        farfree(table[n].org_ptr);
+        while (++n < next_ptr) {
+            table[n-1] = table[n];
+        }
+        next_ptr--;
+        return;
+    }
+    ptr = opaque; /* just to make some compilers happy */
+    Assert(0, "zcfree: ptr not found");
+}
+
+#endif /* __TURBOC__ */
+
+
+#ifdef M_I86
+/* Microsoft C in 16-bit mode */
+
+#  define MY_ZCALLOC
+
+#if (!defined(_MSC_VER) || (_MSC_VER <= 600))
+#  define _halloc  halloc
+#  define _hfree   hfree
+#endif
+
+voidpf ZLIB_INTERNAL zcalloc (voidpf opaque, uInt items, uInt size)
+{
+    if (opaque) opaque = 0; /* to make compiler happy */
+    return _halloc((long)items, size);
+}
+
+void ZLIB_INTERNAL zcfree (voidpf opaque, voidpf ptr)
+{
+    if (opaque) opaque = 0; /* to make compiler happy */
+    _hfree(ptr);
+}
+
+#endif /* M_I86 */
+
+#endif /* SYS16BIT */
+
+
+#ifndef MY_ZCALLOC /* Any system without a special alloc function */
+
+#ifndef STDC
+extern voidp  malloc OF((uInt size));
+extern voidp  calloc OF((uInt items, uInt size));
+extern void   free   OF((voidpf ptr));
+#endif
+
+voidpf ZLIB_INTERNAL zcalloc (opaque, items, size)
+    voidpf opaque;
+    unsigned items;
+    unsigned size;
+{
+    if (opaque) items += size - size; /* make compiler happy */
+    return sizeof(uInt) > 2 ? (voidpf)malloc(items * size) :
+                              (voidpf)calloc(items, size);
+}
+
+void ZLIB_INTERNAL zcfree (opaque, ptr)
+    voidpf opaque;
+    voidpf ptr;
+{
+    free(ptr);
+    if (opaque) return; /* make compiler happy */
+}
+
+#endif /* MY_ZCALLOC */
+
+#endif /* !Z_SOLO */
diff --git a/ml/dlib/dlib/external/zlib/zutil.h b/ml/dlib/dlib/external/zlib/zutil.h
new file mode 100644
index 000000000..24ab06b1c
--- /dev/null
+++ b/ml/dlib/dlib/external/zlib/zutil.h
@@ -0,0 +1,253 @@
+/* zutil.h -- internal interface and configuration of the compression library
+ * Copyright (C) 1995-2013 Jean-loup Gailly.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+/* WARNING: this file should *not* be used by applications. It is
+   part of the implementation of the compression library and is
+   subject to change. Applications should only use zlib.h.
+ */
+
+/* @(#) $Id$ */
+
+#ifndef ZUTIL_H
+#define ZUTIL_H
+
+#ifdef HAVE_HIDDEN
+#  define ZLIB_INTERNAL __attribute__((visibility ("hidden")))
+#else
+#  define ZLIB_INTERNAL
+#endif
+
+#include "zlib.h"
+
+#if defined(STDC) && !defined(Z_SOLO)
+#  if !(defined(_WIN32_WCE) && defined(_MSC_VER))
+#    include <stddef.h>
+#  endif
+#  include <string.h>
+#  include <stdlib.h>
+#endif
+
+#ifdef Z_SOLO
+   typedef long ptrdiff_t;  /* guess -- will be caught if guess is wrong */
+#endif
+
+#ifndef local
+#  define local static
+#endif
+/* compile with -Dlocal if your debugger can't find static symbols */
+
+typedef unsigned char  uch;
+typedef uch FAR uchf;
+typedef unsigned short ush;
+typedef ush FAR ushf;
+typedef unsigned long  ulg;
+
+extern z_const char * const z_errmsg[10]; /* indexed by 2-zlib_error */
+/* (size given to avoid silly warnings with Visual C++) */
+
+#define ERR_MSG(err) z_errmsg[Z_NEED_DICT-(err)]
+
+#define ERR_RETURN(strm,err) \
+  return (strm->msg = ERR_MSG(err), (err))
+/* To be used only when the state is known to be valid */
+
+        /* common constants */
+
+#ifndef DEF_WBITS
+#  define DEF_WBITS MAX_WBITS
+#endif
+/* default windowBits for decompression. MAX_WBITS is for compression only */
+
+#if MAX_MEM_LEVEL >= 8
+#  define DEF_MEM_LEVEL 8
+#else
+#  define DEF_MEM_LEVEL  MAX_MEM_LEVEL
+#endif
+/* default memLevel */
+
+#define STORED_BLOCK 0
+#define STATIC_TREES 1
+#define DYN_TREES    2
+/* The three kinds of block type */
+
+#define MIN_MATCH  3
+#define MAX_MATCH  258
+/* The minimum and maximum match lengths */
+
+#define PRESET_DICT 0x20 /* preset dictionary flag in zlib header */
+
+        /* target dependencies */
+
+#if defined(MSDOS) || (defined(WINDOWS) && !defined(WIN32))
+#  define OS_CODE  0x00
+#  ifndef Z_SOLO
+#    if defined(__TURBOC__) || defined(__BORLANDC__)
+#      if (__STDC__ == 1) && (defined(__LARGE__) || defined(__COMPACT__))
+         /* Allow compilation with ANSI keywords only enabled */
+         void _Cdecl farfree( void *block );
+         void *_Cdecl farmalloc( unsigned long nbytes );
+#      else
+#        include <alloc.h>
+#      endif
+#    else /* MSC or DJGPP */
+#      include <malloc.h>
+#    endif
+#  endif
+#endif
+
+#ifdef AMIGA
+#  define OS_CODE  0x01
+#endif
+
+#if defined(VAXC) || defined(VMS)
+#  define OS_CODE  0x02
+#  define F_OPEN(name, mode) \
+     fopen((name), (mode), "mbc=60", "ctx=stm", "rfm=fix", "mrs=512")
+#endif
+
+#if defined(ATARI) || defined(atarist)
+#  define OS_CODE  0x05
+#endif
+
+#ifdef OS2
+#  define OS_CODE  0x06
+#  if defined(M_I86) && !defined(Z_SOLO)
+#    include <malloc.h>
+#  endif
+#endif
+
+#if defined(MACOS) || defined(TARGET_OS_MAC)
+#  define OS_CODE  0x07
+#  ifndef Z_SOLO
+#    if defined(__MWERKS__) && __dest_os != __be_os && __dest_os != __win32_os
+#      include <unix.h> /* for fdopen */
+#    else
+#      ifndef fdopen
+#        define fdopen(fd,mode) NULL /* No fdopen() */
+#      endif
+#    endif
+#  endif
+#endif
+
+#ifdef TOPS20
+#  define OS_CODE  0x0a
+#endif
+
+#ifdef WIN32
+#  ifndef __CYGWIN__  /* Cygwin is Unix, not Win32 */
+#    define OS_CODE  0x0b
+#  endif
+#endif
+
+#ifdef __50SERIES /* Prime/PRIMOS */
+#  define OS_CODE  0x0f
+#endif
+
+#if defined(_BEOS_) || defined(RISCOS)
+#  define fdopen(fd,mode) NULL /* No fdopen() */
+#endif
+
+#if (defined(_MSC_VER) && (_MSC_VER > 600)) && !defined __INTERIX
+#  if defined(_WIN32_WCE)
+#    define fdopen(fd,mode) NULL /* No fdopen() */
+#    ifndef _PTRDIFF_T_DEFINED
+       typedef int ptrdiff_t;
+#      define _PTRDIFF_T_DEFINED
+#    endif
+#  else
+#    define fdopen(fd,type)  _fdopen(fd,type)
+#  endif
+#endif
+
+#if defined(__BORLANDC__) && !defined(MSDOS)
+  #pragma warn -8004
+  #pragma warn -8008
+  #pragma warn -8066
+#endif
+
+/* provide prototypes for these when building zlib without LFS */
+#if !defined(_WIN32) && \
+    (!defined(_LARGEFILE64_SOURCE) || _LFS64_LARGEFILE-0 == 0)
+    ZEXTERN uLong ZEXPORT adler32_combine64 OF((uLong, uLong, z_off_t));
+    ZEXTERN uLong ZEXPORT crc32_combine64 OF((uLong, uLong, z_off_t));
+#endif
+
+        /* common defaults */
+
+#ifndef OS_CODE
+#  define OS_CODE  0x03  /* assume Unix */
+#endif
+
+#ifndef F_OPEN
+#  define F_OPEN(name, mode) fopen((name), (mode))
+#endif
+
+         /* functions */
+
+#if defined(pyr) || defined(Z_SOLO)
+#  define NO_MEMCPY
+#endif
+#if defined(SMALL_MEDIUM) && !defined(_MSC_VER) && !defined(__SC__)
+ /* Use our own functions for small and medium model with MSC <= 5.0.
+  * You may have to use the same strategy for Borland C (untested).
+  * The __SC__ check is for Symantec.
+  */
+#  define NO_MEMCPY
+#endif
+#if defined(STDC) && !defined(HAVE_MEMCPY) && !defined(NO_MEMCPY)
+#  define HAVE_MEMCPY
+#endif
+#ifdef HAVE_MEMCPY
+#  ifdef SMALL_MEDIUM /* MSDOS small or medium model */
+#    define zmemcpy _fmemcpy
+#    define zmemcmp _fmemcmp
+#    define zmemzero(dest, len) _fmemset(dest, 0, len)
+#  else
+#    define zmemcpy memcpy
+#    define zmemcmp memcmp
+#    define zmemzero(dest, len) memset(dest, 0, len)
+#  endif
+#else
+   void ZLIB_INTERNAL zmemcpy OF((Bytef* dest, const Bytef* source, uInt len));
+   int ZLIB_INTERNAL zmemcmp OF((const Bytef* s1, const Bytef* s2, uInt len));
+   void ZLIB_INTERNAL zmemzero OF((Bytef* dest, uInt len));
+#endif
+
+/* Diagnostic functions */
+#ifdef DEBUG
+#  include <stdio.h>
+   extern int ZLIB_INTERNAL z_verbose;
+   extern void ZLIB_INTERNAL z_error OF((char *m));
+#  define Assert(cond,msg) {if(!(cond)) z_error(msg);}
+#  define Trace(x) {if (z_verbose>=0) fprintf x ;}
+#  define Tracev(x) {if (z_verbose>0) fprintf x ;}
+#  define Tracevv(x) {if (z_verbose>1) fprintf x ;}
+#  define Tracec(c,x) {if (z_verbose>0 && (c)) fprintf x ;}
+#  define Tracecv(c,x) {if (z_verbose>1 && (c)) fprintf x ;}
+#else
+#  define Assert(cond,msg)
+#  define Trace(x)
+#  define Tracev(x)
+#  define Tracevv(x)
+#  define Tracec(c,x)
+#  define Tracecv(c,x)
+#endif
+
+#ifndef Z_SOLO
+   voidpf ZLIB_INTERNAL zcalloc OF((voidpf opaque, unsigned items,
+                                    unsigned size));
+   void ZLIB_INTERNAL zcfree  OF((voidpf opaque, voidpf ptr));
+#endif
+
+#define ZALLOC(strm, items, size) \
+           (*((strm)->zalloc))((strm)->opaque, (items), (size))
+#define ZFREE(strm, addr)  (*((strm)->zfree))((strm)->opaque, (voidpf)(addr))
+#define TRY_FREE(s, p) {if (p) ZFREE(s, p);}
+
+/* Reverse the bytes in a 32-bit value */
+#define ZSWAP32(q) ((((q) >> 24) & 0xff) + (((q) >> 8) & 0xff00) + \
+                    (((q) & 0xff00) << 8) + (((q) & 0xff) << 24))
+
+#endif /* ZUTIL_H */